diff --git a/kernel/include/hfi1/sdma.h b/kernel/include/hfi1/sdma.h new file mode 100644 index 00000000..7edcaaa7 --- /dev/null +++ b/kernel/include/hfi1/sdma.h @@ -0,0 +1,1149 @@ +#ifndef _HFI1_SDMA_H +#define _HFI1_SDMA_H +/* + * Copyright(c) 2015, 2016 Intel Corporation. + * + * This file is provided under a dual BSD/GPLv2 license. When using or + * redistributing this file, you may do so under either license. + * + * GPL LICENSE SUMMARY + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License for more details. + * + * BSD LICENSE + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in + * the documentation and/or other materials provided with the + * distribution. + * - Neither the name of Intel Corporation nor the names of its + * contributors may be used to endorse or promote products derived + * from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#include +#include + +#ifdef __HFI1_ORIG__ + +#include +#include +#include +#include +#include + +#include "hfi.h" +#include "verbs.h" +#include "sdma_txreq.h" + +#define hfi1_cdbg(which, fmt, ...) \ + __hfi1_trace_##which(__func__, fmt, ##__VA_ARGS__) +extern void __hfi1_trace_AIOWRITE(const char *func, char *fmt, ...); +#endif /* __HFI1_ORIG__ */ + +/* Hardware limit */ +#define MAX_DESC 64 +/* Hardware limit for SDMA packet size */ +#define MAX_SDMA_PKT_SIZE ((16 * 1024) - 1) + +#define SDMA_TXREQ_S_OK 0 +#define SDMA_TXREQ_S_SENDERROR 1 +#define SDMA_TXREQ_S_ABORTED 2 +#define SDMA_TXREQ_S_SHUTDOWN 3 + +/* flags bits */ +#define SDMA_TXREQ_F_URGENT 0x0001 +#define SDMA_TXREQ_F_AHG_COPY 0x0002 +#define SDMA_TXREQ_F_USE_AHG 0x0004 + +#define SDMA_MAP_NONE 0 +#define SDMA_MAP_SINGLE 1 +#define SDMA_MAP_PAGE 2 + +#define SDMA_AHG_VALUE_MASK 0xffff +#define SDMA_AHG_VALUE_SHIFT 0 +#define SDMA_AHG_INDEX_MASK 0xf +#define SDMA_AHG_INDEX_SHIFT 16 +#define SDMA_AHG_FIELD_LEN_MASK 0xf +#define SDMA_AHG_FIELD_LEN_SHIFT 20 +#define SDMA_AHG_FIELD_START_MASK 0x1f +#define SDMA_AHG_FIELD_START_SHIFT 24 +#define SDMA_AHG_UPDATE_ENABLE_MASK 0x1 +#define SDMA_AHG_UPDATE_ENABLE_SHIFT 31 + +/* AHG modes */ + +/* + * Be aware the ordering and values + * for SDMA_AHG_APPLY_UPDATE[123] + * are assumed in generating a skip + * count in submit_tx() in sdma.c + */ +#define SDMA_AHG_NO_AHG 0 +#define SDMA_AHG_COPY 1 +#define SDMA_AHG_APPLY_UPDATE1 2 +#define SDMA_AHG_APPLY_UPDATE2 3 +#define SDMA_AHG_APPLY_UPDATE3 4 + +/* + * Bits defined in the send DMA descriptor. + */ +#define SDMA_DESC0_FIRST_DESC_FLAG BIT_ULL(63) +#define SDMA_DESC0_LAST_DESC_FLAG BIT_ULL(62) +#define SDMA_DESC0_BYTE_COUNT_SHIFT 48 +#define SDMA_DESC0_BYTE_COUNT_WIDTH 14 +#define SDMA_DESC0_BYTE_COUNT_MASK \ + ((1ULL << SDMA_DESC0_BYTE_COUNT_WIDTH) - 1) +#define SDMA_DESC0_BYTE_COUNT_SMASK \ + (SDMA_DESC0_BYTE_COUNT_MASK << SDMA_DESC0_BYTE_COUNT_SHIFT) +#define SDMA_DESC0_PHY_ADDR_SHIFT 0 +#define SDMA_DESC0_PHY_ADDR_WIDTH 48 +#define SDMA_DESC0_PHY_ADDR_MASK \ + ((1ULL << SDMA_DESC0_PHY_ADDR_WIDTH) - 1) +#define SDMA_DESC0_PHY_ADDR_SMASK \ + (SDMA_DESC0_PHY_ADDR_MASK << SDMA_DESC0_PHY_ADDR_SHIFT) + +#define SDMA_DESC1_HEADER_UPDATE1_SHIFT 32 +#define SDMA_DESC1_HEADER_UPDATE1_WIDTH 32 +#define SDMA_DESC1_HEADER_UPDATE1_MASK \ + ((1ULL << SDMA_DESC1_HEADER_UPDATE1_WIDTH) - 1) +#define SDMA_DESC1_HEADER_UPDATE1_SMASK \ + (SDMA_DESC1_HEADER_UPDATE1_MASK << SDMA_DESC1_HEADER_UPDATE1_SHIFT) +#define SDMA_DESC1_HEADER_MODE_SHIFT 13 +#define SDMA_DESC1_HEADER_MODE_WIDTH 3 +#define SDMA_DESC1_HEADER_MODE_MASK \ + ((1ULL << SDMA_DESC1_HEADER_MODE_WIDTH) - 1) +#define SDMA_DESC1_HEADER_MODE_SMASK \ + (SDMA_DESC1_HEADER_MODE_MASK << SDMA_DESC1_HEADER_MODE_SHIFT) +#define SDMA_DESC1_HEADER_INDEX_SHIFT 8 +#define SDMA_DESC1_HEADER_INDEX_WIDTH 5 +#define SDMA_DESC1_HEADER_INDEX_MASK \ + ((1ULL << SDMA_DESC1_HEADER_INDEX_WIDTH) - 1) +#define SDMA_DESC1_HEADER_INDEX_SMASK \ + (SDMA_DESC1_HEADER_INDEX_MASK << SDMA_DESC1_HEADER_INDEX_SHIFT) +#define SDMA_DESC1_HEADER_DWS_SHIFT 4 +#define SDMA_DESC1_HEADER_DWS_WIDTH 4 +#define SDMA_DESC1_HEADER_DWS_MASK \ + ((1ULL << SDMA_DESC1_HEADER_DWS_WIDTH) - 1) +#define SDMA_DESC1_HEADER_DWS_SMASK \ + (SDMA_DESC1_HEADER_DWS_MASK << SDMA_DESC1_HEADER_DWS_SHIFT) +#define SDMA_DESC1_GENERATION_SHIFT 2 +#define SDMA_DESC1_GENERATION_WIDTH 2 +#define SDMA_DESC1_GENERATION_MASK \ + ((1ULL << SDMA_DESC1_GENERATION_WIDTH) - 1) +#define SDMA_DESC1_GENERATION_SMASK \ + (SDMA_DESC1_GENERATION_MASK << SDMA_DESC1_GENERATION_SHIFT) +#define SDMA_DESC1_INT_REQ_FLAG BIT_ULL(1) +#define SDMA_DESC1_HEAD_TO_HOST_FLAG BIT_ULL(0) + +enum sdma_states { + sdma_state_s00_hw_down, + sdma_state_s10_hw_start_up_halt_wait, + sdma_state_s15_hw_start_up_clean_wait, + sdma_state_s20_idle, + sdma_state_s30_sw_clean_up_wait, + sdma_state_s40_hw_clean_up_wait, + sdma_state_s50_hw_halt_wait, + sdma_state_s60_idle_halt_wait, + sdma_state_s80_hw_freeze, + sdma_state_s82_freeze_sw_clean, + sdma_state_s99_running, +}; + +enum sdma_events { + sdma_event_e00_go_hw_down, + sdma_event_e10_go_hw_start, + sdma_event_e15_hw_halt_done, + sdma_event_e25_hw_clean_up_done, + sdma_event_e30_go_running, + sdma_event_e40_sw_cleaned, + sdma_event_e50_hw_cleaned, + sdma_event_e60_hw_halted, + sdma_event_e70_go_idle, + sdma_event_e80_hw_freeze, + sdma_event_e81_hw_frozen, + sdma_event_e82_hw_unfreeze, + sdma_event_e85_link_down, + sdma_event_e90_sw_halted, +}; + +struct sdma_set_state_action { + unsigned op_enable:1; + unsigned op_intenable:1; + unsigned op_halt:1; + unsigned op_cleanup:1; + unsigned go_s99_running_tofalse:1; + unsigned go_s99_running_totrue:1; +}; + +#ifdef __HFI1_ORIG__ +struct sdma_state { + struct kref kref; + struct completion comp; + enum sdma_states current_state; + unsigned current_op; + unsigned go_s99_running; + /* debugging/development */ + enum sdma_states previous_state; + unsigned previous_op; + enum sdma_events last_event; +}; +#else +struct sdma_state { + enum sdma_states current_state; +}; +#endif /* __HFI1_ORIG__ */ + +/** + * DOC: sdma exported routines + * + * These sdma routines fit into three categories: + * - The SDMA API for building and submitting packets + * to the ring + * + * - Initialization and tear down routines to buildup + * and tear down SDMA + * + * - ISR entrances to handle interrupts, state changes + * and errors + */ + +/** + * DOC: sdma PSM/verbs API + * + * The sdma API is designed to be used by both PSM + * and verbs to supply packets to the SDMA ring. + * + * The usage of the API is as follows: + * + * Embed a struct iowait in the QP or + * PQ. The iowait should be initialized with a + * call to iowait_init(). + * + * The user of the API should create an allocation method + * for their version of the txreq. slabs, pre-allocated lists, + * and dma pools can be used. Once the user's overload of + * the sdma_txreq has been allocated, the sdma_txreq member + * must be initialized with sdma_txinit() or sdma_txinit_ahg(). + * + * The txreq must be declared with the sdma_txreq first. + * + * The tx request, once initialized, is manipulated with calls to + * sdma_txadd_daddr(), sdma_txadd_page(), or sdma_txadd_kvaddr() + * for each disjoint memory location. It is the user's responsibility + * to understand the packet boundaries and page boundaries to do the + * appropriate number of sdma_txadd_* calls.. The user + * must be prepared to deal with failures from these routines due to + * either memory allocation or dma_mapping failures. + * + * The mapping specifics for each memory location are recorded + * in the tx. Memory locations added with sdma_txadd_page() + * and sdma_txadd_kvaddr() are automatically mapped when added + * to the tx and nmapped as part of the progress processing in the + * SDMA interrupt handling. + * + * sdma_txadd_daddr() is used to add an dma_addr_t memory to the + * tx. An example of a use case would be a pre-allocated + * set of headers allocated via dma_pool_alloc() or + * dma_alloc_coherent(). For these memory locations, it + * is the responsibility of the user to handle that unmapping. + * (This would usually be at an unload or job termination.) + * + * The routine sdma_send_txreq() is used to submit + * a tx to the ring after the appropriate number of + * sdma_txadd_* have been done. + * + * If it is desired to send a burst of sdma_txreqs, sdma_send_txlist() + * can be used to submit a list of packets. + * + * The user is free to use the link overhead in the struct sdma_txreq as + * long as the tx isn't in flight. + * + * The extreme degenerate case of the number of descriptors + * exceeding the ring size is automatically handled as + * memory locations are added. An overflow of the descriptor + * array that is part of the sdma_txreq is also automatically + * handled. + * + */ + +/** + * DOC: Infrastructure calls + * + * sdma_init() is used to initialize data structures and + * CSRs for the desired number of SDMA engines. + * + * sdma_start() is used to kick the SDMA engines initialized + * with sdma_init(). Interrupts must be enabled at this + * point since aspects of the state machine are interrupt + * driven. + * + * sdma_engine_error() and sdma_engine_interrupt() are + * entrances for interrupts. + * + * sdma_map_init() is for the management of the mapping + * table when the number of vls is changed. + * + */ + +/* + * struct hw_sdma_desc - raw 128 bit SDMA descriptor + * + * This is the raw descriptor in the SDMA ring + */ +struct hw_sdma_desc { + /* private: don't use directly */ + __le64 qw[2]; +}; + +/** + * struct sdma_engine - Data pertaining to each SDMA engine. + * @dd: a back-pointer to the device data + * @ppd: per port back-pointer + * @imask: mask for irq manipulation + * @idle_mask: mask for determining if an interrupt is due to sdma_idle + * + * This structure has the state for each sdma_engine. + * + * Accessing to non public fields are not supported + * since the private members are subject to change. + */ +struct sdma_engine { + /* read mostly */ + struct hfi1_devdata *dd; + struct hfi1_pportdata *ppd; + /* private: */ + void __iomem *tail_csr; + u64 imask; /* clear interrupt mask */ + u64 idle_mask; + u64 progress_mask; + u64 int_mask; + /* private: */ + volatile __le64 *head_dma; /* DMA'ed by chip */ + /* private: */ + dma_addr_t head_phys; + /* private: */ + struct hw_sdma_desc *descq; + /* private: */ + unsigned descq_full_count; + struct sdma_txreq **tx_ring; + /* private: */ + dma_addr_t descq_phys; + /* private */ + u32 sdma_mask; + /* private */ + struct sdma_state state; + /* private */ + int cpu; + /* private: */ + u8 sdma_shift; + /* private: */ + u8 this_idx; /* zero relative engine */ + /* protect changes to senddmactrl shadow */ + spinlock_t senddmactrl_lock; + /* private: */ + u64 p_senddmactrl; /* shadow per-engine SendDmaCtrl */ + + /* read/write using tail_lock */ + spinlock_t tail_lock ____cacheline_aligned_in_smp; +#ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER + /* private: */ + u64 tail_sn; +#endif + /* private: */ + u32 descq_tail; + /* private: */ + unsigned long ahg_bits; + /* private: */ + u16 desc_avail; + /* private: */ + u16 tx_tail; + /* private: */ + u16 descq_cnt; + + /* read/write using head_lock */ + /* private: */ + seqlock_t head_lock ____cacheline_aligned_in_smp; +#ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER + /* private: */ + u64 head_sn; +#endif + /* private: */ + u32 descq_head; + /* private: */ + u16 tx_head; + /* private: */ + u64 last_status; + /* private */ + u64 err_cnt; + /* private */ + u64 sdma_int_cnt; + u64 idle_int_cnt; + u64 progress_int_cnt; + + /* private: */ + struct list_head dmawait; + +#ifdef __HFI1_ORIG__ + /* CONFIG SDMA for now, just blindly duplicate */ + /* private: */ + struct tasklet_struct sdma_hw_clean_up_task + ____cacheline_aligned_in_smp; + + /* private: */ + struct tasklet_struct sdma_sw_clean_up_task + ____cacheline_aligned_in_smp; + /* private: */ + struct work_struct err_halt_worker; + /* private */ + struct timer_list err_progress_check_timer; + u32 progress_check_head; + /* private: */ + struct work_struct flush_worker; +#endif /* __HFI1_ORIG__ */ + /* protect flush list */ + spinlock_t flushlist_lock; + /* private: */ + struct list_head flushlist; +#ifdef __HFI1_ORIG__ + struct cpumask cpu_mask; + struct kobject kobj; +#endif /* __HFI1_ORIG__ */ +}; + + +#ifdef __HFI1_ORIG__ + +int sdma_init(struct hfi1_devdata *dd, u8 port); +void sdma_start(struct hfi1_devdata *dd); +void sdma_exit(struct hfi1_devdata *dd); +void sdma_all_running(struct hfi1_devdata *dd); +void sdma_all_idle(struct hfi1_devdata *dd); +void sdma_freeze_notify(struct hfi1_devdata *dd, int go_idle); +void sdma_freeze(struct hfi1_devdata *dd); +void sdma_unfreeze(struct hfi1_devdata *dd); +void sdma_wait(struct hfi1_devdata *dd); + +/** + * sdma_empty() - idle engine test + * @engine: sdma engine + * + * Currently used by verbs as a latency optimization. + * + * Return: + * 1 - empty, 0 - non-empty + */ +static inline int sdma_empty(struct sdma_engine *sde) +{ + return sde->descq_tail == sde->descq_head; +} + +#endif /* __HFI1_ORIG__ */ +static inline u16 sdma_descq_freecnt(struct sdma_engine *sde) +{ + return sde->descq_cnt - + (sde->descq_tail - + ACCESS_ONCE(sde->descq_head)) - 1; +} + +static inline u16 sdma_descq_inprocess(struct sdma_engine *sde) +{ + return sde->descq_cnt - sdma_descq_freecnt(sde); +} + +/* + * Either head_lock or tail lock required to see + * a steady state. + */ +static inline int __sdma_running(struct sdma_engine *engine) +{ + return engine->state.current_state == sdma_state_s99_running; +} + +/** + * sdma_running() - state suitability test + * @engine: sdma engine + * + * sdma_running probes the internal state to determine if it is suitable + * for submitting packets. + * + * Return: + * 1 - ok to submit, 0 - not ok to submit + * + */ +static inline int sdma_running(struct sdma_engine *engine) +{ + unsigned long flags; + int ret; + + hfi1_cdbg(AIOWRITE, "+"); + spin_lock_irqsave(&engine->tail_lock, flags); + ret = __sdma_running(engine); + spin_unlock_irqrestore(&engine->tail_lock, flags); + hfi1_cdbg(AIOWRITE, "-"); + return ret; +} + +void _sdma_txreq_ahgadd( + struct sdma_txreq *tx, + u8 num_ahg, + u8 ahg_entry, + u32 *ahg, + u8 ahg_hlen); + +/** + * sdma_txinit_ahg() - initialize an sdma_txreq struct with AHG + * @tx: tx request to initialize + * @flags: flags to key last descriptor additions + * @tlen: total packet length (pbc + headers + data) + * @ahg_entry: ahg entry to use (0 - 31) + * @num_ahg: ahg descriptor for first descriptor (0 - 9) + * @ahg: array of AHG descriptors (up to 9 entries) + * @ahg_hlen: number of bytes from ASIC entry to use + * @cb: callback + * + * The allocation of the sdma_txreq and it enclosing structure is user + * dependent. This routine must be called to initialize the user independent + * fields. + * + * The currently supported flags are SDMA_TXREQ_F_URGENT, + * SDMA_TXREQ_F_AHG_COPY, and SDMA_TXREQ_F_USE_AHG. + * + * SDMA_TXREQ_F_URGENT is used for latency sensitive situations where the + * completion is desired as soon as possible. + * + * SDMA_TXREQ_F_AHG_COPY causes the header in the first descriptor to be + * copied to chip entry. SDMA_TXREQ_F_USE_AHG causes the code to add in + * the AHG descriptors into the first 1 to 3 descriptors. + * + * Completions of submitted requests can be gotten on selected + * txreqs by giving a completion routine callback to sdma_txinit() or + * sdma_txinit_ahg(). The environment in which the callback runs + * can be from an ISR, a tasklet, or a thread, so no sleeping + * kernel routines can be used. Aspects of the sdma ring may + * be locked so care should be taken with locking. + * + * The callback pointer can be NULL to avoid any callback for the packet + * being submitted. The callback will be provided this tx, a status, and a flag. + * + * The status will be one of SDMA_TXREQ_S_OK, SDMA_TXREQ_S_SENDERROR, + * SDMA_TXREQ_S_ABORTED, or SDMA_TXREQ_S_SHUTDOWN. + * + * The flag, if the is the iowait had been used, indicates the iowait + * sdma_busy count has reached zero. + * + * user data portion of tlen should be precise. The sdma_txadd_* entrances + * will pad with a descriptor references 1 - 3 bytes when the number of bytes + * specified in tlen have been supplied to the sdma_txreq. + * + * ahg_hlen is used to determine the number of on-chip entry bytes to + * use as the header. This is for cases where the stored header is + * larger than the header to be used in a packet. This is typical + * for verbs where an RDMA_WRITE_FIRST is larger than the packet in + * and RDMA_WRITE_MIDDLE. + * + */ +static inline int sdma_txinit_ahg( + struct sdma_txreq *tx, + u16 flags, + u16 tlen, + u8 ahg_entry, + u8 num_ahg, + u32 *ahg, + u8 ahg_hlen, + void (*cb)(struct sdma_txreq *, int)) +{ + if (tlen == 0) + return -ENODATA; + if (tlen > MAX_SDMA_PKT_SIZE) + return -EMSGSIZE; + tx->desc_limit = ARRAY_SIZE(tx->descs); + tx->descp = &tx->descs[0]; + INIT_LIST_HEAD(&tx->list); + tx->num_desc = 0; + tx->flags = flags; + tx->complete = cb; + tx->coalesce_buf = NULL; + tx->wait = NULL; + tx->packet_len = tlen; + tx->tlen = tx->packet_len; + tx->descs[0].qw[0] = SDMA_DESC0_FIRST_DESC_FLAG; + tx->descs[0].qw[1] = 0; + if (flags & SDMA_TXREQ_F_AHG_COPY) + tx->descs[0].qw[1] |= + (((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK) + << SDMA_DESC1_HEADER_INDEX_SHIFT) | + (((u64)SDMA_AHG_COPY & SDMA_DESC1_HEADER_MODE_MASK) + << SDMA_DESC1_HEADER_MODE_SHIFT); + else if (flags & SDMA_TXREQ_F_USE_AHG && num_ahg) + _sdma_txreq_ahgadd(tx, num_ahg, ahg_entry, ahg, ahg_hlen); + return 0; +} + +/** + * sdma_txinit() - initialize an sdma_txreq struct (no AHG) + * @tx: tx request to initialize + * @flags: flags to key last descriptor additions + * @tlen: total packet length (pbc + headers + data) + * @cb: callback pointer + * + * The allocation of the sdma_txreq and it enclosing structure is user + * dependent. This routine must be called to initialize the user + * independent fields. + * + * The currently supported flags is SDMA_TXREQ_F_URGENT. + * + * SDMA_TXREQ_F_URGENT is used for latency sensitive situations where the + * completion is desired as soon as possible. + * + * Completions of submitted requests can be gotten on selected + * txreqs by giving a completion routine callback to sdma_txinit() or + * sdma_txinit_ahg(). The environment in which the callback runs + * can be from an ISR, a tasklet, or a thread, so no sleeping + * kernel routines can be used. The head size of the sdma ring may + * be locked so care should be taken with locking. + * + * The callback pointer can be NULL to avoid any callback for the packet + * being submitted. + * + * The callback, if non-NULL, will be provided this tx and a status. The + * status will be one of SDMA_TXREQ_S_OK, SDMA_TXREQ_S_SENDERROR, + * SDMA_TXREQ_S_ABORTED, or SDMA_TXREQ_S_SHUTDOWN. + * + */ +static inline int sdma_txinit( + struct sdma_txreq *tx, + u16 flags, + u16 tlen, + void (*cb)(struct sdma_txreq *, int)) +{ + return sdma_txinit_ahg(tx, flags, tlen, 0, 0, NULL, 0, cb); +} +#ifdef __HFI1_ORIG__ + +/* helpers - don't use */ +static inline int sdma_mapping_type(struct sdma_desc *d) +{ + return (d->qw[1] & SDMA_DESC1_GENERATION_SMASK) + >> SDMA_DESC1_GENERATION_SHIFT; +} + +static inline size_t sdma_mapping_len(struct sdma_desc *d) +{ + return (d->qw[0] & SDMA_DESC0_BYTE_COUNT_SMASK) + >> SDMA_DESC0_BYTE_COUNT_SHIFT; +} + +static inline dma_addr_t sdma_mapping_addr(struct sdma_desc *d) +{ + return (d->qw[0] & SDMA_DESC0_PHY_ADDR_SMASK) + >> SDMA_DESC0_PHY_ADDR_SHIFT; +} + +#endif /* __HFI1_ORIG__ */ +static inline void make_tx_sdma_desc( + struct sdma_txreq *tx, + int type, + dma_addr_t addr, + size_t len) +{ + struct sdma_desc *desc = &tx->descp[tx->num_desc]; + + if (!tx->num_desc) { + /* qw[0] zero; qw[1] first, ahg mode already in from init */ + desc->qw[1] |= ((u64)type & SDMA_DESC1_GENERATION_MASK) + << SDMA_DESC1_GENERATION_SHIFT; + } else { + desc->qw[0] = 0; + desc->qw[1] = ((u64)type & SDMA_DESC1_GENERATION_MASK) + << SDMA_DESC1_GENERATION_SHIFT; + } + desc->qw[0] |= (((u64)addr & SDMA_DESC0_PHY_ADDR_MASK) + << SDMA_DESC0_PHY_ADDR_SHIFT) | + (((u64)len & SDMA_DESC0_BYTE_COUNT_MASK) + << SDMA_DESC0_BYTE_COUNT_SHIFT); +} + +/* helper to extend txreq */ +int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx, + int type, void *kvaddr, struct page *page, + unsigned long offset, u16 len); +void __sdma_txclean(struct hfi1_devdata *, struct sdma_txreq *); + +static inline void sdma_txclean(struct hfi1_devdata *dd, struct sdma_txreq *tx) +{ + if (tx->num_desc) + __sdma_txclean(dd, tx); +} +#ifdef __HFI1_ORIG__ +int _pad_sdma_tx_descs(struct hfi1_devdata *, struct sdma_txreq *); + +/* helpers used by public routines */ +static inline void _sdma_close_tx(struct hfi1_devdata *dd, + struct sdma_txreq *tx) +{ + tx->descp[tx->num_desc].qw[0] |= + SDMA_DESC0_LAST_DESC_FLAG; + tx->descp[tx->num_desc].qw[1] |= + dd->default_desc1; + if (tx->flags & SDMA_TXREQ_F_URGENT) + tx->descp[tx->num_desc].qw[1] |= + (SDMA_DESC1_HEAD_TO_HOST_FLAG | + SDMA_DESC1_INT_REQ_FLAG); +} + +#endif /* __HFI1_ORIG__ */ +static inline int _sdma_txadd_daddr( + struct hfi1_devdata *dd, + int type, + struct sdma_txreq *tx, + dma_addr_t addr, + u16 len) +{ + int rval = 0; + + make_tx_sdma_desc( + tx, + type, + addr, len); + WARN_ON(len > tx->tlen); + tx->tlen -= len; + /* special cases for last */ + if (!tx->tlen) { + if (tx->packet_len & (sizeof(u32) - 1)) { + //TODO: _pad_sdma_tx_descs + //rval = _pad_sdma_tx_descs(dd, tx); + if (rval) + return rval; + } else { + //TODO: _sdma_close_tx + //_sdma_close_tx(dd, tx); + } + } + tx->num_desc++; + return rval; +} + +/** + * sdma_txadd_page() - add a page to the sdma_txreq + * @dd: the device to use for mapping + * @tx: tx request to which the page is added + * @page: page to map + * @offset: offset within the page + * @len: length in bytes + * + * This is used to add a page/offset/length descriptor. + * + * The mapping/unmapping of the page/offset/len is automatically handled. + * + * Return: + * 0 - success, -ENOSPC - mapping fail, -ENOMEM - couldn't + * extend/coalesce descriptor array + */ +static inline int sdma_txadd_page( + struct hfi1_devdata *dd, + struct sdma_txreq *tx, + struct page *page, + unsigned long offset, + u16 len) +{ + dma_addr_t addr; + int rval; + hfi1_cdbg(AIOWRITE, "+"); + if ((unlikely(tx->num_desc == tx->desc_limit))) { + rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_PAGE, + NULL, page, offset, len); + if (rval <= 0) + return rval; + } + +#ifdef __HFI1_ORIG__ + addr = dma_map_page( + &dd->pcidev->dev, + page, + offset, + len, + DMA_TO_DEVICE); + + if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) { + __sdma_txclean(dd, tx); + return -ENOSPC; + } +#else + //TODO: dma_map_page +#endif /* __HFI1_ORIG__ */ + + hfi1_cdbg(AIOWRITE, "-"); + /* + * XXX: It seems that this is the place where the reference to + * the payload is added, but addr is kernel virtual here. + * TODO: verify this by printing it out in Linux. + */ + return _sdma_txadd_daddr( + dd, SDMA_MAP_PAGE, tx, addr, len); +} + +/** + * sdma_txadd_daddr() - add a dma address to the sdma_txreq + * @dd: the device to use for mapping + * @tx: sdma_txreq to which the page is added + * @addr: dma address mapped by caller + * @len: length in bytes + * + * This is used to add a descriptor for memory that is already dma mapped. + * + * In this case, there is no unmapping as part of the progress processing for + * this memory location. + * + * Return: + * 0 - success, -ENOMEM - couldn't extend descriptor array + */ + +static inline int sdma_txadd_daddr( + struct hfi1_devdata *dd, + struct sdma_txreq *tx, + dma_addr_t addr, + u16 len) +{ + int rval; + + if ((unlikely(tx->num_desc == tx->desc_limit))) { + rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_NONE, + NULL, NULL, 0, 0); + if (rval <= 0) + return rval; + } + + return _sdma_txadd_daddr(dd, SDMA_MAP_NONE, tx, addr, len); +} + +/** + * sdma_txadd_kvaddr() - add a kernel virtual address to sdma_txreq + * @dd: the device to use for mapping + * @tx: sdma_txreq to which the page is added + * @kvaddr: the kernel virtual address + * @len: length in bytes + * + * This is used to add a descriptor referenced by the indicated kvaddr and + * len. + * + * The mapping/unmapping of the kvaddr and len is automatically handled. + * + * Return: + * 0 - success, -ENOSPC - mapping fail, -ENOMEM - couldn't extend/coalesce + * descriptor array + */ +static inline int sdma_txadd_kvaddr( + struct hfi1_devdata *dd, + struct sdma_txreq *tx, + void *kvaddr, + u16 len) +{ + dma_addr_t addr; + int rval; + + if ((unlikely(tx->num_desc == tx->desc_limit))) { + rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_SINGLE, + kvaddr, NULL, 0, len); + if (rval <= 0) + return rval; + } + +#ifdef __HFI1_ORIG__ + addr = dma_map_single( + &dd->pcidev->dev, + kvaddr, + len, + DMA_TO_DEVICE); + + if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) { + __sdma_txclean(dd, tx); + return -ENOSPC; + } +#else +//TODO: dma_map_single +#endif /* __HFI1_ORIG__ */ + + return _sdma_txadd_daddr( + dd, SDMA_MAP_SINGLE, tx, addr, len); +} + +struct iowait_wait; +int sdma_send_txreq(struct sdma_engine *sde, + struct iowait_work *wait, + struct sdma_txreq *tx); +int sdma_send_txlist(struct sdma_engine *sde, + struct iowait_work *wait, + struct list_head *tx_list, + u32 *count); + +int sdma_ahg_alloc(struct sdma_engine *sde); +void sdma_ahg_free(struct sdma_engine *sde, int ahg_index); + +/** + * sdma_build_ahg - build ahg descriptor + * @data + * @dwindex + * @startbit + * @bits + * + * Build and return a 32 bit descriptor. + */ +static inline u32 sdma_build_ahg_descriptor( + u16 data, + u8 dwindex, + u8 startbit, + u8 bits) +{ + return (u32)(1UL << SDMA_AHG_UPDATE_ENABLE_SHIFT | + ((startbit & SDMA_AHG_FIELD_START_MASK) << + SDMA_AHG_FIELD_START_SHIFT) | + ((bits & SDMA_AHG_FIELD_LEN_MASK) << + SDMA_AHG_FIELD_LEN_SHIFT) | + ((dwindex & SDMA_AHG_INDEX_MASK) << + SDMA_AHG_INDEX_SHIFT) | + ((data & SDMA_AHG_VALUE_MASK) << + SDMA_AHG_VALUE_SHIFT)); +} +#ifdef __HFI1_ORIG__ + +/** + * sdma_progress - use seq number of detect head progress + * @sde: sdma_engine to check + * @seq: base seq count + * @tx: txreq for which we need to check descriptor availability + * + * This is used in the appropriate spot in the sleep routine + * to check for potential ring progress. This routine gets the + * seqcount before queuing the iowait structure for progress. + * + * If the seqcount indicates that progress needs to be checked, + * re-submission is detected by checking whether the descriptor + * queue has enough descriptor for the txreq. + */ +static inline unsigned sdma_progress(struct sdma_engine *sde, unsigned seq, + struct sdma_txreq *tx) +{ + if (read_seqretry(&sde->head_lock, seq)) { + sde->desc_avail = sdma_descq_freecnt(sde); + if (tx->num_desc > sde->desc_avail) + return 0; + return 1; + } + return 0; +} + +/** + * sdma_iowait_schedule() - initialize wait structure + * @sde: sdma_engine to schedule + * @wait: wait struct to schedule + * + * This function initializes the iowait + * structure embedded in the QP or PQ. + * + */ +static inline void sdma_iowait_schedule( + struct sdma_engine *sde, + struct iowait *wait) +{ + struct hfi1_pportdata *ppd = sde->dd->pport; + + iowait_schedule(wait, ppd->hfi1_wq, sde->cpu); +} + +/* for use by interrupt handling */ +void sdma_engine_error(struct sdma_engine *sde, u64 status); +void sdma_engine_interrupt(struct sdma_engine *sde, u64 status); + +/* + * + * The diagram below details the relationship of the mapping structures + * + * Since the mapping now allows for non-uniform engines per vl, the + * number of engines for a vl is either the vl_engines[vl] or + * a computation based on num_sdma/num_vls: + * + * For example: + * nactual = vl_engines ? vl_engines[vl] : num_sdma/num_vls + * + * n = roundup to next highest power of 2 using nactual + * + * In the case where there are num_sdma/num_vls doesn't divide + * evenly, the extras are added from the last vl downward. + * + * For the case where n > nactual, the engines are assigned + * in a round robin fashion wrapping back to the first engine + * for a particular vl. + * + * dd->sdma_map + * | sdma_map_elem[0] + * | +--------------------+ + * v | mask | + * sdma_vl_map |--------------------| + * +--------------------------+ | sde[0] -> eng 1 | + * | list (RCU) | |--------------------| + * |--------------------------| ->| sde[1] -> eng 2 | + * | mask | --/ |--------------------| + * |--------------------------| -/ | * | + * | actual_vls (max 8) | -/ |--------------------| + * |--------------------------| --/ | sde[n] -> eng n | + * | vls (max 8) | -/ +--------------------+ + * |--------------------------| --/ + * | map[0] |-/ + * |--------------------------| +--------------------+ + * | map[1] |--- | mask | + * |--------------------------| \---- |--------------------| + * | * | \-- | sde[0] -> eng 1+n | + * | * | \---- |--------------------| + * | * | \->| sde[1] -> eng 2+n | + * |--------------------------| |--------------------| + * | map[vls - 1] |- | * | + * +--------------------------+ \- |--------------------| + * \- | sde[m] -> eng m+n | + * \ +--------------------+ + * \- + * \ + * \- +--------------------+ + * \- | mask | + * \ |--------------------| + * \- | sde[0] -> eng 1+m+n| + * \- |--------------------| + * >| sde[1] -> eng 2+m+n| + * |--------------------| + * | * | + * |--------------------| + * | sde[o] -> eng o+m+n| + * +--------------------+ + * + */ + +/** + * struct sdma_map_elem - mapping for a vl + * @mask - selector mask + * @sde - array of engines for this vl + * + * The mask is used to "mod" the selector + * to produce index into the trailing + * array of sdes. + */ +struct sdma_map_elem { + u32 mask; + struct sdma_engine *sde[0]; +}; + +/** + * struct sdma_map_el - mapping for a vl + * @engine_to_vl - map of an engine to a vl + * @list - rcu head for free callback + * @mask - vl mask to "mod" the vl to produce an index to map array + * @actual_vls - number of vls + * @vls - number of vls rounded to next power of 2 + * @map - array of sdma_map_elem entries + * + * This is the parent mapping structure. The trailing + * members of the struct point to sdma_map_elem entries, which + * in turn point to an array of sde's for that vl. + */ +struct sdma_vl_map { + s8 engine_to_vl[TXE_NUM_SDMA_ENGINES]; + struct rcu_head list; + u32 mask; + u8 actual_vls; + u8 vls; + struct sdma_map_elem *map[0]; +}; + +int sdma_map_init( + struct hfi1_devdata *dd, + u8 port, + u8 num_vls, + u8 *vl_engines); + +/* slow path */ +void _sdma_engine_progress_schedule(struct sdma_engine *sde); + +/** + * sdma_engine_progress_schedule() - schedule progress on engine + * @sde: sdma_engine to schedule progress + * + * This is the fast path. + * + */ +static inline void sdma_engine_progress_schedule( + struct sdma_engine *sde) +{ + if (!sde || sdma_descq_inprocess(sde) < (sde->descq_cnt / 8)) + return; + _sdma_engine_progress_schedule(sde); +} + +struct sdma_engine *sdma_select_engine_sc( + struct hfi1_devdata *dd, + u32 selector, + u8 sc5); + +#endif /* __HFI1_ORIG__ */ +struct sdma_engine *sdma_select_engine_vl( + struct hfi1_devdata *dd, + u32 selector, + u8 vl); + +struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd, + u32 selector, u8 vl); +#ifdef __HFI1_ORIG__ + +ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf); +ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf, + size_t count); +int sdma_engine_get_vl(struct sdma_engine *sde); +void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *); +void sdma_seqfile_dump_cpu_list(struct seq_file *s, struct hfi1_devdata *dd, + unsigned long cpuid); + +#ifdef CONFIG_SDMA_VERBOSITY +void sdma_dumpstate(struct sdma_engine *); +#endif +static inline char *slashstrip(char *s) +{ + char *r = s; + + while (*s) + if (*s++ == '/') + r = s; + return r; +} + +u16 sdma_get_descq_cnt(void); + +extern uint mod_num_sdma; + +void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid); + +#endif /* __HFI1_ORIG__ */ +#endif diff --git a/kernel/sdma.c b/kernel/sdma.c new file mode 100644 index 00000000..62d8661f --- /dev/null +++ b/kernel/sdma.c @@ -0,0 +1,3470 @@ +/* + * Copyright(c) 2015, 2016 Intel Corporation. + * + * This file is provided under a dual BSD/GPLv2 license. When using or + * redistributing this file, you may do so under either license. + * + * GPL LICENSE SUMMARY + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License for more details. + * + * BSD LICENSE + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in + * the documentation and/or other materials provided with the + * distribution. + * - Neither the name of Intel Corporation nor the names of its + * contributors may be used to endorse or promote products derived + * from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#include +#include +#include +#include + +unsigned long hfi1_cap_mask = HFI1_CAP_MASK_DEFAULT; + +#ifdef __HFI1_ORIG__ + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "hfi.h" +#include "common.h" +#include "qp.h" +#include "sdma.h" +#include "iowait.h" +#include "trace.h" + +#endif /* __HFI1_ORIG__ */ + +/* must be a power of 2 >= 64 <= 32768 */ +#define SDMA_DESCQ_CNT 2048 +#define SDMA_DESC_INTR 64 +#define INVALID_TAIL 0xffff + +#ifdef __HFI1_ORIG__ + +static uint sdma_descq_cnt = SDMA_DESCQ_CNT; +module_param(sdma_descq_cnt, uint, S_IRUGO); +MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries"); + +static uint sdma_idle_cnt = 250; +module_param(sdma_idle_cnt, uint, S_IRUGO); +MODULE_PARM_DESC(sdma_idle_cnt, "sdma interrupt idle delay (ns,default 250)"); + +uint mod_num_sdma; +module_param_named(num_sdma, mod_num_sdma, uint, S_IRUGO); +MODULE_PARM_DESC(num_sdma, "Set max number SDMA engines to use"); + +static uint sdma_desct_intr = SDMA_DESC_INTR; +module_param_named(desct_intr, sdma_desct_intr, uint, S_IRUGO | S_IWUSR); +MODULE_PARM_DESC(desct_intr, "Number of SDMA descriptor before interrupt"); + +#define SDMA_WAIT_BATCH_SIZE 20 +/* max wait time for a SDMA engine to indicate it has halted */ +#define SDMA_ERR_HALT_TIMEOUT 10 /* ms */ +/* all SDMA engine errors that cause a halt */ + +#define SD(name) SEND_DMA_##name +#define ALL_SDMA_ENG_HALT_ERRS \ + (SD(ENG_ERR_STATUS_SDMA_WRONG_DW_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_GEN_MISMATCH_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_TOO_LONG_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_TAIL_OUT_OF_BOUNDS_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_FIRST_DESC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_MEM_READ_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_LENGTH_MISMATCH_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_PACKET_DESC_OVERFLOW_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_SELECT_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_ADDRESS_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_LENGTH_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_TIMEOUT_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_DESC_TABLE_UNC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_ASSEMBLY_UNC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_PACKET_TRACKING_UNC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_STORAGE_UNC_ERR_SMASK) \ + | SD(ENG_ERR_STATUS_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SMASK)) + +/* sdma_sendctrl operations */ +#define SDMA_SENDCTRL_OP_ENABLE BIT(0) +#define SDMA_SENDCTRL_OP_INTENABLE BIT(1) +#define SDMA_SENDCTRL_OP_HALT BIT(2) +#define SDMA_SENDCTRL_OP_CLEANUP BIT(3) + +/* handle long defines */ +#define SDMA_EGRESS_PACKET_OCCUPANCY_SMASK \ +SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SMASK +#define SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT \ +SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT + +static const char * const sdma_state_names[] = { + [sdma_state_s00_hw_down] = "s00_HwDown", + [sdma_state_s10_hw_start_up_halt_wait] = "s10_HwStartUpHaltWait", + [sdma_state_s15_hw_start_up_clean_wait] = "s15_HwStartUpCleanWait", + [sdma_state_s20_idle] = "s20_Idle", + [sdma_state_s30_sw_clean_up_wait] = "s30_SwCleanUpWait", + [sdma_state_s40_hw_clean_up_wait] = "s40_HwCleanUpWait", + [sdma_state_s50_hw_halt_wait] = "s50_HwHaltWait", + [sdma_state_s60_idle_halt_wait] = "s60_IdleHaltWait", + [sdma_state_s80_hw_freeze] = "s80_HwFreeze", + [sdma_state_s82_freeze_sw_clean] = "s82_FreezeSwClean", + [sdma_state_s99_running] = "s99_Running", +}; + +#ifdef CONFIG_SDMA_VERBOSITY +static const char * const sdma_event_names[] = { + [sdma_event_e00_go_hw_down] = "e00_GoHwDown", + [sdma_event_e10_go_hw_start] = "e10_GoHwStart", + [sdma_event_e15_hw_halt_done] = "e15_HwHaltDone", + [sdma_event_e25_hw_clean_up_done] = "e25_HwCleanUpDone", + [sdma_event_e30_go_running] = "e30_GoRunning", + [sdma_event_e40_sw_cleaned] = "e40_SwCleaned", + [sdma_event_e50_hw_cleaned] = "e50_HwCleaned", + [sdma_event_e60_hw_halted] = "e60_HwHalted", + [sdma_event_e70_go_idle] = "e70_GoIdle", + [sdma_event_e80_hw_freeze] = "e80_HwFreeze", + [sdma_event_e81_hw_frozen] = "e81_HwFrozen", + [sdma_event_e82_hw_unfreeze] = "e82_HwUnfreeze", + [sdma_event_e85_link_down] = "e85_LinkDown", + [sdma_event_e90_sw_halted] = "e90_SwHalted", +}; +#endif + +static const struct sdma_set_state_action sdma_action_table[] = { + [sdma_state_s00_hw_down] = { + .go_s99_running_tofalse = 1, + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s10_hw_start_up_halt_wait] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 1, + .op_cleanup = 0, + }, + [sdma_state_s15_hw_start_up_clean_wait] = { + .op_enable = 0, + .op_intenable = 1, + .op_halt = 0, + .op_cleanup = 1, + }, + [sdma_state_s20_idle] = { + .op_enable = 0, + .op_intenable = 1, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s30_sw_clean_up_wait] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s40_hw_clean_up_wait] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 1, + }, + [sdma_state_s50_hw_halt_wait] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s60_idle_halt_wait] = { + .go_s99_running_tofalse = 1, + .op_enable = 0, + .op_intenable = 0, + .op_halt = 1, + .op_cleanup = 0, + }, + [sdma_state_s80_hw_freeze] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s82_freeze_sw_clean] = { + .op_enable = 0, + .op_intenable = 0, + .op_halt = 0, + .op_cleanup = 0, + }, + [sdma_state_s99_running] = { + .op_enable = 1, + .op_intenable = 1, + .op_halt = 0, + .op_cleanup = 0, + .go_s99_running_totrue = 1, + }, +}; + +#endif /* __HFI1_ORIG__ */ +#define SDMA_TAIL_UPDATE_THRESH 0x1F +#ifdef __HFI1_ORIG__ + +/* declare all statics here rather than keep sorting */ +static void sdma_complete(struct kref *); +static void sdma_finalput(struct sdma_state *); +static void sdma_get(struct sdma_state *); +static void sdma_hw_clean_up_task(unsigned long); +static void sdma_put(struct sdma_state *); +static void sdma_set_state(struct sdma_engine *, enum sdma_states); +static void sdma_start_hw_clean_up(struct sdma_engine *); +static void sdma_sw_clean_up_task(unsigned long); +static void sdma_sendctrl(struct sdma_engine *, unsigned); +static void init_sdma_regs(struct sdma_engine *, u32, uint); +static void sdma_process_event( + struct sdma_engine *sde, + enum sdma_events event); +static void __sdma_process_event( + struct sdma_engine *sde, + enum sdma_events event); +static void dump_sdma_state(struct sdma_engine *sde); +static void sdma_make_progress(struct sdma_engine *sde, u64 status); +static void sdma_desc_avail(struct sdma_engine *sde, unsigned avail); +static void sdma_flush_descq(struct sdma_engine *sde); + +/** + * sdma_state_name() - return state string from enum + * @state: state + */ +static const char *sdma_state_name(enum sdma_states state) +{ + return sdma_state_names[state]; +} + +static void sdma_get(struct sdma_state *ss) +{ + kref_get(&ss->kref); +} + +static void sdma_complete(struct kref *kref) +{ + struct sdma_state *ss = + container_of(kref, struct sdma_state, kref); + + complete(&ss->comp); +} + +static void sdma_put(struct sdma_state *ss) +{ + kref_put(&ss->kref, sdma_complete); +} + +static void sdma_finalput(struct sdma_state *ss) +{ + sdma_put(ss); + wait_for_completion(&ss->comp); +} + +static inline void write_sde_csr( + struct sdma_engine *sde, + u32 offset0, + u64 value) +{ + write_kctxt_csr(sde->dd, sde->this_idx, offset0, value); +} + +static inline u64 read_sde_csr( + struct sdma_engine *sde, + u32 offset0) +{ + return read_kctxt_csr(sde->dd, sde->this_idx, offset0); +} + +/* + * sdma_wait_for_packet_egress() - wait for the VL FIFO occupancy for + * sdma engine 'sde' to drop to 0. + */ +static void sdma_wait_for_packet_egress(struct sdma_engine *sde, + int pause) +{ + u64 off = 8 * sde->this_idx; + struct hfi1_devdata *dd = sde->dd; + int lcnt = 0; + u64 reg_prev; + u64 reg = 0; + + while (1) { + reg_prev = reg; + reg = read_csr(dd, off + SEND_EGRESS_SEND_DMA_STATUS); + + reg &= SDMA_EGRESS_PACKET_OCCUPANCY_SMASK; + reg >>= SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT; + if (reg == 0) + break; + /* counter is reest if accupancy count changes */ + if (reg != reg_prev) + lcnt = 0; + if (lcnt++ > 500) { + /* timed out - bounce the link */ + dd_dev_err(dd, "%s: engine %u timeout waiting for packets to egress, remaining count %u, bouncing link\n", + __func__, sde->this_idx, (u32)reg); + queue_work(dd->pport->hfi1_wq, + &dd->pport->link_bounce_work); + break; + } + udelay(1); + } +} + +/* + * sdma_wait() - wait for packet egress to complete for all SDMA engines, + * and pause for credit return. + */ +void sdma_wait(struct hfi1_devdata *dd) +{ + int i; + + for (i = 0; i < dd->num_sdma; i++) { + struct sdma_engine *sde = &dd->per_sdma[i]; + + sdma_wait_for_packet_egress(sde, 0); + } +} + +static inline void sdma_set_desc_cnt(struct sdma_engine *sde, unsigned cnt) +{ + u64 reg; + + if (!(sde->dd->flags & HFI1_HAS_SDMA_TIMEOUT)) + return; + reg = cnt; + reg &= SD(DESC_CNT_CNT_MASK); + reg <<= SD(DESC_CNT_CNT_SHIFT); + write_sde_csr(sde, SD(DESC_CNT), reg); +} + +static inline void complete_tx(struct sdma_engine *sde, + struct sdma_txreq *tx, + int res) +{ + /* protect against complete modifying */ + struct iowait *wait = tx->wait; + callback_t complete = tx->complete; + hfi1_cdbg(AIOWRITE, "+"); +#ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER + trace_hfi1_sdma_out_sn(sde, tx->sn); + if (WARN_ON_ONCE(sde->head_sn != tx->sn)) + dd_dev_err(sde->dd, "expected %llu got %llu\n", + sde->head_sn, tx->sn); + sde->head_sn++; +#endif + __sdma_txclean(sde->dd, tx); + if (complete) + (*complete)(tx, res); + if (iowait_sdma_dec(wait)) + iowait_drain_wakeup(wait); + hfi1_cdbg(AIOWRITE, "-"); +} + +/* + * Complete all the sdma requests with a SDMA_TXREQ_S_ABORTED status + * + * Depending on timing there can be txreqs in two places: + * - in the descq ring + * - in the flush list + * + * To avoid ordering issues the descq ring needs to be flushed + * first followed by the flush list. + * + * This routine is called from two places + * - From a work queue item + * - Directly from the state machine just before setting the + * state to running + * + * Must be called with head_lock held + * + */ +static void sdma_flush(struct sdma_engine *sde) +{ + struct sdma_txreq *txp, *txp_next; + LIST_HEAD(flushlist); + unsigned long flags; + + /* flush from head to tail */ + sdma_flush_descq(sde); + spin_lock_irqsave(&sde->flushlist_lock, flags); + /* copy flush list */ + list_for_each_entry_safe(txp, txp_next, &sde->flushlist, list) { + list_del_init(&txp->list); + list_add_tail(&txp->list, &flushlist); + } + spin_unlock_irqrestore(&sde->flushlist_lock, flags); + /* flush from flush list */ + list_for_each_entry_safe(txp, txp_next, &flushlist, list) + complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED); +} + +/* + * Fields a work request for flushing the descq ring + * and the flush list + * + * If the engine has been brought to running during + * the scheduling delay, the flush is ignored, assuming + * that the process of bringing the engine to running + * would have done this flush prior to going to running. + * + */ +static void sdma_field_flush(struct work_struct *work) +{ + unsigned long flags; + struct sdma_engine *sde = + container_of(work, struct sdma_engine, flush_worker); + + write_seqlock_irqsave(&sde->head_lock, flags); + if (!__sdma_running(sde)) + sdma_flush(sde); + write_sequnlock_irqrestore(&sde->head_lock, flags); +} + +static void sdma_err_halt_wait(struct work_struct *work) +{ + struct sdma_engine *sde = container_of(work, struct sdma_engine, + err_halt_worker); + u64 statuscsr; + unsigned long timeout; + + timeout = jiffies + msecs_to_jiffies(SDMA_ERR_HALT_TIMEOUT); + while (1) { + statuscsr = read_sde_csr(sde, SD(STATUS)); + statuscsr &= SD(STATUS_ENG_HALTED_SMASK); + if (statuscsr) + break; + if (time_after(jiffies, timeout)) { + dd_dev_err(sde->dd, + "SDMA engine %d - timeout waiting for engine to halt\n", + sde->this_idx); + /* + * Continue anyway. This could happen if there was + * an uncorrectable error in the wrong spot. + */ + break; + } + usleep_range(80, 120); + } + + sdma_process_event(sde, sdma_event_e15_hw_halt_done); +} + +static void sdma_err_progress_check_schedule(struct sdma_engine *sde) +{ + if (!is_bx(sde->dd) && HFI1_CAP_IS_KSET(SDMA_AHG)) { + unsigned index; + struct hfi1_devdata *dd = sde->dd; + + for (index = 0; index < dd->num_sdma; index++) { + struct sdma_engine *curr_sdma = &dd->per_sdma[index]; + + if (curr_sdma != sde) + curr_sdma->progress_check_head = + curr_sdma->descq_head; + } + dd_dev_err(sde->dd, + "SDMA engine %d - check scheduled\n", + sde->this_idx); + mod_timer(&sde->err_progress_check_timer, jiffies + 10); + } +} + +static void sdma_err_progress_check(unsigned long data) +{ + unsigned index; + struct sdma_engine *sde = (struct sdma_engine *)data; + + dd_dev_err(sde->dd, "SDE progress check event\n"); + for (index = 0; index < sde->dd->num_sdma; index++) { + struct sdma_engine *curr_sde = &sde->dd->per_sdma[index]; + unsigned long flags; + + /* check progress on each engine except the current one */ + if (curr_sde == sde) + continue; + /* + * We must lock interrupts when acquiring sde->lock, + * to avoid a deadlock if interrupt triggers and spins on + * the same lock on same CPU + */ + spin_lock_irqsave(&curr_sde->tail_lock, flags); + write_seqlock(&curr_sde->head_lock); + + /* skip non-running queues */ + if (curr_sde->state.current_state != sdma_state_s99_running) { + write_sequnlock(&curr_sde->head_lock); + spin_unlock_irqrestore(&curr_sde->tail_lock, flags); + continue; + } + + if ((curr_sde->descq_head != curr_sde->descq_tail) && + (curr_sde->descq_head == + curr_sde->progress_check_head)) + __sdma_process_event(curr_sde, + sdma_event_e90_sw_halted); + write_sequnlock(&curr_sde->head_lock); + spin_unlock_irqrestore(&curr_sde->tail_lock, flags); + } + schedule_work(&sde->err_halt_worker); +} + +static void sdma_hw_clean_up_task(unsigned long opaque) +{ + struct sdma_engine *sde = (struct sdma_engine *)opaque; + u64 statuscsr; + + while (1) { +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", + sde->this_idx, slashstrip(__FILE__), __LINE__, + __func__); +#endif + statuscsr = read_sde_csr(sde, SD(STATUS)); + statuscsr &= SD(STATUS_ENG_CLEANED_UP_SMASK); + if (statuscsr) + break; + udelay(10); + } + + sdma_process_event(sde, sdma_event_e25_hw_clean_up_done); +} + +static inline struct sdma_txreq *get_txhead(struct sdma_engine *sde) +{ + smp_read_barrier_depends(); /* see sdma_update_tail() */ + return sde->tx_ring[sde->tx_head & sde->sdma_mask]; +} + +/* + * flush ring for recovery + */ +static void sdma_flush_descq(struct sdma_engine *sde) +{ + u16 head, tail; + int progress = 0; + struct sdma_txreq *txp = get_txhead(sde); + + /* The reason for some of the complexity of this code is that + * not all descriptors have corresponding txps. So, we have to + * be able to skip over descs until we wander into the range of + * the next txp on the list. + */ + head = sde->descq_head & sde->sdma_mask; + tail = sde->descq_tail & sde->sdma_mask; + while (head != tail) { + /* advance head, wrap if needed */ + head = ++sde->descq_head & sde->sdma_mask; + /* if now past this txp's descs, do the callback */ + if (txp && txp->next_descq_idx == head) { + /* remove from list */ + sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL; + complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED); + trace_hfi1_sdma_progress(sde, head, tail, txp); + txp = get_txhead(sde); + } + progress++; + } + if (progress) + sdma_desc_avail(sde, sdma_descq_freecnt(sde)); +} + +static void sdma_sw_clean_up_task(unsigned long opaque) +{ + struct sdma_engine *sde = (struct sdma_engine *)opaque; + unsigned long flags; + + spin_lock_irqsave(&sde->tail_lock, flags); + write_seqlock(&sde->head_lock); + + /* + * At this point, the following should always be true: + * - We are halted, so no more descriptors are getting retired. + * - We are not running, so no one is submitting new work. + * - Only we can send the e40_sw_cleaned, so we can't start + * running again until we say so. So, the active list and + * descq are ours to play with. + */ + + /* + * In the error clean up sequence, software clean must be called + * before the hardware clean so we can use the hardware head in + * the progress routine. A hardware clean or SPC unfreeze will + * reset the hardware head. + * + * Process all retired requests. The progress routine will use the + * latest physical hardware head - we are not running so speed does + * not matter. + */ + sdma_make_progress(sde, 0); + + sdma_flush(sde); + + /* + * Reset our notion of head and tail. + * Note that the HW registers have been reset via an earlier + * clean up. + */ + sde->descq_tail = 0; + sde->descq_head = 0; + sde->desc_avail = sdma_descq_freecnt(sde); + *sde->head_dma = 0; + + __sdma_process_event(sde, sdma_event_e40_sw_cleaned); + + write_sequnlock(&sde->head_lock); + spin_unlock_irqrestore(&sde->tail_lock, flags); +} + +static void sdma_sw_tear_down(struct sdma_engine *sde) +{ + struct sdma_state *ss = &sde->state; + + /* Releasing this reference means the state machine has stopped. */ + sdma_put(ss); + + /* stop waiting for all unfreeze events to complete */ + atomic_set(&sde->dd->sdma_unfreeze_count, -1); + wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); +} + +static void sdma_start_hw_clean_up(struct sdma_engine *sde) +{ + tasklet_hi_schedule(&sde->sdma_hw_clean_up_task); +} + +static void sdma_set_state(struct sdma_engine *sde, + enum sdma_states next_state) +{ + struct sdma_state *ss = &sde->state; + const struct sdma_set_state_action *action = sdma_action_table; + unsigned op = 0; + + trace_hfi1_sdma_state( + sde, + sdma_state_names[ss->current_state], + sdma_state_names[next_state]); + + /* debugging bookkeeping */ + ss->previous_state = ss->current_state; + ss->previous_op = ss->current_op; + ss->current_state = next_state; + + if (ss->previous_state != sdma_state_s99_running && + next_state == sdma_state_s99_running) + sdma_flush(sde); + + if (action[next_state].op_enable) + op |= SDMA_SENDCTRL_OP_ENABLE; + + if (action[next_state].op_intenable) + op |= SDMA_SENDCTRL_OP_INTENABLE; + + if (action[next_state].op_halt) + op |= SDMA_SENDCTRL_OP_HALT; + + if (action[next_state].op_cleanup) + op |= SDMA_SENDCTRL_OP_CLEANUP; + + if (action[next_state].go_s99_running_tofalse) + ss->go_s99_running = 0; + + if (action[next_state].go_s99_running_totrue) + ss->go_s99_running = 1; + + ss->current_op = op; + sdma_sendctrl(sde, ss->current_op); +} + +/** + * sdma_get_descq_cnt() - called when device probed + * + * Return a validated descq count. + * + * This is currently only used in the verbs initialization to build the tx + * list. + * + * This will probably be deleted in favor of a more scalable approach to + * alloc tx's. + * + */ +u16 sdma_get_descq_cnt(void) +{ + u16 count = sdma_descq_cnt; + + if (!count) + return SDMA_DESCQ_CNT; + /* count must be a power of 2 greater than 64 and less than + * 32768. Otherwise return default. + */ + if (!is_power_of_2(count)) + return SDMA_DESCQ_CNT; + if (count < 64 || count > 32768) + return SDMA_DESCQ_CNT; + return count; +} + +/** + * sdma_engine_get_vl() - return vl for a given sdma engine + * @sde: sdma engine + * + * This function returns the vl mapped to a given engine, or an error if + * the mapping can't be found. The mapping fields are protected by RCU. + */ +int sdma_engine_get_vl(struct sdma_engine *sde) +{ + struct hfi1_devdata *dd = sde->dd; + struct sdma_vl_map *m; + u8 vl; + + if (sde->this_idx >= TXE_NUM_SDMA_ENGINES) + return -EINVAL; + + rcu_read_lock(); + m = rcu_dereference(dd->sdma_map); + if (unlikely(!m)) { + rcu_read_unlock(); + return -EINVAL; + } + vl = m->engine_to_vl[sde->this_idx]; + rcu_read_unlock(); + + return vl; +} + +/** + * sdma_select_engine_vl() - select sdma engine + * @dd: devdata + * @selector: a spreading factor + * @vl: this vl + * + * + * This function returns an engine based on the selector and a vl. The + * mapping fields are protected by RCU. + */ +struct sdma_engine *sdma_select_engine_vl( + struct hfi1_devdata *dd, + u32 selector, + u8 vl) +{ + struct sdma_vl_map *m; + struct sdma_map_elem *e; + struct sdma_engine *rval; + + hfi1_cdbg(AIOWRITE, "+"); + /* NOTE This should only happen if SC->VL changed after the initial + * checks on the QP/AH + * Default will return engine 0 below + */ +#ifdef __HFI1_ORIG__ + if (vl >= num_vls) { +#else + if (vl >= HFI1_MAX_VLS_SUPPORTED) { +#endif /* __HFI1_ORIG__ */ + rval = NULL; + goto done; + } + + rcu_read_lock(); + m = rcu_dereference(dd->sdma_map); + if (unlikely(!m)) { + rcu_read_unlock(); + return &dd->per_sdma[0]; + } + e = m->map[vl & m->mask]; + rval = e->sde[selector & e->mask]; + rcu_read_unlock(); + +done: + rval = !rval ? &dd->per_sdma[0] : rval; + trace_hfi1_sdma_engine_select(dd, selector, vl, rval->this_idx); + hfi1_cdbg(AIOWRITE, "-"); + return rval; +} + +/** + * sdma_select_engine_sc() - select sdma engine + * @dd: devdata + * @selector: a spreading factor + * @sc5: the 5 bit sc + * + * + * This function returns an engine based on the selector and an sc. + */ +struct sdma_engine *sdma_select_engine_sc( + struct hfi1_devdata *dd, + u32 selector, + u8 sc5) +{ + u8 vl = sc_to_vlt(dd, sc5); + + return sdma_select_engine_vl(dd, selector, vl); +} + +struct sdma_rht_map_elem { + u32 mask; + u8 ctr; + struct sdma_engine *sde[0]; +}; + +struct sdma_rht_node { + unsigned long cpu_id; + struct sdma_rht_map_elem *map[HFI1_MAX_VLS_SUPPORTED]; + struct rhash_head node; +}; + +#define NR_CPUS_HINT 192 + +static int info_mutex_is_held(void) +{ + return 1; +} + +static struct rhashtable_params sdma_rht_params = { + .nelem_hint = NR_CPUS_HINT, + .head_offset = offsetof(struct sdma_rht_node, node), + .key_offset = offsetof(struct sdma_rht_node, cpu_id), + .key_len = sizeof(unsigned long), + .hashfn = jhash, + .mutex_is_held = info_mutex_is_held, +}; + +/* + * sdma_select_user_engine() - select sdma engine based on user setup + * @dd: devdata + * @selector: a spreading factor + * @vl: this vl + * + * This function returns an sdma engine for a user sdma request. + * User defined sdma engine affinity setting is honored when applicable, + * otherwise system default sdma engine mapping is used. To ensure correct + * ordering, the mapping from to sde must remain unchanged. + */ +struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd, + u32 selector, u8 vl) +{ + struct sdma_rht_node *rht_node; + struct sdma_engine *sde = NULL; + const struct cpumask *current_mask = tsk_cpus_allowed(current); + unsigned long cpu_id; + + hfi1_cdbg(AIOWRITE, "+"); + /* + * To ensure that always the same sdma engine(s) will be + * selected make sure the process is pinned to this CPU only. + */ + if (cpumask_weight(current_mask) != 1) + goto out; + + cpu_id = smp_processor_id(); + rcu_read_lock(); + rht_node = rhashtable_lookup(&dd->sdma_rht, &cpu_id); + + if (rht_node && rht_node->map[vl]) { + struct sdma_rht_map_elem *map = rht_node->map[vl]; + + sde = map->sde[selector & map->mask]; + } + rcu_read_unlock(); + + if (sde) + return sde; + +out: + return sdma_select_engine_vl(dd, selector, vl); +} + +static void sdma_populate_sde_map(struct sdma_rht_map_elem *map) +{ + int i; + + for (i = 0; i < roundup_pow_of_two(map->ctr ? : 1) - map->ctr; i++) + map->sde[map->ctr + i] = map->sde[i]; +} + +static void sdma_cleanup_sde_map(struct sdma_rht_map_elem *map, + struct sdma_engine *sde) +{ + unsigned int i, pow; + + /* only need to check the first ctr entries for a match */ + for (i = 0; i < map->ctr; i++) { + if (map->sde[i] == sde) { + memmove(&map->sde[i], &map->sde[i + 1], + (map->ctr - i - 1) * sizeof(map->sde[0])); + map->ctr--; + pow = roundup_pow_of_two(map->ctr ? : 1); + map->mask = pow - 1; + sdma_populate_sde_map(map); + break; + } + } +} + +/* + * Prevents concurrent reads and writes of the sdma engine cpu_mask + */ +static DEFINE_MUTEX(process_to_sde_mutex); + +ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf, + size_t count) +{ + struct hfi1_devdata *dd = sde->dd; + cpumask_var_t mask, new_mask; + unsigned long cpu; + int ret, vl, sz; + + vl = sdma_engine_get_vl(sde); + if (unlikely(vl < 0)) + return -EINVAL; + + ret = zalloc_cpumask_var(&mask, GFP_KERNEL); + if (!ret) + return -ENOMEM; + + ret = zalloc_cpumask_var(&new_mask, GFP_KERNEL); + if (!ret) { + free_cpumask_var(mask); + return -ENOMEM; + } + ret = cpulist_parse(buf, mask); + if (ret) + goto out_free; + + if (!cpumask_subset(mask, cpu_online_mask)) { + dd_dev_warn(sde->dd, "Invalid CPU mask\n"); + ret = -EINVAL; + goto out_free; + } + + sz = sizeof(struct sdma_rht_map_elem) + + (TXE_NUM_SDMA_ENGINES * sizeof(struct sdma_engine *)); + + mutex_lock(&process_to_sde_mutex); + + for_each_cpu(cpu, mask) { + struct sdma_rht_node *rht_node; + + /* Check if we have this already mapped */ + if (cpumask_test_cpu(cpu, &sde->cpu_mask)) { + cpumask_set_cpu(cpu, new_mask); + continue; + } + + rht_node = rhashtable_lookup(&dd->sdma_rht, &cpu); + if (!rht_node) { + rht_node = kzalloc(sizeof(*rht_node), GFP_KERNEL); + if (!rht_node) { + ret = -ENOMEM; + goto out; + } + + if (vl >= ARRAY_SIZE(rht_node->map)) { + ret = -EINVAL; + goto out; + } + + rht_node->map[vl] = kzalloc(sz, GFP_KERNEL); + if (!rht_node->map[vl]) { + kfree(rht_node); + ret = -ENOMEM; + goto out; + } + rht_node->cpu_id = cpu; + rht_node->map[vl]->mask = 0; + rht_node->map[vl]->ctr = 1; + rht_node->map[vl]->sde[0] = sde; + + rhashtable_insert(&dd->sdma_rht, + &rht_node->node, GFP_KERNEL); + } else { + int ctr, pow; + + /* Add new user mappings */ + if (!rht_node->map[vl]) + rht_node->map[vl] = kzalloc(sz, GFP_KERNEL); + + if (!rht_node->map[vl]) { + ret = -ENOMEM; + goto out; + } + + rht_node->map[vl]->ctr++; + ctr = rht_node->map[vl]->ctr; + rht_node->map[vl]->sde[ctr - 1] = sde; + pow = roundup_pow_of_two(ctr); + rht_node->map[vl]->mask = pow - 1; + + /* Populate the sde map table */ + sdma_populate_sde_map(rht_node->map[vl]); + } + cpumask_set_cpu(cpu, new_mask); + } + + /* Clean up old mappings */ + for_each_cpu(cpu, cpu_online_mask) { + struct sdma_rht_node *rht_node; + + /* Don't cleanup sdes that are set in the new mask */ + if (cpumask_test_cpu(cpu, mask)) + continue; + + rht_node = rhashtable_lookup(&dd->sdma_rht, &cpu); + if (rht_node) { + bool empty = true; + int i; + + /* Remove mappings for old sde */ + for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) + if (rht_node->map[i]) + sdma_cleanup_sde_map(rht_node->map[i], + sde); + + /* Free empty hash table entries */ + for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) { + if (!rht_node->map[i]) + continue; + + if (rht_node->map[i]->ctr) { + empty = false; + break; + } + } + + if (empty) { + ret = rhashtable_remove(&dd->sdma_rht, + &rht_node->node, + GFP_KERNEL); + /* + * The return check WARN_ON is inverted for + * RHEL 7.2 backport since the return value + * from rhashtable_remove() is different from + * rhashtable_remove_fast() upstream, + * rhashtable_remove() returns bool true on + * success, while rhashtable_remove_fast() + * returns 0 on success. + */ + WARN_ON(!ret); + + for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) + kfree(rht_node->map[i]); + + kfree(rht_node); + } + } + } + + cpumask_copy(&sde->cpu_mask, new_mask); +out: + mutex_unlock(&process_to_sde_mutex); +out_free: + free_cpumask_var(mask); + free_cpumask_var(new_mask); + return ret ? : strnlen(buf, PAGE_SIZE); +} + +ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf) +{ + mutex_lock(&process_to_sde_mutex); + if (cpumask_empty(&sde->cpu_mask)) { + snprintf(buf, PAGE_SIZE, "%s\n", "empty"); + } else { + int n = cpulist_scnprintf(buf, PAGE_SIZE - 2, &sde->cpu_mask); + + buf[n++] = '\n'; + buf[n] = '\0'; + } + mutex_unlock(&process_to_sde_mutex); + return strnlen(buf, PAGE_SIZE); +} + +static void sdma_rht_free(void *ptr, void *arg) +{ + struct sdma_rht_node *rht_node = ptr; + int i; + + for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) + kfree(rht_node->map[i]); + + kfree(rht_node); +} + +/** + * sdma_seqfile_dump_cpu_list() - debugfs dump the cpu to sdma mappings + * @s: seq file + * @dd: hfi1_devdata + * @cpuid: cpu id + * + * This routine dumps the process to sde mappings per cpu + */ +void sdma_seqfile_dump_cpu_list(struct seq_file *s, + struct hfi1_devdata *dd, + unsigned long cpuid) +{ + struct sdma_rht_node *rht_node; + int i, j; + + rht_node = rhashtable_lookup(&dd->sdma_rht, &cpuid); + if (!rht_node) + return; + + seq_printf(s, "cpu%3lu: ", cpuid); + for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) { + if (!rht_node->map[i] || !rht_node->map[i]->ctr) + continue; + + seq_printf(s, " vl%d: [", i); + + for (j = 0; j < rht_node->map[i]->ctr; j++) { + if (!rht_node->map[i]->sde[j]) + continue; + + if (j > 0) + seq_puts(s, ","); + + seq_printf(s, " sdma%2d", + rht_node->map[i]->sde[j]->this_idx); + } + seq_puts(s, " ]"); + } + + seq_puts(s, "\n"); +} + +/* + * Free the indicated map struct + */ +static void sdma_map_free(struct sdma_vl_map *m) +{ + int i; + + for (i = 0; m && i < m->actual_vls; i++) + kfree(m->map[i]); + kfree(m); +} + +/* + * Handle RCU callback + */ +static void sdma_map_rcu_callback(struct rcu_head *list) +{ + struct sdma_vl_map *m = container_of(list, struct sdma_vl_map, list); + + sdma_map_free(m); +} + +/** + * sdma_map_init - called when # vls change + * @dd: hfi1_devdata + * @port: port number + * @num_vls: number of vls + * @vl_engines: per vl engine mapping (optional) + * + * This routine changes the mapping based on the number of vls. + * + * vl_engines is used to specify a non-uniform vl/engine loading. NULL + * implies auto computing the loading and giving each VLs a uniform + * distribution of engines per VL. + * + * The auto algorithm computes the sde_per_vl and the number of extra + * engines. Any extra engines are added from the last VL on down. + * + * rcu locking is used here to control access to the mapping fields. + * + * If either the num_vls or num_sdma are non-power of 2, the array sizes + * in the struct sdma_vl_map and the struct sdma_map_elem are rounded + * up to the next highest power of 2 and the first entry is reused + * in a round robin fashion. + * + * If an error occurs the map change is not done and the mapping is + * not changed. + * + */ +int sdma_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_engines) +{ + int i, j; + int extra, sde_per_vl; + int engine = 0; + u8 lvl_engines[OPA_MAX_VLS]; + struct sdma_vl_map *oldmap, *newmap; + + if (!(dd->flags & HFI1_HAS_SEND_DMA)) + return 0; + + if (!vl_engines) { + /* truncate divide */ + sde_per_vl = dd->num_sdma / num_vls; + /* extras */ + extra = dd->num_sdma % num_vls; + vl_engines = lvl_engines; + /* add extras from last vl down */ + for (i = num_vls - 1; i >= 0; i--, extra--) + vl_engines[i] = sde_per_vl + (extra > 0 ? 1 : 0); + } + /* build new map */ + newmap = kzalloc( + sizeof(struct sdma_vl_map) + + roundup_pow_of_two(num_vls) * + sizeof(struct sdma_map_elem *), + GFP_KERNEL); + if (!newmap) + goto bail; + newmap->actual_vls = num_vls; + newmap->vls = roundup_pow_of_two(num_vls); + newmap->mask = (1 << ilog2(newmap->vls)) - 1; + /* initialize back-map */ + for (i = 0; i < TXE_NUM_SDMA_ENGINES; i++) + newmap->engine_to_vl[i] = -1; + for (i = 0; i < newmap->vls; i++) { + /* save for wrap around */ + int first_engine = engine; + + if (i < newmap->actual_vls) { + int sz = roundup_pow_of_two(vl_engines[i]); + + /* only allocate once */ + newmap->map[i] = kzalloc( + sizeof(struct sdma_map_elem) + + sz * sizeof(struct sdma_engine *), + GFP_KERNEL); + if (!newmap->map[i]) + goto bail; + newmap->map[i]->mask = (1 << ilog2(sz)) - 1; + /* assign engines */ + for (j = 0; j < sz; j++) { + newmap->map[i]->sde[j] = + &dd->per_sdma[engine]; + if (++engine >= first_engine + vl_engines[i]) + /* wrap back to first engine */ + engine = first_engine; + } + /* assign back-map */ + for (j = 0; j < vl_engines[i]; j++) + newmap->engine_to_vl[first_engine + j] = i; + } else { + /* just re-use entry without allocating */ + newmap->map[i] = newmap->map[i % num_vls]; + } + engine = first_engine + vl_engines[i]; + } + /* newmap in hand, save old map */ + spin_lock_irq(&dd->sde_map_lock); + oldmap = rcu_dereference_protected(dd->sdma_map, + lockdep_is_held(&dd->sde_map_lock)); + + /* publish newmap */ + rcu_assign_pointer(dd->sdma_map, newmap); + + spin_unlock_irq(&dd->sde_map_lock); + /* success, free any old map after grace period */ + if (oldmap) + call_rcu(&oldmap->list, sdma_map_rcu_callback); + return 0; +bail: + /* free any partial allocation */ + sdma_map_free(newmap); + return -ENOMEM; +} + +/* + * Clean up allocated memory. + * + * This routine is can be called regardless of the success of sdma_init() + * + */ +static void sdma_clean(struct hfi1_devdata *dd, size_t num_engines) +{ + size_t i; + struct sdma_engine *sde; + + if (dd->sdma_pad_dma) { + dma_free_coherent(&dd->pcidev->dev, 4, + (void *)dd->sdma_pad_dma, + dd->sdma_pad_phys); + dd->sdma_pad_dma = NULL; + dd->sdma_pad_phys = 0; + } + if (dd->sdma_heads_dma) { + dma_free_coherent(&dd->pcidev->dev, dd->sdma_heads_size, + (void *)dd->sdma_heads_dma, + dd->sdma_heads_phys); + dd->sdma_heads_dma = NULL; + dd->sdma_heads_phys = 0; + } + for (i = 0; dd->per_sdma && i < num_engines; ++i) { + sde = &dd->per_sdma[i]; + + sde->head_dma = NULL; + sde->head_phys = 0; + + if (sde->descq) { + dma_free_coherent( + &dd->pcidev->dev, + sde->descq_cnt * sizeof(u64[2]), + sde->descq, + sde->descq_phys + ); + sde->descq = NULL; + sde->descq_phys = 0; + } + kvfree(sde->tx_ring); + sde->tx_ring = NULL; + } + spin_lock_irq(&dd->sde_map_lock); + sdma_map_free(rcu_access_pointer(dd->sdma_map)); + RCU_INIT_POINTER(dd->sdma_map, NULL); + spin_unlock_irq(&dd->sde_map_lock); + synchronize_rcu(); + kfree(dd->per_sdma); + dd->per_sdma = NULL; +} + +/** + * sdma_init() - called when device probed + * @dd: hfi1_devdata + * @port: port number (currently only zero) + * + * sdma_init initializes the specified number of engines. + * + * The code initializes each sde, its csrs. Interrupts + * are not required to be enabled. + * + * Returns: + * 0 - success, -errno on failure + */ +int sdma_init(struct hfi1_devdata *dd, u8 port) +{ + unsigned this_idx; + struct sdma_engine *sde; + u16 descq_cnt; + void *curr_head; + struct hfi1_pportdata *ppd = dd->pport + port; + u32 per_sdma_credits; + uint idle_cnt = sdma_idle_cnt; + size_t num_engines = dd->chip_sdma_engines; + + if (!HFI1_CAP_IS_KSET(SDMA)) { + HFI1_CAP_CLEAR(SDMA_AHG); + return 0; + } + if (mod_num_sdma && + /* can't exceed chip support */ + mod_num_sdma <= dd->chip_sdma_engines && + /* count must be >= vls */ + mod_num_sdma >= num_vls) + num_engines = mod_num_sdma; + + dd_dev_info(dd, "SDMA mod_num_sdma: %u\n", mod_num_sdma); + dd_dev_info(dd, "SDMA chip_sdma_engines: %u\n", dd->chip_sdma_engines); + dd_dev_info(dd, "SDMA chip_sdma_mem_size: %u\n", + dd->chip_sdma_mem_size); + + per_sdma_credits = + dd->chip_sdma_mem_size / (num_engines * SDMA_BLOCK_SIZE); + + /* set up freeze waitqueue */ + init_waitqueue_head(&dd->sdma_unfreeze_wq); + atomic_set(&dd->sdma_unfreeze_count, 0); + + descq_cnt = sdma_get_descq_cnt(); + dd_dev_info(dd, "SDMA engines %zu descq_cnt %u\n", + num_engines, descq_cnt); + + /* alloc memory for array of send engines */ + dd->per_sdma = kcalloc(num_engines, sizeof(*dd->per_sdma), GFP_KERNEL); + if (!dd->per_sdma) + return -ENOMEM; + + idle_cnt = ns_to_cclock(dd, idle_cnt); + if (!sdma_desct_intr) + sdma_desct_intr = SDMA_DESC_INTR; + + /* Allocate memory for SendDMA descriptor FIFOs */ + for (this_idx = 0; this_idx < num_engines; ++this_idx) { + sde = &dd->per_sdma[this_idx]; + sde->dd = dd; + sde->ppd = ppd; + sde->this_idx = this_idx; + sde->descq_cnt = descq_cnt; + sde->desc_avail = sdma_descq_freecnt(sde); + sde->sdma_shift = ilog2(descq_cnt); + sde->sdma_mask = (1 << sde->sdma_shift) - 1; + + /* Create a mask specifically for each interrupt source */ + sde->int_mask = (u64)1 << (0 * TXE_NUM_SDMA_ENGINES + + this_idx); + sde->progress_mask = (u64)1 << (1 * TXE_NUM_SDMA_ENGINES + + this_idx); + sde->idle_mask = (u64)1 << (2 * TXE_NUM_SDMA_ENGINES + + this_idx); + /* Create a combined mask to cover all 3 interrupt sources */ + sde->imask = sde->int_mask | sde->progress_mask | + sde->idle_mask; + + spin_lock_init(&sde->tail_lock); + seqlock_init(&sde->head_lock); + spin_lock_init(&sde->senddmactrl_lock); + spin_lock_init(&sde->flushlist_lock); + /* insure there is always a zero bit */ + sde->ahg_bits = 0xfffffffe00000000ULL; + + sdma_set_state(sde, sdma_state_s00_hw_down); + + /* set up reference counting */ + kref_init(&sde->state.kref); + init_completion(&sde->state.comp); + + INIT_LIST_HEAD(&sde->flushlist); + INIT_LIST_HEAD(&sde->dmawait); + + sde->tail_csr = + get_kctxt_csr_addr(dd, this_idx, SD(TAIL)); + + if (idle_cnt) + dd->default_desc1 = + SDMA_DESC1_HEAD_TO_HOST_FLAG; + else + dd->default_desc1 = + SDMA_DESC1_INT_REQ_FLAG; + + tasklet_init(&sde->sdma_hw_clean_up_task, sdma_hw_clean_up_task, + (unsigned long)sde); + + tasklet_init(&sde->sdma_sw_clean_up_task, sdma_sw_clean_up_task, + (unsigned long)sde); + INIT_WORK(&sde->err_halt_worker, sdma_err_halt_wait); + INIT_WORK(&sde->flush_worker, sdma_field_flush); + + sde->progress_check_head = 0; + + setup_timer(&sde->err_progress_check_timer, + sdma_err_progress_check, (unsigned long)sde); + + sde->descq = dma_zalloc_coherent( + &dd->pcidev->dev, + descq_cnt * sizeof(u64[2]), + &sde->descq_phys, + GFP_KERNEL + ); + if (!sde->descq) + goto bail; + sde->tx_ring = + kcalloc(descq_cnt, sizeof(struct sdma_txreq *), + GFP_KERNEL); + if (!sde->tx_ring) + sde->tx_ring = + vzalloc( + sizeof(struct sdma_txreq *) * + descq_cnt); + if (!sde->tx_ring) + goto bail; + } + + dd->sdma_heads_size = L1_CACHE_BYTES * num_engines; + /* Allocate memory for DMA of head registers to memory */ + dd->sdma_heads_dma = dma_zalloc_coherent( + &dd->pcidev->dev, + dd->sdma_heads_size, + &dd->sdma_heads_phys, + GFP_KERNEL + ); + if (!dd->sdma_heads_dma) { + dd_dev_err(dd, "failed to allocate SendDMA head memory\n"); + goto bail; + } + + /* Allocate memory for pad */ + dd->sdma_pad_dma = dma_zalloc_coherent( + &dd->pcidev->dev, + sizeof(u32), + &dd->sdma_pad_phys, + GFP_KERNEL + ); + if (!dd->sdma_pad_dma) { + dd_dev_err(dd, "failed to allocate SendDMA pad memory\n"); + goto bail; + } + + /* assign each engine to different cacheline and init registers */ + curr_head = (void *)dd->sdma_heads_dma; + for (this_idx = 0; this_idx < num_engines; ++this_idx) { + unsigned long phys_offset; + + sde = &dd->per_sdma[this_idx]; + + sde->head_dma = curr_head; + curr_head += L1_CACHE_BYTES; + phys_offset = (unsigned long)sde->head_dma - + (unsigned long)dd->sdma_heads_dma; + sde->head_phys = dd->sdma_heads_phys + phys_offset; + init_sdma_regs(sde, per_sdma_credits, idle_cnt); + } + dd->flags |= HFI1_HAS_SEND_DMA; + dd->flags |= idle_cnt ? HFI1_HAS_SDMA_TIMEOUT : 0; + dd->num_sdma = num_engines; + if (sdma_map_init(dd, port, ppd->vls_operational, NULL)) + goto bail; + + if (rhashtable_init(&dd->sdma_rht, &sdma_rht_params)) + goto bail; + + dd_dev_info(dd, "SDMA num_sdma: %u\n", dd->num_sdma); + return 0; + +bail: + sdma_clean(dd, num_engines); + return -ENOMEM; +} + +/** + * sdma_all_running() - called when the link goes up + * @dd: hfi1_devdata + * + * This routine moves all engines to the running state. + */ +void sdma_all_running(struct hfi1_devdata *dd) +{ + struct sdma_engine *sde; + unsigned int i; + + /* move all engines to running */ + for (i = 0; i < dd->num_sdma; ++i) { + sde = &dd->per_sdma[i]; + sdma_process_event(sde, sdma_event_e30_go_running); + } +} + +/** + * sdma_all_idle() - called when the link goes down + * @dd: hfi1_devdata + * + * This routine moves all engines to the idle state. + */ +void sdma_all_idle(struct hfi1_devdata *dd) +{ + struct sdma_engine *sde; + unsigned int i; + + /* idle all engines */ + for (i = 0; i < dd->num_sdma; ++i) { + sde = &dd->per_sdma[i]; + sdma_process_event(sde, sdma_event_e70_go_idle); + } +} + +/** + * sdma_start() - called to kick off state processing for all engines + * @dd: hfi1_devdata + * + * This routine is for kicking off the state processing for all required + * sdma engines. Interrupts need to be working at this point. + * + */ +void sdma_start(struct hfi1_devdata *dd) +{ + unsigned i; + struct sdma_engine *sde; + + /* kick off the engines state processing */ + for (i = 0; i < dd->num_sdma; ++i) { + sde = &dd->per_sdma[i]; + sdma_process_event(sde, sdma_event_e10_go_hw_start); + } +} + +/* Needed for RHEL 7.2 backport */ +static void sdma_rht_destroy(struct rhashtable *ht) +{ + const struct bucket_table *tbl; + struct sdma_rht_node *rht_node, *next; + unsigned int i; + + tbl = rht_dereference(ht->tbl, ht); + for (i = 0; i < tbl->size; i++) { + rht_for_each_entry_safe(rht_node, next, tbl->buckets[i], + ht, node) + sdma_rht_free(rht_node, NULL); + } + + rhashtable_destroy(ht); +} + +/** + * sdma_exit() - used when module is removed + * @dd: hfi1_devdata + */ +void sdma_exit(struct hfi1_devdata *dd) +{ + unsigned this_idx; + struct sdma_engine *sde; + + for (this_idx = 0; dd->per_sdma && this_idx < dd->num_sdma; + ++this_idx) { + sde = &dd->per_sdma[this_idx]; + if (!list_empty(&sde->dmawait)) + dd_dev_err(dd, "sde %u: dmawait list not empty!\n", + sde->this_idx); + sdma_process_event(sde, sdma_event_e00_go_hw_down); + + del_timer_sync(&sde->err_progress_check_timer); + + /* + * This waits for the state machine to exit so it is not + * necessary to kill the sdma_sw_clean_up_task to make sure + * it is not running. + */ + sdma_finalput(&sde->state); + } + sdma_clean(dd, dd->num_sdma); + sdma_rht_destroy(&dd->sdma_rht); +} + +/* + * unmap the indicated descriptor + */ +static inline void sdma_unmap_desc( + struct hfi1_devdata *dd, + struct sdma_desc *descp) +{ + switch (sdma_mapping_type(descp)) { + case SDMA_MAP_SINGLE: + dma_unmap_single( + &dd->pcidev->dev, + sdma_mapping_addr(descp), + sdma_mapping_len(descp), + DMA_TO_DEVICE); + break; + case SDMA_MAP_PAGE: + dma_unmap_page( + &dd->pcidev->dev, + sdma_mapping_addr(descp), + sdma_mapping_len(descp), + DMA_TO_DEVICE); + break; + } +} +#endif /* __HFI1_ORIG__ */ + +/* + * return the mode as indicated by the first + * descriptor in the tx. + */ +static inline u8 ahg_mode(struct sdma_txreq *tx) +{ + return (tx->descp[0].qw[1] & SDMA_DESC1_HEADER_MODE_SMASK) + >> SDMA_DESC1_HEADER_MODE_SHIFT; +} + +/** + * __sdma_txclean() - clean tx of mappings, descp *kmalloc's + * @dd: hfi1_devdata for unmapping + * @tx: tx request to clean + * + * This is used in the progress routine to clean the tx or + * by the ULP to toss an in-process tx build. + * + * The code can be called multiple times without issue. + * + */ +void __sdma_txclean( + struct hfi1_devdata *dd, + struct sdma_txreq *tx) +{ + u16 i; + + if (tx->num_desc) { + u8 skip = 0, mode = ahg_mode(tx); + + /* TODO: enable sdma_unmap_desc */ + /* unmap first */ + //sdma_unmap_desc(dd, &tx->descp[0]); + /* determine number of AHG descriptors to skip */ + if (mode > SDMA_AHG_APPLY_UPDATE1) + skip = mode >> 1; + /* TODO: enable sdma_unmap_desc */ + // for (i = 1 + skip; i < tx->num_desc; i++) + // sdma_unmap_desc(dd, &tx->descp[i]); + tx->num_desc = 0; + } + kfree(tx->coalesce_buf); + tx->coalesce_buf = NULL; + /* kmalloc'ed descp */ + if (unlikely(tx->desc_limit > ARRAY_SIZE(tx->descs))) { + tx->desc_limit = ARRAY_SIZE(tx->descs); + kfree(tx->descp); + } +} +#ifdef __HFI1_ORIG__ + +static inline u16 sdma_gethead(struct sdma_engine *sde) +{ + struct hfi1_devdata *dd = sde->dd; + int use_dmahead; + u16 hwhead; + +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", + sde->this_idx, slashstrip(__FILE__), __LINE__, __func__); +#endif + +retry: + use_dmahead = HFI1_CAP_IS_KSET(USE_SDMA_HEAD) && __sdma_running(sde) && + (dd->flags & HFI1_HAS_SDMA_TIMEOUT); + hwhead = use_dmahead ? + (u16)le64_to_cpu(*sde->head_dma) : + (u16)read_sde_csr(sde, SD(HEAD)); + + if (unlikely(HFI1_CAP_IS_KSET(SDMA_HEAD_CHECK))) { + u16 cnt; + u16 swtail; + u16 swhead; + int sane; + + swhead = sde->descq_head & sde->sdma_mask; + /* this code is really bad for cache line trading */ + swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask; + cnt = sde->descq_cnt; + + if (swhead < swtail) + /* not wrapped */ + sane = (hwhead >= swhead) & (hwhead <= swtail); + else if (swhead > swtail) + /* wrapped around */ + sane = ((hwhead >= swhead) && (hwhead < cnt)) || + (hwhead <= swtail); + else + /* empty */ + sane = (hwhead == swhead); + + if (unlikely(!sane)) { + dd_dev_err(dd, "SDMA(%u) bad head (%s) hwhd=%hu swhd=%hu swtl=%hu cnt=%hu\n", + sde->this_idx, + use_dmahead ? "dma" : "kreg", + hwhead, swhead, swtail, cnt); + if (use_dmahead) { + /* try one more time, using csr */ + use_dmahead = 0; + goto retry; + } + /* proceed as if no progress */ + hwhead = swhead; + } + } + return hwhead; +} + +/* + * This is called when there are send DMA descriptors that might be + * available. + * + * This is called with head_lock held. + */ +static void sdma_desc_avail(struct sdma_engine *sde, unsigned avail) +{ + struct iowait *wait, *nw; + struct iowait *waits[SDMA_WAIT_BATCH_SIZE]; + unsigned i, n = 0, seq; + struct hfi1_ibdev *dev = &sde->dd->verbs_dev; + +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", sde->this_idx, + slashstrip(__FILE__), __LINE__, __func__); + dd_dev_err(sde->dd, "avail: %u\n", avail); +#endif + + do { + seq = read_seqbegin(&dev->iowait_lock); + if (!list_empty(&sde->dmawait)) { + /* at least one item */ + write_seqlock(&dev->iowait_lock); + /* Harvest waiters wanting DMA descriptors */ + list_for_each_entry_safe( + wait, + nw, + &sde->dmawait, + list) { + u32 num_desc; + + if (!wait->wakeup) + continue; + if (n == ARRAY_SIZE(waits)) + break; + num_desc = iowait_get_all_desc(wait); + if (num_desc > avail) + break; + avail -= num_desc; + list_del_init(&wait->list); + waits[n++] = wait; + } + write_sequnlock(&dev->iowait_lock); + break; + } + } while (read_seqretry(&dev->iowait_lock, seq)); + + for (i = 0; i < n; i++) + waits[i]->wakeup(waits[i], SDMA_AVAIL_REASON); +} + +/* head_lock must be held */ +static void sdma_make_progress(struct sdma_engine *sde, u64 status) +{ + struct sdma_txreq *txp = NULL; + int progress = 0; + u16 hwhead, swhead; + int idle_check_done = 0; + + hfi1_cdbg(AIOWRITE, "+"); + hwhead = sdma_gethead(sde); + + /* The reason for some of the complexity of this code is that + * not all descriptors have corresponding txps. So, we have to + * be able to skip over descs until we wander into the range of + * the next txp on the list. + */ + +retry: + txp = get_txhead(sde); + swhead = sde->descq_head & sde->sdma_mask; + trace_hfi1_sdma_progress(sde, hwhead, swhead, txp); + while (swhead != hwhead) { + /* advance head, wrap if needed */ + swhead = ++sde->descq_head & sde->sdma_mask; + + /* if now past this txp's descs, do the callback */ + if (txp && txp->next_descq_idx == swhead) { + /* remove from list */ + sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL; + complete_tx(sde, txp, SDMA_TXREQ_S_OK); + /* see if there is another txp */ + txp = get_txhead(sde); + } + trace_hfi1_sdma_progress(sde, hwhead, swhead, txp); + progress++; + } + + /* + * The SDMA idle interrupt is not guaranteed to be ordered with respect + * to updates to the the dma_head location in host memory. The head + * value read might not be fully up to date. If there are pending + * descriptors and the SDMA idle interrupt fired then read from the + * CSR SDMA head instead to get the latest value from the hardware. + * The hardware SDMA head should be read at most once in this invocation + * of sdma_make_progress(..) which is ensured by idle_check_done flag + */ + if ((status & sde->idle_mask) && !idle_check_done) { + u16 swtail; + + swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask; + if (swtail != hwhead) { + hwhead = (u16)read_sde_csr(sde, SD(HEAD)); + idle_check_done = 1; + goto retry; + } + } + + sde->last_status = status; + if (progress) + sdma_desc_avail(sde, sdma_descq_freecnt(sde)); + hfi1_cdbg(AIOWRITE, "-"); +} + +/* + * sdma_engine_interrupt() - interrupt handler for engine + * @sde: sdma engine + * @status: sdma interrupt reason + * + * Status is a mask of the 3 possible interrupts for this engine. It will + * contain bits _only_ for this SDMA engine. It will contain at least one + * bit, it may contain more. + */ +void sdma_engine_interrupt(struct sdma_engine *sde, u64 status) +{ + hfi1_cdbg(AIOWRITE, "+"); + trace_hfi1_sdma_engine_interrupt(sde, status); + write_seqlock(&sde->head_lock); + sdma_set_desc_cnt(sde, sdma_desct_intr); + if (status & sde->idle_mask) + sde->idle_int_cnt++; + else if (status & sde->progress_mask) + sde->progress_int_cnt++; + else if (status & sde->int_mask) + sde->sdma_int_cnt++; + sdma_make_progress(sde, status); + write_sequnlock(&sde->head_lock); + hfi1_cdbg(AIOWRITE, "-"); +} + +/** + * sdma_engine_error() - error handler for engine + * @sde: sdma engine + * @status: sdma interrupt reason + */ +void sdma_engine_error(struct sdma_engine *sde, u64 status) +{ + unsigned long flags; + +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) error status 0x%llx state %s\n", + sde->this_idx, + (unsigned long long)status, + sdma_state_names[sde->state.current_state]); +#endif + spin_lock_irqsave(&sde->tail_lock, flags); + write_seqlock(&sde->head_lock); + if (status & ALL_SDMA_ENG_HALT_ERRS) + __sdma_process_event(sde, sdma_event_e60_hw_halted); + if (status & ~SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK)) { + dd_dev_err(sde->dd, + "SDMA (%u) engine error: 0x%llx state %s\n", + sde->this_idx, + (unsigned long long)status, + sdma_state_names[sde->state.current_state]); + dump_sdma_state(sde); + } + write_sequnlock(&sde->head_lock); + spin_unlock_irqrestore(&sde->tail_lock, flags); +} + +static void sdma_sendctrl(struct sdma_engine *sde, unsigned op) +{ + u64 set_senddmactrl = 0; + u64 clr_senddmactrl = 0; + unsigned long flags; + +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) senddmactrl E=%d I=%d H=%d C=%d\n", + sde->this_idx, + (op & SDMA_SENDCTRL_OP_ENABLE) ? 1 : 0, + (op & SDMA_SENDCTRL_OP_INTENABLE) ? 1 : 0, + (op & SDMA_SENDCTRL_OP_HALT) ? 1 : 0, + (op & SDMA_SENDCTRL_OP_CLEANUP) ? 1 : 0); +#endif + + if (op & SDMA_SENDCTRL_OP_ENABLE) + set_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK); + else + clr_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK); + + if (op & SDMA_SENDCTRL_OP_INTENABLE) + set_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK); + else + clr_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK); + + if (op & SDMA_SENDCTRL_OP_HALT) + set_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK); + else + clr_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK); + + spin_lock_irqsave(&sde->senddmactrl_lock, flags); + + sde->p_senddmactrl |= set_senddmactrl; + sde->p_senddmactrl &= ~clr_senddmactrl; + + if (op & SDMA_SENDCTRL_OP_CLEANUP) + write_sde_csr(sde, SD(CTRL), + sde->p_senddmactrl | + SD(CTRL_SDMA_CLEANUP_SMASK)); + else + write_sde_csr(sde, SD(CTRL), sde->p_senddmactrl); + + spin_unlock_irqrestore(&sde->senddmactrl_lock, flags); + +#ifdef CONFIG_SDMA_VERBOSITY + sdma_dumpstate(sde); +#endif +} + +static void sdma_setlengen(struct sdma_engine *sde) +{ +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", + sde->this_idx, slashstrip(__FILE__), __LINE__, __func__); +#endif + + /* + * Set SendDmaLenGen and clear-then-set the MSB of the generation + * count to enable generation checking and load the internal + * generation counter. + */ + write_sde_csr(sde, SD(LEN_GEN), + (sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT)); + write_sde_csr(sde, SD(LEN_GEN), + ((sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT)) | + (4ULL << SD(LEN_GEN_GENERATION_SHIFT))); +} + +#endif /* __HFI1_ORIG__ */ +static inline void sdma_update_tail(struct sdma_engine *sde, u16 tail) +{ + hfi1_cdbg(AIOWRITE, "."); + /* Commit writes to memory and advance the tail on the chip */ + smp_wmb(); /* see get_txhead() */ + writeq(tail, sde->tail_csr); +} +#ifdef __HFI1_ORIG__ + +/* + * This is called when changing to state s10_hw_start_up_halt_wait as + * a result of send buffer errors or send DMA descriptor errors. + */ +static void sdma_hw_start_up(struct sdma_engine *sde) +{ + u64 reg; + +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", + sde->this_idx, slashstrip(__FILE__), __LINE__, __func__); +#endif + + sdma_setlengen(sde); + sdma_update_tail(sde, 0); /* Set SendDmaTail */ + *sde->head_dma = 0; + + reg = SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_MASK) << + SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SHIFT); + write_sde_csr(sde, SD(ENG_ERR_CLEAR), reg); +} + +/* + * set_sdma_integrity + * + * Set the SEND_DMA_CHECK_ENABLE register for send DMA engine 'sde'. + */ +static void set_sdma_integrity(struct sdma_engine *sde) +{ + struct hfi1_devdata *dd = sde->dd; + + write_sde_csr(sde, SD(CHECK_ENABLE), + hfi1_pkt_base_sdma_integrity(dd)); +} + +static void init_sdma_regs( + struct sdma_engine *sde, + u32 credits, + uint idle_cnt) +{ + u8 opval, opmask; +#ifdef CONFIG_SDMA_VERBOSITY + struct hfi1_devdata *dd = sde->dd; + + dd_dev_err(dd, "CONFIG SDMA(%u) %s:%d %s()\n", + sde->this_idx, slashstrip(__FILE__), __LINE__, __func__); +#endif + + write_sde_csr(sde, SD(BASE_ADDR), sde->descq_phys); + sdma_setlengen(sde); + sdma_update_tail(sde, 0); /* Set SendDmaTail */ + write_sde_csr(sde, SD(RELOAD_CNT), idle_cnt); + write_sde_csr(sde, SD(DESC_CNT), 0); + write_sde_csr(sde, SD(HEAD_ADDR), sde->head_phys); + write_sde_csr(sde, SD(MEMORY), + ((u64)credits << SD(MEMORY_SDMA_MEMORY_CNT_SHIFT)) | + ((u64)(credits * sde->this_idx) << + SD(MEMORY_SDMA_MEMORY_INDEX_SHIFT))); + write_sde_csr(sde, SD(ENG_ERR_MASK), ~0ull); + set_sdma_integrity(sde); + opmask = OPCODE_CHECK_MASK_DISABLED; + opval = OPCODE_CHECK_VAL_DISABLED; + write_sde_csr(sde, SD(CHECK_OPCODE), + (opmask << SEND_CTXT_CHECK_OPCODE_MASK_SHIFT) | + (opval << SEND_CTXT_CHECK_OPCODE_VALUE_SHIFT)); +} + +#ifdef CONFIG_SDMA_VERBOSITY + +#define sdma_dumpstate_helper0(reg) do { \ + csr = read_csr(sde->dd, reg); \ + dd_dev_err(sde->dd, "%36s 0x%016llx\n", #reg, csr); \ + } while (0) + +#define sdma_dumpstate_helper(reg) do { \ + csr = read_sde_csr(sde, reg); \ + dd_dev_err(sde->dd, "%36s[%02u] 0x%016llx\n", \ + #reg, sde->this_idx, csr); \ + } while (0) + +#define sdma_dumpstate_helper2(reg) do { \ + csr = read_csr(sde->dd, reg + (8 * i)); \ + dd_dev_err(sde->dd, "%33s_%02u 0x%016llx\n", \ + #reg, i, csr); \ + } while (0) + +void sdma_dumpstate(struct sdma_engine *sde) +{ + u64 csr; + unsigned i; + + sdma_dumpstate_helper(SD(CTRL)); + sdma_dumpstate_helper(SD(STATUS)); + sdma_dumpstate_helper0(SD(ERR_STATUS)); + sdma_dumpstate_helper0(SD(ERR_MASK)); + sdma_dumpstate_helper(SD(ENG_ERR_STATUS)); + sdma_dumpstate_helper(SD(ENG_ERR_MASK)); + + for (i = 0; i < CCE_NUM_INT_CSRS; ++i) { + sdma_dumpstate_helper2(CCE_INT_STATUS); + sdma_dumpstate_helper2(CCE_INT_MASK); + sdma_dumpstate_helper2(CCE_INT_BLOCKED); + } + + sdma_dumpstate_helper(SD(TAIL)); + sdma_dumpstate_helper(SD(HEAD)); + sdma_dumpstate_helper(SD(PRIORITY_THLD)); + sdma_dumpstate_helper(SD(IDLE_CNT)); + sdma_dumpstate_helper(SD(RELOAD_CNT)); + sdma_dumpstate_helper(SD(DESC_CNT)); + sdma_dumpstate_helper(SD(DESC_FETCHED_CNT)); + sdma_dumpstate_helper(SD(MEMORY)); + sdma_dumpstate_helper0(SD(ENGINES)); + sdma_dumpstate_helper0(SD(MEM_SIZE)); + /* sdma_dumpstate_helper(SEND_EGRESS_SEND_DMA_STATUS); */ + sdma_dumpstate_helper(SD(BASE_ADDR)); + sdma_dumpstate_helper(SD(LEN_GEN)); + sdma_dumpstate_helper(SD(HEAD_ADDR)); + sdma_dumpstate_helper(SD(CHECK_ENABLE)); + sdma_dumpstate_helper(SD(CHECK_VL)); + sdma_dumpstate_helper(SD(CHECK_JOB_KEY)); + sdma_dumpstate_helper(SD(CHECK_PARTITION_KEY)); + sdma_dumpstate_helper(SD(CHECK_SLID)); + sdma_dumpstate_helper(SD(CHECK_OPCODE)); +} +#endif + +static void dump_sdma_state(struct sdma_engine *sde) +{ + struct hw_sdma_desc *descq; + struct hw_sdma_desc *descqp; + u64 desc[2]; + u64 addr; + u8 gen; + u16 len; + u16 head, tail, cnt; + + head = sde->descq_head & sde->sdma_mask; + tail = sde->descq_tail & sde->sdma_mask; + cnt = sdma_descq_freecnt(sde); + descq = sde->descq; + + dd_dev_err(sde->dd, + "SDMA (%u) descq_head: %u descq_tail: %u freecnt: %u FLE %d\n", + sde->this_idx, head, tail, cnt, + !list_empty(&sde->flushlist)); + + /* print info for each entry in the descriptor queue */ + while (head != tail) { + char flags[6] = { 'x', 'x', 'x', 'x', 0 }; + + descqp = &sde->descq[head]; + desc[0] = le64_to_cpu(descqp->qw[0]); + desc[1] = le64_to_cpu(descqp->qw[1]); + flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-'; + flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ? + 'H' : '-'; + flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-'; + flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-'; + addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT) + & SDMA_DESC0_PHY_ADDR_MASK; + gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT) + & SDMA_DESC1_GENERATION_MASK; + len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT) + & SDMA_DESC0_BYTE_COUNT_MASK; + dd_dev_err(sde->dd, + "SDMA sdmadesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n", + head, flags, addr, gen, len); + dd_dev_err(sde->dd, + "\tdesc0:0x%016llx desc1 0x%016llx\n", + desc[0], desc[1]); + if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) + dd_dev_err(sde->dd, + "\taidx: %u amode: %u alen: %u\n", + (u8)((desc[1] & + SDMA_DESC1_HEADER_INDEX_SMASK) >> + SDMA_DESC1_HEADER_INDEX_SHIFT), + (u8)((desc[1] & + SDMA_DESC1_HEADER_MODE_SMASK) >> + SDMA_DESC1_HEADER_MODE_SHIFT), + (u8)((desc[1] & + SDMA_DESC1_HEADER_DWS_SMASK) >> + SDMA_DESC1_HEADER_DWS_SHIFT)); + head++; + head &= sde->sdma_mask; + } +} + +#define SDE_FMT \ + "SDE %u CPU %d STE %s C 0x%llx S 0x%016llx E 0x%llx T(HW) 0x%llx T(SW) 0x%x H(HW) 0x%llx H(SW) 0x%x H(D) 0x%llx DM 0x%llx GL 0x%llx R 0x%llx LIS 0x%llx AHGI 0x%llx TXT %u TXH %u DT %u DH %u FLNE %d DQF %u SLC 0x%llx\n" +/** + * sdma_seqfile_dump_sde() - debugfs dump of sde + * @s: seq file + * @sde: send dma engine to dump + * + * This routine dumps the sde to the indicated seq file. + */ +void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *sde) +{ + u16 head, tail; + struct hw_sdma_desc *descqp; + u64 desc[2]; + u64 addr; + u8 gen; + u16 len; + + head = sde->descq_head & sde->sdma_mask; + tail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask; + seq_printf(s, SDE_FMT, sde->this_idx, + sde->cpu, + sdma_state_name(sde->state.current_state), + (unsigned long long)read_sde_csr(sde, SD(CTRL)), + (unsigned long long)read_sde_csr(sde, SD(STATUS)), + (unsigned long long)read_sde_csr(sde, SD(ENG_ERR_STATUS)), + (unsigned long long)read_sde_csr(sde, SD(TAIL)), tail, + (unsigned long long)read_sde_csr(sde, SD(HEAD)), head, + (unsigned long long)le64_to_cpu(*sde->head_dma), + (unsigned long long)read_sde_csr(sde, SD(MEMORY)), + (unsigned long long)read_sde_csr(sde, SD(LEN_GEN)), + (unsigned long long)read_sde_csr(sde, SD(RELOAD_CNT)), + (unsigned long long)sde->last_status, + (unsigned long long)sde->ahg_bits, + sde->tx_tail, + sde->tx_head, + sde->descq_tail, + sde->descq_head, + !list_empty(&sde->flushlist), + sde->descq_full_count, + (unsigned long long)read_sde_csr(sde, SEND_DMA_CHECK_SLID)); + + /* print info for each entry in the descriptor queue */ + while (head != tail) { + char flags[6] = { 'x', 'x', 'x', 'x', 0 }; + + descqp = &sde->descq[head]; + desc[0] = le64_to_cpu(descqp->qw[0]); + desc[1] = le64_to_cpu(descqp->qw[1]); + flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-'; + flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ? + 'H' : '-'; + flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-'; + flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-'; + addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT) + & SDMA_DESC0_PHY_ADDR_MASK; + gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT) + & SDMA_DESC1_GENERATION_MASK; + len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT) + & SDMA_DESC0_BYTE_COUNT_MASK; + seq_printf(s, + "\tdesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n", + head, flags, addr, gen, len); + if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) + seq_printf(s, "\t\tahgidx: %u ahgmode: %u\n", + (u8)((desc[1] & + SDMA_DESC1_HEADER_INDEX_SMASK) >> + SDMA_DESC1_HEADER_INDEX_SHIFT), + (u8)((desc[1] & + SDMA_DESC1_HEADER_MODE_SMASK) >> + SDMA_DESC1_HEADER_MODE_SHIFT)); + head = (head + 1) & sde->sdma_mask; + } +} + +#endif /* __HFI1_ORIG__ */ +/* + * add the generation number into + * the qw1 and return + */ +static inline u64 add_gen(struct sdma_engine *sde, u64 qw1) +{ + u8 generation = (sde->descq_tail >> sde->sdma_shift) & 3; + + qw1 &= ~SDMA_DESC1_GENERATION_SMASK; + qw1 |= ((u64)generation & SDMA_DESC1_GENERATION_MASK) + << SDMA_DESC1_GENERATION_SHIFT; + return qw1; +} + +/* + * This routine submits the indicated tx + * + * Space has already been guaranteed and + * tail side of ring is locked. + * + * The hardware tail update is done + * in the caller and that is facilitated + * by returning the new tail. + * + * There is special case logic for ahg + * to not add the generation number for + * up to 2 descriptors that follow the + * first descriptor. + * + */ +static inline u16 submit_tx(struct sdma_engine *sde, struct sdma_txreq *tx) +{ + int i; + u16 tail; + struct sdma_desc *descp = tx->descp; + u8 skip = 0, mode = ahg_mode(tx); + hfi1_cdbg(AIOWRITE, "+"); + tail = sde->descq_tail & sde->sdma_mask; + sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]); + sde->descq[tail].qw[1] = cpu_to_le64(add_gen(sde, descp->qw[1])); + // trace_hfi1_sdma_descriptor(sde, descp->qw[0], descp->qw[1], + // tail, &sde->descq[tail]); + tail = ++sde->descq_tail & sde->sdma_mask; + descp++; + if (mode > SDMA_AHG_APPLY_UPDATE1) + skip = mode >> 1; + for (i = 1; i < tx->num_desc; i++, descp++) { + u64 qw1; + + sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]); + if (skip) { + /* edits don't have generation */ + qw1 = descp->qw[1]; + skip--; + } else { + /* replace generation with real one for non-edits */ + qw1 = add_gen(sde, descp->qw[1]); + } + sde->descq[tail].qw[1] = cpu_to_le64(qw1); + // trace_hfi1_sdma_descriptor(sde, descp->qw[0], qw1, + // tail, &sde->descq[tail]); + tail = ++sde->descq_tail & sde->sdma_mask; + } + tx->next_descq_idx = tail; +#ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER + tx->sn = sde->tail_sn++; + // trace_hfi1_sdma_in_sn(sde, tx->sn); + WARN_ON_ONCE(sde->tx_ring[sde->tx_tail & sde->sdma_mask]); +#endif + sde->tx_ring[sde->tx_tail++ & sde->sdma_mask] = tx; + sde->desc_avail -= tx->num_desc; + hfi1_cdbg(AIOWRITE, "-"); + return tail; +} + +/* + * Check for progress + */ +static int sdma_check_progress( + struct sdma_engine *sde, + struct iowait_work *wait, + struct sdma_txreq *tx) +{ + int ret; + + hfi1_cdbg(AIOWRITE, "+"); + sde->desc_avail = sdma_descq_freecnt(sde); + if (tx->num_desc <= sde->desc_avail) + return -EAGAIN; + /* pulse the head_lock */ + if (wait && iowait_ioww_to_iow(wait)->sleep) { + unsigned seq; + + seq = raw_seqcount_begin( + (const seqcount_t *)&sde->head_lock.seqcount); + ret = wait->iow->sleep(sde, wait, tx, seq); + if (ret == -EAGAIN) + sde->desc_avail = sdma_descq_freecnt(sde); + } else { + ret = -EBUSY; + } + hfi1_cdbg(AIOWRITE, "-"); + return ret; +} +#ifdef __HFI1_ORIG__ + +/** + * sdma_send_txreq() - submit a tx req to ring + * @sde: sdma engine to use + * @wait: SE wait structure to use when full (may be NULL) + * @tx: sdma_txreq to submit + * + * The call submits the tx into the ring. If a iowait structure is non-NULL + * the packet will be queued to the list in wait. + * + * Return: + * 0 - Success, -EINVAL - sdma_txreq incomplete, -EBUSY - no space in + * ring (wait == NULL) + * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state + */ +int sdma_send_txreq(struct sdma_engine *sde, + struct iowait_work *wait, + struct sdma_txreq *tx) +{ + int ret = 0; + u16 tail; + unsigned long flags; + + hfi1_cdbg(AIOWRITE, "+"); + /* user should have supplied entire packet */ + if (unlikely(tx->tlen)) + return -EINVAL; + tx->wait = iowait_ioww_to_iow(wait); + spin_lock_irqsave(&sde->tail_lock, flags); +retry: + if (unlikely(!__sdma_running(sde))) + goto unlock_noconn; + if (unlikely(tx->num_desc > sde->desc_avail)) + goto nodesc; + tail = submit_tx(sde, tx); + if (wait) + iowait_sdma_inc(iowait_ioww_to_iow(wait)); + sdma_update_tail(sde, tail); +unlock: + spin_unlock_irqrestore(&sde->tail_lock, flags); + hfi1_cdbg(AIOWRITE, "-"); + return ret; +unlock_noconn: + if (wait) + iowait_sdma_inc(iowait_ioww_to_iow(wait)); + tx->next_descq_idx = 0; +#ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER + tx->sn = sde->tail_sn++; + trace_hfi1_sdma_in_sn(sde, tx->sn); +#endif + spin_lock(&sde->flushlist_lock); + list_add_tail(&tx->list, &sde->flushlist); + spin_unlock(&sde->flushlist_lock); + iowait_inc_wait_count(wait, tx->num_desc); + schedule_work(&sde->flush_worker); + ret = -ECOMM; + goto unlock; +nodesc: + ret = sdma_check_progress(sde, wait, tx); + if (ret == -EAGAIN) { + ret = 0; + goto retry; + } + sde->descq_full_count++; + goto unlock; +} + +#endif /* __HFI1_ORIG__ */ +/** + * sdma_send_txlist() - submit a list of tx req to ring + * @sde: sdma engine to use + * @wait: SE wait structure to use when full (may be NULL) + * @tx_list: list of sdma_txreqs to submit + * @count: pointer to a u32 which, after return will contain the total number of + * sdma_txreqs removed from the tx_list. This will include sdma_txreqs + * whose SDMA descriptors are submitted to the ring and the sdma_txreqs + * which are added to SDMA engine flush list if the SDMA engine state is + * not running. + * + * The call submits the list into the ring. + * + * If the iowait structure is non-NULL and not equal to the iowait list + * the unprocessed part of the list will be appended to the list in wait. + * + * In all cases, the tx_list will be updated so the head of the tx_list is + * the list of descriptors that have yet to be transmitted. + * + * The intent of this call is to provide a more efficient + * way of submitting multiple packets to SDMA while holding the tail + * side locking. + * + * Return: + * 0 - Success, + * -EINVAL - sdma_txreq incomplete, -EBUSY - no space in ring (wait == NULL) + * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state + */ +int sdma_send_txlist(struct sdma_engine *sde, struct iowait_work *wait, + struct list_head *tx_list, u32 *count_out) +{ + struct sdma_txreq *tx, *tx_next; + int ret = 0; + unsigned long flags; + u16 tail = INVALID_TAIL; + u32 submit_count = 0, flush_count = 0, total_count; + + hfi1_cdbg(AIOWRITE, "+"); + spin_lock_irqsave(&sde->tail_lock, flags); +retry: + /* + * XXX: do we really need to build tx_list?? + * We could just submit the requests straight to the HW.. + */ + list_for_each_entry_safe(tx, tx_next, tx_list, list) { + tx->wait = iowait_ioww_to_iow(wait); + if (unlikely(!__sdma_running(sde))) + goto unlock_noconn; + if (unlikely(tx->num_desc > sde->desc_avail)) + goto nodesc; + if (unlikely(tx->tlen)) { + ret = -EINVAL; + goto update_tail; + } + list_del_init(&tx->list); + tail = submit_tx(sde, tx); + submit_count++; + if (tail != INVALID_TAIL && + (submit_count & SDMA_TAIL_UPDATE_THRESH) == 0) { + sdma_update_tail(sde, tail); + tail = INVALID_TAIL; + } + } +update_tail: + total_count = submit_count + flush_count; + if (wait) + iowait_sdma_add(iowait_ioww_to_iow(wait), total_count); + if (tail != INVALID_TAIL) + sdma_update_tail(sde, tail); + spin_unlock_irqrestore(&sde->tail_lock, flags); + *count_out = total_count; + hfi1_cdbg(AIOWRITE, "-"); + return ret; +unlock_noconn: + spin_lock(&sde->flushlist_lock); + list_for_each_entry_safe(tx, tx_next, tx_list, list) { + tx->wait = iowait_ioww_to_iow(wait); + list_del_init(&tx->list); + tx->next_descq_idx = 0; +#ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER + tx->sn = sde->tail_sn++; + // trace_hfi1_sdma_in_sn(sde, tx->sn); +#endif + list_add_tail(&tx->list, &sde->flushlist); + flush_count++; + iowait_inc_wait_count(wait, tx->num_desc); + } + spin_unlock(&sde->flushlist_lock); + // TODO: schedule_work + //schedule_work(&sde->flush_worker); + ret = -ECOMM; + goto update_tail; +nodesc: + ret = sdma_check_progress(sde, wait, tx); + if (ret == -EAGAIN) { + ret = 0; + goto retry; + } + sde->descq_full_count++; + goto update_tail; +} +#ifdef __HFI1_ORIG__ + +static void sdma_process_event(struct sdma_engine *sde, enum sdma_events event) +{ + unsigned long flags; + + spin_lock_irqsave(&sde->tail_lock, flags); + write_seqlock(&sde->head_lock); + + __sdma_process_event(sde, event); + + if (sde->state.current_state == sdma_state_s99_running) + sdma_desc_avail(sde, sdma_descq_freecnt(sde)); + + write_sequnlock(&sde->head_lock); + spin_unlock_irqrestore(&sde->tail_lock, flags); +} + +static void __sdma_process_event(struct sdma_engine *sde, + enum sdma_events event) +{ + struct sdma_state *ss = &sde->state; + int need_progress = 0; + + /* CONFIG SDMA temporary */ +#ifdef CONFIG_SDMA_VERBOSITY + dd_dev_err(sde->dd, "CONFIG SDMA(%u) [%s] %s\n", sde->this_idx, + sdma_state_names[ss->current_state], + sdma_event_names[event]); +#endif + + switch (ss->current_state) { + case sdma_state_s00_hw_down: + switch (event) { + case sdma_event_e00_go_hw_down: + break; + case sdma_event_e30_go_running: + /* + * If down, but running requested (usually result + * of link up, then we need to start up. + * This can happen when hw down is requested while + * bringing the link up with traffic active on + * 7220, e.g. + */ + ss->go_s99_running = 1; + /* fall through and start dma engine */ + case sdma_event_e10_go_hw_start: + /* This reference means the state machine is started */ + sdma_get(&sde->state); + sdma_set_state(sde, + sdma_state_s10_hw_start_up_halt_wait); + break; + case sdma_event_e15_hw_halt_done: + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e40_sw_cleaned: + sdma_sw_tear_down(sde); + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + break; + case sdma_event_e70_go_idle: + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e85_link_down: + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s10_hw_start_up_halt_wait: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + sdma_sw_tear_down(sde); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + sdma_set_state(sde, + sdma_state_s15_hw_start_up_clean_wait); + sdma_start_hw_clean_up(sde); + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e30_go_running: + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + schedule_work(&sde->err_halt_worker); + break; + case sdma_event_e70_go_idle: + ss->go_s99_running = 0; + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e85_link_down: + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s15_hw_start_up_clean_wait: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + sdma_sw_tear_down(sde); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + break; + case sdma_event_e25_hw_clean_up_done: + sdma_hw_start_up(sde); + sdma_set_state(sde, ss->go_s99_running ? + sdma_state_s99_running : + sdma_state_s20_idle); + break; + case sdma_event_e30_go_running: + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + break; + case sdma_event_e70_go_idle: + ss->go_s99_running = 0; + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e85_link_down: + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s20_idle: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + sdma_sw_tear_down(sde); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e30_go_running: + sdma_set_state(sde, sdma_state_s99_running); + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + sdma_set_state(sde, sdma_state_s50_hw_halt_wait); + schedule_work(&sde->err_halt_worker); + break; + case sdma_event_e70_go_idle: + break; + case sdma_event_e85_link_down: + /* fall through */ + case sdma_event_e80_hw_freeze: + sdma_set_state(sde, sdma_state_s80_hw_freeze); + atomic_dec(&sde->dd->sdma_unfreeze_count); + wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s30_sw_clean_up_wait: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e30_go_running: + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + sdma_set_state(sde, sdma_state_s40_hw_clean_up_wait); + sdma_start_hw_clean_up(sde); + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + break; + case sdma_event_e70_go_idle: + ss->go_s99_running = 0; + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e85_link_down: + ss->go_s99_running = 0; + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s40_hw_clean_up_wait: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + break; + case sdma_event_e25_hw_clean_up_done: + sdma_hw_start_up(sde); + sdma_set_state(sde, ss->go_s99_running ? + sdma_state_s99_running : + sdma_state_s20_idle); + break; + case sdma_event_e30_go_running: + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + break; + case sdma_event_e70_go_idle: + ss->go_s99_running = 0; + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e85_link_down: + ss->go_s99_running = 0; + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s50_hw_halt_wait: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e30_go_running: + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + schedule_work(&sde->err_halt_worker); + break; + case sdma_event_e70_go_idle: + ss->go_s99_running = 0; + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e85_link_down: + ss->go_s99_running = 0; + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s60_idle_halt_wait: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e30_go_running: + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + schedule_work(&sde->err_halt_worker); + break; + case sdma_event_e70_go_idle: + ss->go_s99_running = 0; + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e85_link_down: + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s80_hw_freeze: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e30_go_running: + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + break; + case sdma_event_e70_go_idle: + ss->go_s99_running = 0; + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + sdma_set_state(sde, sdma_state_s82_freeze_sw_clean); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e82_hw_unfreeze: + break; + case sdma_event_e85_link_down: + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s82_freeze_sw_clean: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e30_go_running: + ss->go_s99_running = 1; + break; + case sdma_event_e40_sw_cleaned: + /* notify caller this engine is done cleaning */ + atomic_dec(&sde->dd->sdma_unfreeze_count); + wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + break; + case sdma_event_e70_go_idle: + ss->go_s99_running = 0; + break; + case sdma_event_e80_hw_freeze: + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + sdma_hw_start_up(sde); + sdma_set_state(sde, ss->go_s99_running ? + sdma_state_s99_running : + sdma_state_s20_idle); + break; + case sdma_event_e85_link_down: + break; + case sdma_event_e90_sw_halted: + break; + } + break; + + case sdma_state_s99_running: + switch (event) { + case sdma_event_e00_go_hw_down: + sdma_set_state(sde, sdma_state_s00_hw_down); + tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); + break; + case sdma_event_e10_go_hw_start: + break; + case sdma_event_e15_hw_halt_done: + break; + case sdma_event_e25_hw_clean_up_done: + break; + case sdma_event_e30_go_running: + break; + case sdma_event_e40_sw_cleaned: + break; + case sdma_event_e50_hw_cleaned: + break; + case sdma_event_e60_hw_halted: + need_progress = 1; + sdma_err_progress_check_schedule(sde); + case sdma_event_e90_sw_halted: + /* + * SW initiated halt does not perform engines + * progress check + */ + sdma_set_state(sde, sdma_state_s50_hw_halt_wait); + schedule_work(&sde->err_halt_worker); + break; + case sdma_event_e70_go_idle: + sdma_set_state(sde, sdma_state_s60_idle_halt_wait); + break; + case sdma_event_e85_link_down: + ss->go_s99_running = 0; + /* fall through */ + case sdma_event_e80_hw_freeze: + sdma_set_state(sde, sdma_state_s80_hw_freeze); + atomic_dec(&sde->dd->sdma_unfreeze_count); + wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); + break; + case sdma_event_e81_hw_frozen: + break; + case sdma_event_e82_hw_unfreeze: + break; + } + break; + } + + ss->last_event = event; + if (need_progress) + sdma_make_progress(sde, 0); +} + +/* + * _extend_sdma_tx_descs() - helper to extend txreq + * + * This is called once the initial nominal allocation + * of descriptors in the sdma_txreq is exhausted. + * + * The code will bump the allocation up to the max + * of MAX_DESC (64) descriptors. There doesn't seem + * much point in an interim step. The last descriptor + * is reserved for coalesce buffer in order to support + * cases where input packet has >MAX_DESC iovecs. + * + */ +static int _extend_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx) +{ + int i; + + /* Handle last descriptor */ + if (unlikely((tx->num_desc == (MAX_DESC - 1)))) { + /* if tlen is 0, it is for padding, release last descriptor */ + if (!tx->tlen) { + tx->desc_limit = MAX_DESC; + } else if (!tx->coalesce_buf) { + /* allocate coalesce buffer with space for padding */ + tx->coalesce_buf = kmalloc(tx->tlen + sizeof(u32), + GFP_ATOMIC); + if (!tx->coalesce_buf) + goto enomem; + tx->coalesce_idx = 0; + } + return 0; + } + + if (unlikely(tx->num_desc == MAX_DESC)) + goto enomem; + + tx->descp = kmalloc_array( + MAX_DESC, + sizeof(struct sdma_desc), + GFP_ATOMIC); + if (!tx->descp) + goto enomem; + + /* reserve last descriptor for coalescing */ + tx->desc_limit = MAX_DESC - 1; + /* copy ones already built */ + for (i = 0; i < tx->num_desc; i++) + tx->descp[i] = tx->descs[i]; + return 0; +enomem: + __sdma_txclean(dd, tx); + return -ENOMEM; +} + +#endif /* __HFI1_ORIG__ */ +/* + * ext_coal_sdma_tx_descs() - extend or coalesce sdma tx descriptors + * + * This is called once the initial nominal allocation of descriptors + * in the sdma_txreq is exhausted. + * + * This function calls _extend_sdma_tx_descs to extend or allocate + * coalesce buffer. If there is a allocated coalesce buffer, it will + * copy the input packet data into the coalesce buffer. It also adds + * coalesce buffer descriptor once when whole packet is received. + * + * Return: + * <0 - error + * 0 - coalescing, don't populate descriptor + * 1 - continue with populating descriptor + */ +int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx, + int type, void *kvaddr, struct page *page, + unsigned long offset, u16 len) +{ +//TODO: ext_coal_sdma_tx_descs +#ifdef __HFI1_ORIG__ + int pad_len, rval; + dma_addr_t addr; + + rval = _extend_sdma_tx_descs(dd, tx); + if (rval) { + __sdma_txclean(dd, tx); + return rval; + } + + /* If coalesce buffer is allocated, copy data into it */ + if (tx->coalesce_buf) { + if (type == SDMA_MAP_NONE) { + __sdma_txclean(dd, tx); + return -EINVAL; + } + + if (type == SDMA_MAP_PAGE) { + kvaddr = kmap(page); + kvaddr += offset; + } else if (WARN_ON(!kvaddr)) { + __sdma_txclean(dd, tx); + return -EINVAL; + } + + memcpy(tx->coalesce_buf + tx->coalesce_idx, kvaddr, len); + tx->coalesce_idx += len; + if (type == SDMA_MAP_PAGE) + kunmap(page); + + /* If there is more data, return */ + if (tx->tlen - tx->coalesce_idx) + return 0; + + /* Whole packet is received; add any padding */ + pad_len = tx->packet_len & (sizeof(u32) - 1); + if (pad_len) { + pad_len = sizeof(u32) - pad_len; + memset(tx->coalesce_buf + tx->coalesce_idx, 0, pad_len); + /* padding is taken care of for coalescing case */ + tx->packet_len += pad_len; + tx->tlen += pad_len; + } + + /* dma map the coalesce buffer */ + addr = dma_map_single(&dd->pcidev->dev, + tx->coalesce_buf, + tx->tlen, + DMA_TO_DEVICE); + + if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) { + __sdma_txclean(dd, tx); + return -ENOSPC; + } + + /* Add descriptor for coalesce buffer */ + tx->desc_limit = MAX_DESC; + return _sdma_txadd_daddr(dd, SDMA_MAP_SINGLE, tx, + addr, tx->tlen); + } +#endif /* __HFI1_ORIG__ */ + return 1; +} +#ifdef __HFI1_ORIG__ + +/* Update sdes when the lmc changes */ +void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid) +{ + struct sdma_engine *sde; + int i; + u64 sreg; + + sreg = ((mask & SD(CHECK_SLID_MASK_MASK)) << + SD(CHECK_SLID_MASK_SHIFT)) | + (((lid & mask) & SD(CHECK_SLID_VALUE_MASK)) << + SD(CHECK_SLID_VALUE_SHIFT)); + + for (i = 0; i < dd->num_sdma; i++) { + hfi1_cdbg(LINKVERB, "SendDmaEngine[%d].SLID_CHECK = 0x%x", + i, (u32)sreg); + sde = &dd->per_sdma[i]; + write_sde_csr(sde, SD(CHECK_SLID), sreg); + } +} + +/* tx not dword sized - pad */ +int _pad_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx) +{ + int rval = 0; + + tx->num_desc++; + if ((unlikely(tx->num_desc == tx->desc_limit))) { + rval = _extend_sdma_tx_descs(dd, tx); + if (rval) { + __sdma_txclean(dd, tx); + return rval; + } + } + /* finish the one just added */ + make_tx_sdma_desc( + tx, + SDMA_MAP_NONE, + dd->sdma_pad_phys, + sizeof(u32) - (tx->packet_len & (sizeof(u32) - 1))); + _sdma_close_tx(dd, tx); + return rval; +} + +#endif /* __HFI1_ORIG__ */ +/* + * Add ahg to the sdma_txreq + * + * The logic will consume up to 3 + * descriptors at the beginning of + * sdma_txreq. + */ +void _sdma_txreq_ahgadd( + struct sdma_txreq *tx, + u8 num_ahg, + u8 ahg_entry, + u32 *ahg, + u8 ahg_hlen) +{ + u32 i, shift = 0, desc = 0; + u8 mode; + + WARN_ON_ONCE(num_ahg > 9 || (ahg_hlen & 3) || ahg_hlen == 4); + /* compute mode */ + if (num_ahg == 1) + mode = SDMA_AHG_APPLY_UPDATE1; + else if (num_ahg <= 5) + mode = SDMA_AHG_APPLY_UPDATE2; + else + mode = SDMA_AHG_APPLY_UPDATE3; + tx->num_desc++; + /* initialize to consumed descriptors to zero */ + switch (mode) { + case SDMA_AHG_APPLY_UPDATE3: + tx->num_desc++; + tx->descs[2].qw[0] = 0; + tx->descs[2].qw[1] = 0; + /* FALLTHROUGH */ + case SDMA_AHG_APPLY_UPDATE2: + tx->num_desc++; + tx->descs[1].qw[0] = 0; + tx->descs[1].qw[1] = 0; + break; + } + ahg_hlen >>= 2; + tx->descs[0].qw[1] |= + (((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK) + << SDMA_DESC1_HEADER_INDEX_SHIFT) | + (((u64)ahg_hlen & SDMA_DESC1_HEADER_DWS_MASK) + << SDMA_DESC1_HEADER_DWS_SHIFT) | + (((u64)mode & SDMA_DESC1_HEADER_MODE_MASK) + << SDMA_DESC1_HEADER_MODE_SHIFT) | + (((u64)ahg[0] & SDMA_DESC1_HEADER_UPDATE1_MASK) + << SDMA_DESC1_HEADER_UPDATE1_SHIFT); + for (i = 0; i < (num_ahg - 1); i++) { + if (!shift && !(i & 2)) + desc++; + tx->descs[desc].qw[!!(i & 2)] |= + (((u64)ahg[i + 1]) + << shift); + shift = (shift + 32) & 63; + } +} + +/** + * sdma_ahg_alloc - allocate an AHG entry + * @sde: engine to allocate from + * + * Return: + * 0-31 when successful, -EOPNOTSUPP if AHG is not enabled, + * -ENOSPC if an entry is not available + */ +int sdma_ahg_alloc(struct sdma_engine *sde) +{ + int nr; + int oldbit; + + if (!sde) { + trace_hfi1_ahg_allocate(sde, -EINVAL); + return -EINVAL; + } + while (1) { + nr = ffz(ACCESS_ONCE(sde->ahg_bits)); + if (nr > 31) { + trace_hfi1_ahg_allocate(sde, -ENOSPC); + return -ENOSPC; + } + oldbit = test_and_set_bit(nr, &sde->ahg_bits); + if (!oldbit) + break; + cpu_relax(); + } + trace_hfi1_ahg_allocate(sde, nr); + return nr; +} + +/** + * sdma_ahg_free - free an AHG entry + * @sde: engine to return AHG entry + * @ahg_index: index to free + * + * This routine frees the indicate AHG entry. + */ +void sdma_ahg_free(struct sdma_engine *sde, int ahg_index) +{ + if (!sde) + return; + trace_hfi1_ahg_deallocate(sde, ahg_index); + if (ahg_index < 0 || ahg_index > 31) + return; + clear_bit(ahg_index, &sde->ahg_bits); +} +#ifdef __HFI1_ORIG__ + +/* + * SPC freeze handling for SDMA engines. Called when the driver knows + * the SPC is going into a freeze but before the freeze is fully + * settled. Generally an error interrupt. + * + * This event will pull the engine out of running so no more entries can be + * added to the engine's queue. + */ +void sdma_freeze_notify(struct hfi1_devdata *dd, int link_down) +{ + int i; + enum sdma_events event = link_down ? sdma_event_e85_link_down : + sdma_event_e80_hw_freeze; + + /* set up the wait but do not wait here */ + atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma); + + /* tell all engines to stop running and wait */ + for (i = 0; i < dd->num_sdma; i++) + sdma_process_event(&dd->per_sdma[i], event); + + /* sdma_freeze() will wait for all engines to have stopped */ +} + +/* + * SPC freeze handling for SDMA engines. Called when the driver knows + * the SPC is fully frozen. + */ +void sdma_freeze(struct hfi1_devdata *dd) +{ + int i; + int ret; + + /* + * Make sure all engines have moved out of the running state before + * continuing. + */ + ret = wait_event_interruptible(dd->sdma_unfreeze_wq, + atomic_read(&dd->sdma_unfreeze_count) <= + 0); + /* interrupted or count is negative, then unloading - just exit */ + if (ret || atomic_read(&dd->sdma_unfreeze_count) < 0) + return; + + /* set up the count for the next wait */ + atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma); + + /* tell all engines that the SPC is frozen, they can start cleaning */ + for (i = 0; i < dd->num_sdma; i++) + sdma_process_event(&dd->per_sdma[i], sdma_event_e81_hw_frozen); + + /* + * Wait for everyone to finish software clean before exiting. The + * software clean will read engine CSRs, so must be completed before + * the next step, which will clear the engine CSRs. + */ + (void)wait_event_interruptible(dd->sdma_unfreeze_wq, + atomic_read(&dd->sdma_unfreeze_count) <= 0); + /* no need to check results - done no matter what */ +} + +/* + * SPC freeze handling for the SDMA engines. Called after the SPC is unfrozen. + * + * The SPC freeze acts like a SDMA halt and a hardware clean combined. All + * that is left is a software clean. We could do it after the SPC is fully + * frozen, but then we'd have to add another state to wait for the unfreeze. + * Instead, just defer the software clean until the unfreeze step. + */ +void sdma_unfreeze(struct hfi1_devdata *dd) +{ + int i; + + /* tell all engines start freeze clean up */ + for (i = 0; i < dd->num_sdma; i++) + sdma_process_event(&dd->per_sdma[i], + sdma_event_e82_hw_unfreeze); +} + +/** + * _sdma_engine_progress_schedule() - schedule progress on engine + * @sde: sdma_engine to schedule progress + * + */ +void _sdma_engine_progress_schedule( + struct sdma_engine *sde) +{ + trace_hfi1_sdma_engine_progress(sde, sde->progress_mask); + /* assume we have selected a good cpu */ + write_csr(sde->dd, + CCE_INT_FORCE + (8 * (IS_SDMA_START / 64)), + sde->progress_mask); +} + +#endif /* __HFI1_ORIG__ */ diff --git a/kernel/user_sdma.c b/kernel/user_sdma.c new file mode 100644 index 00000000..aae53e5e --- /dev/null +++ b/kernel/user_sdma.c @@ -0,0 +1,1746 @@ +/* + * Copyright(c) 2015, 2016 Intel Corporation. + * + * This file is provided under a dual BSD/GPLv2 license. When using or + * redistributing this file, you may do so under either license. + * + * GPL LICENSE SUMMARY + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License for more details. + * + * BSD LICENSE + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in + * the documentation and/or other materials provided with the + * distribution. + * - Neither the name of Intel Corporation nor the names of its + * contributors may be used to endorse or promote products derived + * from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#include +#include +#include + +#ifdef __HFI1_ORIG__ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "hfi.h" +#include "sdma.h" +#include "user_sdma.h" +#include "verbs.h" /* for the headers */ +#include "common.h" /* for struct hfi1_tid_info */ +#include "trace.h" +#include "mmu_rb.h" + + +module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO); +MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128"); +#endif /* __HFI1_ORIG__ */ +static uint hfi1_sdma_comp_ring_size = 128; + +/* The maximum number of Data io vectors per message/request */ +#define MAX_VECTORS_PER_REQ 8 +/* + * Maximum number of packet to send from each message/request + * before moving to the next one. + */ +#define MAX_PKTS_PER_QUEUE 16 + +#define num_pages(x) (1 + ((((x) - 1) & PAGE_MASK) >> PAGE_SHIFT)) + +#define req_opcode(x) \ + (((x) >> HFI1_SDMA_REQ_OPCODE_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK) +#define req_version(x) \ + (((x) >> HFI1_SDMA_REQ_VERSION_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK) +#define req_iovcnt(x) \ + (((x) >> HFI1_SDMA_REQ_IOVCNT_SHIFT) & HFI1_SDMA_REQ_IOVCNT_MASK) + +#define PBC2LRH(x) ((((x) & 0xfff) << 2) - 4) +#define LRH2PBC(x) ((((x) >> 2) + 1) & 0xfff) + +#define AHG_HEADER_SET(arr, idx, dw, bit, width, value) \ + do { \ + if ((idx) < ARRAY_SIZE((arr))) \ + (arr)[(idx++)] = sdma_build_ahg_descriptor( \ + (__force u16)(value), (dw), (bit), \ + (width)); \ + else \ + return -ERANGE; \ + } while (0) + +/* KDETH OM multipliers and switch over point */ +#define KDETH_OM_SMALL 4 +#define KDETH_OM_LARGE 64 +#define KDETH_OM_MAX_SIZE (1 << ((KDETH_OM_LARGE / KDETH_OM_SMALL) + 1)) + +/* Tx request flag bits */ +#define TXREQ_FLAGS_REQ_ACK BIT(0) /* Set the ACK bit in the header */ +#define TXREQ_FLAGS_REQ_DISABLE_SH BIT(1) /* Disable header suppression */ + +/* SDMA request flag bits */ +#define SDMA_REQ_FOR_THREAD 1 +#define SDMA_REQ_SEND_DONE 2 +#define SDMA_REQ_HAVE_AHG 3 +#define SDMA_REQ_HAS_ERROR 4 +#define SDMA_REQ_DONE_ERROR 5 + +#define SDMA_PKT_Q_INACTIVE BIT(0) +#define SDMA_PKT_Q_ACTIVE BIT(1) +#define SDMA_PKT_Q_DEFERRED BIT(2) + +/* + * Maximum retry attempts to submit a TX request + * before putting the process to sleep. + */ +#define MAX_DEFER_RETRY_COUNT 1 + +static unsigned initial_pkt_count = 8; + +#define SDMA_IOWAIT_TIMEOUT 1000 /* in milliseconds */ + +#ifdef __HFI1_ORIG__ + +struct sdma_mmu_node; + +#endif /* __HFI1_ORIG__ */ +struct user_sdma_iovec { + struct list_head list; + struct iovec iov; + /* number of pages in this vector */ + unsigned npages; + /* array of pinned pages for this vector */ + struct page **pages; + /* + * offset into the virtual address space of the vector at + * which we last left off. + */ + u64 offset; + struct sdma_mmu_node *node; +}; +#ifdef __HFI1_ORIG__ + +struct sdma_mmu_node { + struct mmu_rb_node rb; + struct hfi1_user_sdma_pkt_q *pq; + atomic_t refcount; + struct page **pages; + unsigned npages; +}; + +/* evict operation argument */ +struct evict_data { + u32 cleared; /* count evicted so far */ + u32 target; /* target count to evict */ +}; + +#endif /* __HFI1_ORIG__ */ +struct user_sdma_request { + struct sdma_req_info info; + struct hfi1_user_sdma_pkt_q *pq; + struct hfi1_user_sdma_comp_q *cq; + /* This is the original header from user space */ + struct hfi1_pkt_header hdr; + /* + * Pointer to the SDMA engine for this request. + * Since different request could be on different VLs, + * each request will need it's own engine pointer. + */ + struct sdma_engine *sde; + u8 ahg_idx; + u32 ahg[9]; + /* + * KDETH.Offset (Eager) field + * We need to remember the initial value so the headers + * can be updated properly. + */ + u32 koffset; + /* + * KDETH.OFFSET (TID) field + * The offset can cover multiple packets, depending on the + * size of the TID entry. + */ + u32 tidoffset; + /* + * KDETH.OM + * Remember this because the header template always sets it + * to 0. + */ + u8 omfactor; + /* + * We copy the iovs for this request (based on + * info.iovcnt). These are only the data vectors + */ + unsigned data_iovs; + /* total length of the data in the request */ + u32 data_len; + /* progress index moving along the iovs array */ + unsigned iov_idx; + struct user_sdma_iovec iovs[MAX_VECTORS_PER_REQ]; + /* number of elements copied to the tids array */ + u16 n_tids; + /* TID array values copied from the tid_iov vector */ + u32 *tids; + u16 tididx; + u32 sent; + u64 seqnum; + u64 seqcomp; + u64 seqsubmitted; + struct list_head txps; + unsigned long flags; + /* status of the last txreq completed */ + int status; +}; + +/* + * A single txreq could span up to 3 physical pages when the MTU + * is sufficiently large (> 4K). Each of the IOV pointers also + * needs it's own set of flags so the vector has been handled + * independently of each other. + */ +struct user_sdma_txreq { + /* Packet header for the txreq */ + struct hfi1_pkt_header hdr; + struct sdma_txreq txreq; + struct list_head list; + struct user_sdma_request *req; + u16 flags; + unsigned busycount; + u64 seqnum; +}; +#ifdef __HFI1_ORIG__ + +#define SDMA_DBG(req, fmt, ...) \ + hfi1_cdbg(SDMA, "[%u:%u:%u:%u] " fmt, (req)->pq->dd->unit, \ + (req)->pq->ctxt, (req)->pq->subctxt, (req)->info.comp_idx, \ + ##__VA_ARGS__) +#define SDMA_Q_DBG(pq, fmt, ...) \ + hfi1_cdbg(SDMA, "[%u:%u:%u] " fmt, (pq)->dd->unit, (pq)->ctxt, \ + (pq)->subctxt, ##__VA_ARGS__) + +#endif /* __HFI1_ORIG__ */ + +static int user_sdma_send_pkts(struct user_sdma_request *, unsigned); +static inline void pq_update(struct hfi1_user_sdma_pkt_q *); +static int check_header_template(struct user_sdma_request *, + struct hfi1_pkt_header *, u32, u32); +static int set_txreq_header(struct user_sdma_request *, + struct user_sdma_txreq *, u32); +static int set_txreq_header_ahg(struct user_sdma_request *, + struct user_sdma_txreq *, u32); +static void user_sdma_free_request(struct user_sdma_request *, bool); +static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *, + struct hfi1_user_sdma_comp_q *, + u16, enum hfi1_sdma_comp_state, int); +static void user_sdma_txreq_cb(struct sdma_txreq *, int); + +#ifdef __HFI1_ORIG__ +static int num_user_pages(const struct iovec *); +static int pin_vector_pages(struct user_sdma_request *, + struct user_sdma_iovec *); +static void unpin_vector_pages(struct mm_struct *, struct page **, unsigned, + unsigned); +static inline u32 set_pkt_bth_psn(__be32, u8, u32); +static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len); + +static int defer_packet_queue( + struct sdma_engine *, + struct iowait_work *, + struct sdma_txreq *, + unsigned seq); +static void activate_packet_queue(struct iowait *, int); +static bool sdma_rb_filter(struct mmu_rb_node *, unsigned long, unsigned long); +static int sdma_rb_insert(void *, struct mmu_rb_node *); +static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode, + void *arg2, bool *stop); +static void sdma_rb_remove(void *, struct mmu_rb_node *); +static int sdma_rb_invalidate(void *, struct mmu_rb_node *); + +static struct mmu_rb_ops sdma_rb_ops = { + .filter = sdma_rb_filter, + .insert = sdma_rb_insert, + .evict = sdma_rb_evict, + .remove = sdma_rb_remove, + .invalidate = sdma_rb_invalidate +}; + +static int defer_packet_queue( + struct sdma_engine *sde, + struct iowait_work *wait, + struct sdma_txreq *txreq, + unsigned seq) +{ + struct hfi1_user_sdma_pkt_q *pq = + container_of(wait->iow, struct hfi1_user_sdma_pkt_q, busy); + struct hfi1_ibdev *dev = &pq->dd->verbs_dev; + struct user_sdma_txreq *tx = + container_of(txreq, struct user_sdma_txreq, txreq); + + if (sdma_progress(sde, seq, txreq)) { + if (tx->busycount++ < MAX_DEFER_RETRY_COUNT) + goto eagain; + } + /* + * We are assuming that if the list is enqueued somewhere, it + * is to the dmawait list since that is the only place where + * it is supposed to be enqueued. + */ + xchg(&pq->state, SDMA_PKT_Q_DEFERRED); + write_seqlock(&dev->iowait_lock); + if (list_empty(&pq->busy.list)) + list_add_tail(&pq->busy.list, &sde->dmawait); + write_sequnlock(&dev->iowait_lock); + return -EBUSY; +eagain: + return -EAGAIN; +} + +static void activate_packet_queue(struct iowait *wait, int reason) +{ + struct hfi1_user_sdma_pkt_q *pq = + container_of(wait, struct hfi1_user_sdma_pkt_q, busy); + xchg(&pq->state, SDMA_PKT_Q_ACTIVE); + wake_up(&wait->wait_dma); +}; + +static void sdma_kmem_cache_ctor(void *obj) +{ + struct user_sdma_txreq *tx = obj; + + memset(tx, 0, sizeof(*tx)); +} + +int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt, struct file *fp) +{ + struct hfi1_filedata *fd; + int ret = 0; + unsigned memsize; + char buf[64]; + struct hfi1_devdata *dd; + struct hfi1_user_sdma_comp_q *cq; + struct hfi1_user_sdma_pkt_q *pq; + unsigned long flags; + + if (!uctxt || !fp) { + ret = -EBADF; + goto done; + } + + fd = fp->private_data; + + if (!hfi1_sdma_comp_ring_size) { + ret = -EINVAL; + goto done; + } + + dd = uctxt->dd; + + pq = kzalloc(sizeof(*pq), GFP_KERNEL); + if (!pq) + goto pq_nomem; + + memsize = sizeof(*pq->reqs) * hfi1_sdma_comp_ring_size; + pq->reqs = kzalloc(memsize, GFP_KERNEL); + if (!pq->reqs) + goto pq_reqs_nomem; + + memsize = BITS_TO_LONGS(hfi1_sdma_comp_ring_size) * sizeof(long); + pq->req_in_use = kzalloc(memsize, GFP_KERNEL); + if (!pq->req_in_use) + goto pq_reqs_no_in_use; + + INIT_LIST_HEAD(&pq->list); + pq->dd = dd; + pq->ctxt = uctxt->ctxt; + pq->subctxt = fd->subctxt; + pq->n_max_reqs = hfi1_sdma_comp_ring_size; + pq->state = SDMA_PKT_Q_INACTIVE; + atomic_set(&pq->n_reqs, 0); + init_waitqueue_head(&pq->wait); + atomic_set(&pq->n_locked, 0); + pq->mm = fd->mm; + + iowait_init(&pq->busy, 0, NULL, NULL, defer_packet_queue, + activate_packet_queue, NULL); + pq->reqidx = 0; + snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt, + fd->subctxt); + pq->txreq_cache = kmem_cache_create(buf, + sizeof(struct user_sdma_txreq), + L1_CACHE_BYTES, + SLAB_HWCACHE_ALIGN, + sdma_kmem_cache_ctor); + if (!pq->txreq_cache) { + dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n", + uctxt->ctxt); + goto pq_txreq_nomem; + } + fd->pq = pq; + cq = kzalloc(sizeof(*cq), GFP_KERNEL); + if (!cq) + goto cq_nomem; + + memsize = PAGE_ALIGN(sizeof(*cq->comps) * hfi1_sdma_comp_ring_size); + cq->comps = vmalloc_user(memsize); + if (!cq->comps) + goto cq_comps_nomem; + + cq->nentries = hfi1_sdma_comp_ring_size; + fd->cq = cq; + + ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq, + &pq->handler); + if (ret) { + dd_dev_err(dd, "Failed to register with MMU %d", ret); + goto done; + } + + spin_lock_irqsave(&uctxt->sdma_qlock, flags); + list_add(&pq->list, &uctxt->sdma_queues); + spin_unlock_irqrestore(&uctxt->sdma_qlock, flags); + goto done; + +cq_comps_nomem: + kfree(cq); +cq_nomem: + kmem_cache_destroy(pq->txreq_cache); +pq_txreq_nomem: + kfree(pq->req_in_use); +pq_reqs_no_in_use: + kfree(pq->reqs); +pq_reqs_nomem: + kfree(pq); + fd->pq = NULL; +pq_nomem: + ret = -ENOMEM; +done: + return ret; +} + +int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd) +{ + struct hfi1_ctxtdata *uctxt = fd->uctxt; + struct hfi1_user_sdma_pkt_q *pq; + unsigned long flags; + + hfi1_cdbg(SDMA, "[%u:%u:%u] Freeing user SDMA queues", uctxt->dd->unit, + uctxt->ctxt, fd->subctxt); + pq = fd->pq; + if (pq) { + if (pq->handler) + hfi1_mmu_rb_unregister(pq->handler); + spin_lock_irqsave(&uctxt->sdma_qlock, flags); + if (!list_empty(&pq->list)) + list_del_init(&pq->list); + spin_unlock_irqrestore(&uctxt->sdma_qlock, flags); + iowait_sdma_drain(&pq->busy); + /* Wait until all requests have been freed. */ + wait_event_interruptible( + pq->wait, + (ACCESS_ONCE(pq->state) == SDMA_PKT_Q_INACTIVE)); + kfree(pq->reqs); + kfree(pq->req_in_use); + kmem_cache_destroy(pq->txreq_cache); + kfree(pq); + fd->pq = NULL; + } + if (fd->cq) { + vfree(fd->cq->comps); + kfree(fd->cq); + fd->cq = NULL; + } + return 0; +} + +#endif /* __HFI1_ORIG__ */ + +static u8 dlid_to_selector(u16 dlid) +{ + static u8 mapping[256]; + static int initialized; + static u8 next; + int hash; + + if (!initialized) { + memset(mapping, 0xFF, 256); + initialized = 1; + } + + hash = ((dlid >> 8) ^ dlid) & 0xFF; + if (mapping[hash] == 0xFF) { + mapping[hash] = next; + next = (next + 1) & 0x7F; + } + + return mapping[hash]; +} + +#ifdef __HFI1_ORIG__ +int hfi1_user_sdma_process_request(struct file *fp, struct iovec *iovec, + unsigned long dim, unsigned long *count) +{ + int ret = 0, i; + struct hfi1_filedata *fd = fp->private_data; +#else +int hfi1_user_sdma_process_request(void *private_data, struct iovec *iovec, + unsigned long dim, unsigned long *count) +{ + int ret = 0, i; + struct hfi1_filedata *fd = private_data; +#endif /* __HFI1_ORIG__ */ + struct hfi1_ctxtdata *uctxt = fd->uctxt; + struct hfi1_user_sdma_pkt_q *pq = fd->pq; + struct hfi1_user_sdma_comp_q *cq = fd->cq; + struct hfi1_devdata *dd = pq->dd; + unsigned long idx = 0; + u8 pcount = initial_pkt_count; + struct sdma_req_info info; + struct user_sdma_request *req; + u8 opcode, sc, vl; + int req_queued = 0; + u16 dlid; + u32 selector; + + hfi1_cdbg(AIOWRITE, "+"); + if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) { + hfi1_cdbg( + SDMA, + "[%u:%u:%u] First vector not big enough for header %lu/%lu", + dd->unit, uctxt->ctxt, fd->subctxt, + iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr)); + return -EINVAL; + } + ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info)); + if (ret) { + hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)", + dd->unit, uctxt->ctxt, fd->subctxt, ret); + return -EFAULT; + } + + // trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt, + // (u16 *)&info); + + if (info.comp_idx >= hfi1_sdma_comp_ring_size) { + hfi1_cdbg(SDMA, + "[%u:%u:%u:%u] Invalid comp index", + dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx); + return -EINVAL; + } + + /* + * Sanity check the header io vector count. Need at least 1 vector + * (header) and cannot be larger than the actual io vector count. + */ + if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) { + hfi1_cdbg(SDMA, + "[%u:%u:%u:%u] Invalid iov count %d, dim %ld", + dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx, + req_iovcnt(info.ctrl), dim); + return -EINVAL; + } + + if (!info.fragsize) { + hfi1_cdbg(SDMA, + "[%u:%u:%u:%u] Request does not specify fragsize", + dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx); + return -EINVAL; + } + + /* Try to claim the request. */ + if (test_and_set_bit(info.comp_idx, pq->req_in_use)) { + hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use", + dd->unit, uctxt->ctxt, fd->subctxt, + info.comp_idx); + return -EBADSLT; + } + /* + * All safety checks have been done and this request has been claimed. + */ + hfi1_cdbg(SDMA, "[%u:%u:%u] Using req/comp entry %u\n", dd->unit, + uctxt->ctxt, fd->subctxt, info.comp_idx); + req = pq->reqs + info.comp_idx; + memset(req, 0, sizeof(*req)); + req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */ + req->pq = pq; + req->cq = cq; + req->status = -1; + INIT_LIST_HEAD(&req->txps); + + memcpy(&req->info, &info, sizeof(info)); + + /* + * Seems like iovec[0] is always the header and if EXPECTED, + * TID is present is in iovec[dim - 1] + * The rest in between are data. + */ + + if (req_opcode(info.ctrl) == EXPECTED) { + /* expected must have a TID info and at least one data vector */ + if (req->data_iovs < 2) { + SDMA_DBG(req, + "Not enough vectors for expected request"); + ret = -EINVAL; + goto free_req; + } + req->data_iovs--; + } + + if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) { + SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs, + MAX_VECTORS_PER_REQ); + ret = -EINVAL; + goto free_req; + } + /* Copy the header from the user buffer */ + ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info), + sizeof(req->hdr)); + if (ret) { + SDMA_DBG(req, "Failed to copy header template (%d)", ret); + ret = -EFAULT; + goto free_req; + } + + /* If Static rate control is not enabled, sanitize the header. */ + if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL)) + req->hdr.pbc[2] = 0; + + /* Validate the opcode. Do not trust packets from user space blindly. */ + opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff; + if ((opcode & USER_OPCODE_CHECK_MASK) != + USER_OPCODE_CHECK_VAL) { + SDMA_DBG(req, "Invalid opcode (%d)", opcode); + ret = -EINVAL; + goto free_req; + } + +// TODO: Enable this validation and checking +#ifdef __HFI1_ORIG__ + /* + * Validate the vl. Do not trust packets from user space blindly. + * VL comes from PBC, SC comes from LRH, and the VL needs to + * match the SC look up. + */ + vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF; + sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) | + (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4)); + if (vl >= dd->pport->vls_operational || + vl != sc_to_vlt(dd, sc)) { + SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl); + ret = -EINVAL; + goto free_req; + } + + /* Checking P_KEY for requests from user-space */ + if (egress_pkey_check(dd->pport, req->hdr.lrh, req->hdr.bth, sc, + PKEY_CHECK_INVALID)) { + ret = -EINVAL; + goto free_req; + } +#endif /* __HFI1_ORIG__ */ + + /* + * Also should check the BTH.lnh. If it says the next header is GRH then + * the RXE parsing will be off and will land in the middle of the KDETH + * or miss it entirely. + */ + if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) { + SDMA_DBG(req, "User tried to pass in a GRH"); + ret = -EINVAL; + goto free_req; + } + + req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]); + /* + * Calculate the initial TID offset based on the values of + * KDETH.OFFSET and KDETH.OM that are passed in. + */ + req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) * + (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ? + KDETH_OM_LARGE : KDETH_OM_SMALL); + SDMA_DBG(req, "Initial TID offset %u", req->tidoffset); + idx++; + + /* Save all the IO vector structures */ + for (i = 0; i < req->data_iovs; i++) { + INIT_LIST_HEAD(&req->iovs[i].list); + /* + * XXX: iovec is still user-space in McKernel here!! + * + * req->iovs[] contain only the data. + */ + memcpy(&req->iovs[i].iov, iovec + idx++, sizeof(struct iovec)); +#ifdef __HFI1_ORIG__ + hfi1_cdbg(AIOWRITE, "+pin_vector_pages"); + // TODO: pin_vector_pages + ret = pin_vector_pages(req, &req->iovs[i]); + hfi1_cdbg(AIOWRITE, "-pin_vector_pages"); + if (ret) { + req->status = ret; + goto free_req; + } +#endif /* __HFI1_ORIG__ */ + req->data_len += req->iovs[i].iov.iov_len; + } + SDMA_DBG(req, "total data length %u", req->data_len); + + if (pcount > req->info.npkts) + pcount = req->info.npkts; + /* + * Copy any TID info + * User space will provide the TID info only when the + * request type is EXPECTED. This is true even if there is + * only one packet in the request and the header is already + * setup. The reason for the singular TID case is that the + * driver needs to perform safety checks. + */ + if (req_opcode(req->info.ctrl) == EXPECTED) { + u16 ntids = iovec[idx].iov_len / sizeof(*req->tids); + + if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) { + ret = -EINVAL; + goto free_req; + } + req->tids = kcalloc(ntids, sizeof(*req->tids), GFP_KERNEL); + if (!req->tids) { + ret = -ENOMEM; + goto free_req; + } + /* + * We have to copy all of the tids because they may vary + * in size and, therefore, the TID count might not be + * equal to the pkt count. However, there is no way to + * tell at this point. + */ + ret = copy_from_user(req->tids, iovec[idx].iov_base, + ntids * sizeof(*req->tids)); + if (ret) { + SDMA_DBG(req, "Failed to copy %d TIDs (%d)", + ntids, ret); + ret = -EFAULT; + goto free_req; + } + req->n_tids = ntids; + idx++; + } + + dlid = be16_to_cpu(req->hdr.lrh[1]); + selector = dlid_to_selector(dlid); + selector += uctxt->ctxt + fd->subctxt; + /* TODO: check the rcu stuff */ + /* + * XXX: didn't we conclude that we don't need to worry about RCU here? + * the mapping is created at driver initialization, the rest of the + * accesses are read-only + */ + //req->sde = sdma_select_user_engine(dd, selector, vl); + + if (!req->sde || !sdma_running(req->sde)) { + ret = -ECOMM; + goto free_req; + } + + /* We don't need an AHG entry if the request contains only one packet */ + if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG)) { + int ahg = sdma_ahg_alloc(req->sde); + + if (likely(ahg >= 0)) { + req->ahg_idx = (u8)ahg; + set_bit(SDMA_REQ_HAVE_AHG, &req->flags); + } + } + + set_comp_state(pq, cq, info.comp_idx, QUEUED, 0); + atomic_inc(&pq->n_reqs); + req_queued = 1; + /* Send the first N packets in the request to buy us some time */ + ret = user_sdma_send_pkts(req, pcount); + if (unlikely(ret < 0 && ret != -EBUSY)) { + req->status = ret; + goto free_req; + } + + /* + * It is possible that the SDMA engine would have processed all the + * submitted packets by the time we get here. Therefore, only set + * packet queue state to ACTIVE if there are still uncompleted + * requests. + */ + if (atomic_read(&pq->n_reqs)) + xchg(&pq->state, SDMA_PKT_Q_ACTIVE); + + /* + * This is a somewhat blocking send implementation. + * The driver will block the caller until all packets of the + * request have been submitted to the SDMA engine. However, it + * will not wait for send completions. + */ + while (!test_bit(SDMA_REQ_SEND_DONE, &req->flags)) { + ret = user_sdma_send_pkts(req, pcount); + if (ret < 0) { + if (ret != -EBUSY) { + req->status = ret; + set_bit(SDMA_REQ_DONE_ERROR, &req->flags); + if (ACCESS_ONCE(req->seqcomp) == + req->seqsubmitted - 1) + goto free_req; + return ret; + } +#ifdef __HFI1_ORIG__ + hfi1_cdbg(AIOWRITE, "+wait_event_interruptible_timeout"); + wait_event_interruptible_timeout( + pq->busy.wait_dma, + (pq->state == SDMA_PKT_Q_ACTIVE), + msecs_to_jiffies( + SDMA_IOWAIT_TIMEOUT)); + hfi1_cdbg(AIOWRITE, "-wait_event_interruptible_timeout"); +#else + while (pq->state != SDMA_PKT_Q_ACTIVE) cpu_pause(); +#endif /* __HFI1_ORIG__ */ + } + } + *count += idx; + hfi1_cdbg(AIOWRITE, "-"); + return 0; +free_req: + user_sdma_free_request(req, true); + if (req_queued) + pq_update(pq); + set_comp_state(pq, cq, info.comp_idx, ERROR, req->status); + hfi1_cdbg(AIOWRITE, "-"); + return ret; +} + +static inline u32 compute_data_length(struct user_sdma_request *req, + struct user_sdma_txreq *tx) +{ + /* + * Determine the proper size of the packet data. + * The size of the data of the first packet is in the header + * template. However, it includes the header and ICRC, which need + * to be subtracted. + * The minimum representable packet data length in a header is 4 bytes, + * therefore, when the data length request is less than 4 bytes, there's + * only one packet, and the packet data length is equal to that of the + * request data length. + * The size of the remaining packets is the minimum of the frag + * size (MTU) or remaining data in the request. + */ + u32 len; + + if (!req->seqnum) { + if (req->data_len < sizeof(u32)) + len = req->data_len; + else + len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) - + (sizeof(tx->hdr) - 4)); + } else if (req_opcode(req->info.ctrl) == EXPECTED) { + u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) * + PAGE_SIZE; + /* + * Get the data length based on the remaining space in the + * TID pair. + */ + len = min(tidlen - req->tidoffset, (u32)req->info.fragsize); + /* If we've filled up the TID pair, move to the next one. */ + if (unlikely(!len) && ++req->tididx < req->n_tids && + req->tids[req->tididx]) { + tidlen = EXP_TID_GET(req->tids[req->tididx], + LEN) * PAGE_SIZE; + req->tidoffset = 0; + len = min_t(u32, tidlen, req->info.fragsize); + } + /* + * Since the TID pairs map entire pages, make sure that we + * are not going to try to send more data that we have + * remaining. + */ + len = min(len, req->data_len - req->sent); + } else { + len = min(req->data_len - req->sent, (u32)req->info.fragsize); + } + SDMA_DBG(req, "Data Length = %u", len); + return len; +} + +static inline u32 pad_len(u32 len) +{ + if (len & (sizeof(u32) - 1)) + len += sizeof(u32) - (len & (sizeof(u32) - 1)); + return len; +} + +static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len) +{ + /* (Size of complete header - size of PBC) + 4B ICRC + data length */ + return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len); +} + +static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts) +{ + int ret = 0, count; + unsigned npkts = 0; + struct user_sdma_txreq *tx = NULL; + struct hfi1_user_sdma_pkt_q *pq = NULL; + struct user_sdma_iovec *iovec = NULL; + + hfi1_cdbg(AIOWRITE, "+"); + if (!req->pq) + return -EINVAL; + + pq = req->pq; + + /* If tx completion has reported an error, we are done. */ + if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) { + set_bit(SDMA_REQ_DONE_ERROR, &req->flags); + return -EFAULT; + } + + /* + * Check if we might have sent the entire request already + */ + if (unlikely(req->seqnum == req->info.npkts)) { + if (!list_empty(&req->txps)) + goto dosend; + return ret; + } + + if (!maxpkts || maxpkts > req->info.npkts - req->seqnum) + maxpkts = req->info.npkts - req->seqnum; + + while (npkts < maxpkts) { + u32 datalen = 0, queued = 0, data_sent = 0; + u64 iov_offset = 0; + + /* + * Check whether any of the completions have come back + * with errors. If so, we are not going to process any + * more packets from this request. + */ + if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) { + set_bit(SDMA_REQ_DONE_ERROR, &req->flags); + return -EFAULT; + } + +#ifdef __HFI1_ORIG__ + tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL); +#else + tx = kmalloc(sizeof(struct user_sdma_txreq), GFP_KERNEL | __GFP_ZERO); +#endif /* __HFI1_ORIG__ */ + if (!tx) + return -ENOMEM; + + tx->flags = 0; + tx->req = req; + tx->busycount = 0; + INIT_LIST_HEAD(&tx->list); + + /* + * For the last packet set the ACK request + * and disable header suppression. + */ + if (req->seqnum == req->info.npkts - 1) + tx->flags |= (TXREQ_FLAGS_REQ_ACK | + TXREQ_FLAGS_REQ_DISABLE_SH); + + /* + * Calculate the payload size - this is min of the fragment + * (MTU) size or the remaining bytes in the request but only + * if we have payload data. + */ + if (req->data_len) { + iovec = &req->iovs[req->iov_idx]; + if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) { + if (++req->iov_idx == req->data_iovs) { + ret = -EFAULT; + goto free_txreq; + } + iovec = &req->iovs[req->iov_idx]; + WARN_ON(iovec->offset); + } + + datalen = compute_data_length(req, tx); + + /* + * Disable header suppression for the payload <= 8DWS. + * If there is an uncorrectable error in the receive + * data FIFO when the received payload size is less than + * or equal to 8DWS then the RxDmaDataFifoRdUncErr is + * not reported.There is set RHF.EccErr if the header + * is not suppressed. + */ + if (!datalen) { + SDMA_DBG(req, + "Request has data but pkt len is 0"); + ret = -EFAULT; + goto free_tx; + } else if (datalen <= 32) { + tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH; + } + } + + if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags)) { + if (!req->seqnum) { + u16 pbclen = le16_to_cpu(req->hdr.pbc[0]); + u32 lrhlen = get_lrh_len(req->hdr, + pad_len(datalen)); + /* + * Copy the request header into the tx header + * because the HW needs a cacheline-aligned + * address. + * This copy can be optimized out if the hdr + * member of user_sdma_request were also + * cacheline aligned. + */ + memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr)); + if (PBC2LRH(pbclen) != lrhlen) { + pbclen = (pbclen & 0xf000) | + LRH2PBC(lrhlen); + tx->hdr.pbc[0] = cpu_to_le16(pbclen); + } + ret = check_header_template(req, &tx->hdr, + lrhlen, datalen); + if (ret) + goto free_tx; + ret = sdma_txinit_ahg(&tx->txreq, + SDMA_TXREQ_F_AHG_COPY, + sizeof(tx->hdr) + datalen, + req->ahg_idx, 0, NULL, 0, + user_sdma_txreq_cb); + if (ret) + goto free_tx; + ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq, + &tx->hdr, + sizeof(tx->hdr)); + if (ret) + goto free_txreq; + } else { + int changes; + + changes = set_txreq_header_ahg(req, tx, + datalen); + if (changes < 0) + goto free_tx; + sdma_txinit_ahg(&tx->txreq, + SDMA_TXREQ_F_USE_AHG, + datalen, req->ahg_idx, changes, + req->ahg, sizeof(req->hdr), + user_sdma_txreq_cb); + } + } else { + ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) + + datalen, user_sdma_txreq_cb); + if (ret) + goto free_tx; + /* + * Modify the header for this packet. This only needs + * to be done if we are not going to use AHG. Otherwise, + * the HW will do it based on the changes we gave it + * during sdma_txinit_ahg(). + */ + ret = set_txreq_header(req, tx, datalen); + if (ret) + goto free_txreq; + } + + /* + * If the request contains any data vectors, add up to + * fragsize bytes to the descriptor. + */ + while (queued < datalen && + (req->sent + data_sent) < req->data_len) { + unsigned long base, offset; + unsigned pageidx, len; + + /* + * XXX: this seems to operate with the assumption + * that all user buffers need to be processed in + * PAGE_SIZE steps, how about large pages? + */ + base = (unsigned long)iovec->iov.iov_base; + offset = offset_in_page(base + iovec->offset + + iov_offset); + pageidx = (((iovec->offset + iov_offset + + base) - (base & PAGE_MASK)) >> PAGE_SHIFT); + len = offset + req->info.fragsize > PAGE_SIZE ? + PAGE_SIZE - offset : req->info.fragsize; + len = min((datalen - queued), len); + ret = sdma_txadd_page(pq->dd, &tx->txreq, + iovec->pages[pageidx], + offset, len); + if (ret) { + SDMA_DBG(req, "SDMA txreq add page failed %d\n", + ret); + goto free_txreq; + } + iov_offset += len; + queued += len; + data_sent += len; + if (unlikely(queued < datalen && + pageidx == iovec->npages && + req->iov_idx < req->data_iovs - 1)) { + iovec->offset += iov_offset; + iovec = &req->iovs[++req->iov_idx]; + iov_offset = 0; + } + } + /* + * The txreq was submitted successfully so we can update + * the counters. + */ + req->koffset += datalen; + if (req_opcode(req->info.ctrl) == EXPECTED) + req->tidoffset += datalen; + req->sent += data_sent; + if (req->data_len) + iovec->offset += iov_offset; + list_add_tail(&tx->txreq.list, &req->txps); + /* + * It is important to increment this here as it is used to + * generate the BTH.PSN and, therefore, can't be bulk-updated + * outside of the loop. + */ + tx->seqnum = req->seqnum++; + npkts++; + } +dosend: + ret = sdma_send_txlist(req->sde, + iowait_get_ib_work(&pq->busy), + &req->txps, &count); + req->seqsubmitted += count; + if (req->seqsubmitted == req->info.npkts) { + set_bit(SDMA_REQ_SEND_DONE, &req->flags); + /* + * The txreq has already been submitted to the HW queue + * so we can free the AHG entry now. Corruption will not + * happen due to the sequential manner in which + * descriptors are processed. + */ + if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags)) + sdma_ahg_free(req->sde, req->ahg_idx); + } + hfi1_cdbg(AIOWRITE, "-"); + return ret; + +free_txreq: + sdma_txclean(pq->dd, &tx->txreq); +free_tx: +#ifdef __HFI1_ORIG__ + kmem_cache_free(pq->txreq_cache, tx); + hfi1_cdbg(AIOWRITE, "-"); +#else + kfree(tx); +#endif /* __HFI1_ORIG__ */ + return ret; +} +#ifdef __HFI1_ORIG__ + +/* + * How many pages in this iovec element? + */ +static inline int num_user_pages(const struct iovec *iov) +{ + const unsigned long addr = (unsigned long)iov->iov_base; + const unsigned long len = iov->iov_len; + const unsigned long spage = addr & PAGE_MASK; + const unsigned long epage = (addr + len - 1) & PAGE_MASK; + + return 1 + ((epage - spage) >> PAGE_SHIFT); +} + +static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages) +{ + struct evict_data evict_data; + + evict_data.cleared = 0; + evict_data.target = npages; + hfi1_mmu_rb_evict(pq->handler, &evict_data); + return evict_data.cleared; +} + +static int pin_vector_pages(struct user_sdma_request *req, + struct user_sdma_iovec *iovec) +{ + int ret = 0, pinned, npages, cleared; + struct page **pages; + struct hfi1_user_sdma_pkt_q *pq = req->pq; + struct sdma_mmu_node *node = NULL; + struct mmu_rb_node *rb_node; + + rb_node = hfi1_mmu_rb_extract(pq->handler, + (unsigned long)iovec->iov.iov_base, + iovec->iov.iov_len); + if (rb_node) + node = container_of(rb_node, struct sdma_mmu_node, rb); + else + rb_node = NULL; + + if (!node) { + node = kzalloc(sizeof(*node), GFP_KERNEL); + if (!node) + return -ENOMEM; + + node->rb.addr = (unsigned long)iovec->iov.iov_base; + node->pq = pq; + atomic_set(&node->refcount, 0); + } + + npages = num_user_pages(&iovec->iov); + if (node->npages < npages) { + pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL); + if (!pages) { + SDMA_DBG(req, "Failed page array alloc"); + ret = -ENOMEM; + goto bail; + } + memcpy(pages, node->pages, node->npages * sizeof(*pages)); + + npages -= node->npages; + +retry: + if (!hfi1_can_pin_pages(pq->dd, pq->mm, + atomic_read(&pq->n_locked), npages)) { + cleared = sdma_cache_evict(pq, npages); + if (cleared >= npages) + goto retry; + } + pinned = hfi1_acquire_user_pages(pq->mm, + ((unsigned long)iovec->iov.iov_base + + (node->npages * PAGE_SIZE)), npages, 0, + pages + node->npages); + if (pinned < 0) { + kfree(pages); + ret = pinned; + goto bail; + } + if (pinned != npages) { + unpin_vector_pages(pq->mm, pages, node->npages, + pinned); + ret = -EFAULT; + goto bail; + } + kfree(node->pages); + node->rb.len = iovec->iov.iov_len; + node->pages = pages; + node->npages += pinned; + npages = node->npages; + atomic_add(pinned, &pq->n_locked); + } + iovec->pages = node->pages; + iovec->npages = npages; + iovec->node = node; + + ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb); + if (ret) { + atomic_sub(node->npages, &pq->n_locked); + iovec->node = NULL; + goto bail; + } + return 0; +bail: + if (rb_node) + unpin_vector_pages(pq->mm, node->pages, 0, node->npages); + kfree(node); + return ret; +} + +static void unpin_vector_pages(struct mm_struct *mm, struct page **pages, + unsigned start, unsigned npages) +{ + hfi1_release_user_pages(mm, pages + start, npages, false); + kfree(pages); +} + +#endif /* __HFI1_ORIG__ */ +static int check_header_template(struct user_sdma_request *req, + struct hfi1_pkt_header *hdr, u32 lrhlen, + u32 datalen) +{ + /* + * Perform safety checks for any type of packet: + * - transfer size is multiple of 64bytes + * - packet length is multiple of 4 bytes + * - packet length is not larger than MTU size + * + * These checks are only done for the first packet of the + * transfer since the header is "given" to us by user space. + * For the remainder of the packets we compute the values. + */ + if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 || + lrhlen > get_lrh_len(*hdr, req->info.fragsize)) + return -EINVAL; + + if (req_opcode(req->info.ctrl) == EXPECTED) { + /* + * The header is checked only on the first packet. Furthermore, + * we ensure that at least one TID entry is copied when the + * request is submitted. Therefore, we don't have to verify that + * tididx points to something sane. + */ + u32 tidval = req->tids[req->tididx], + tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE, + tididx = EXP_TID_GET(tidval, IDX), + tidctrl = EXP_TID_GET(tidval, CTRL), + tidoff; + __le32 kval = hdr->kdeth.ver_tid_offset; + + tidoff = KDETH_GET(kval, OFFSET) * + (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ? + KDETH_OM_LARGE : KDETH_OM_SMALL); + /* + * Expected receive packets have the following + * additional checks: + * - offset is not larger than the TID size + * - TIDCtrl values match between header and TID array + * - TID indexes match between header and TID array + */ + if ((tidoff + datalen > tidlen) || + KDETH_GET(kval, TIDCTRL) != tidctrl || + KDETH_GET(kval, TID) != tididx) + return -EINVAL; + } + return 0; +} + +/* + * Correctly set the BTH.PSN field based on type of + * transfer - eager packets can just increment the PSN but + * expected packets encode generation and sequence in the + * BTH.PSN field so just incrementing will result in errors. + */ +static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags) +{ + u32 val = be32_to_cpu(bthpsn), + mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull : + 0xffffffull), + psn = val & mask; + if (expct) + psn = (psn & ~HFI1_KDETH_BTH_SEQ_MASK) | + ((psn + frags) & HFI1_KDETH_BTH_SEQ_MASK); + else + psn = psn + frags; + return psn & mask; +} + +static int set_txreq_header(struct user_sdma_request *req, + struct user_sdma_txreq *tx, u32 datalen) +{ + struct hfi1_user_sdma_pkt_q *pq = req->pq; + struct hfi1_pkt_header *hdr = &tx->hdr; + u16 pbclen; + int ret; + u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen)); + + /* Copy the header template to the request before modification */ + memcpy(hdr, &req->hdr, sizeof(*hdr)); + + /* + * Check if the PBC and LRH length are mismatched. If so + * adjust both in the header. + */ + pbclen = le16_to_cpu(hdr->pbc[0]); + if (PBC2LRH(pbclen) != lrhlen) { + pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen); + hdr->pbc[0] = cpu_to_le16(pbclen); + hdr->lrh[2] = cpu_to_be16(lrhlen >> 2); + /* + * Third packet + * This is the first packet in the sequence that has + * a "static" size that can be used for the rest of + * the packets (besides the last one). + */ + if (unlikely(req->seqnum == 2)) { + /* + * From this point on the lengths in both the + * PBC and LRH are the same until the last + * packet. + * Adjust the template so we don't have to update + * every packet + */ + req->hdr.pbc[0] = hdr->pbc[0]; + req->hdr.lrh[2] = hdr->lrh[2]; + } + } + /* + * We only have to modify the header if this is not the + * first packet in the request. Otherwise, we use the + * header given to us. + */ + if (unlikely(!req->seqnum)) { + ret = check_header_template(req, hdr, lrhlen, datalen); + if (ret) + return ret; + goto done; + } + + hdr->bth[2] = cpu_to_be32( + set_pkt_bth_psn(hdr->bth[2], + (req_opcode(req->info.ctrl) == EXPECTED), + req->seqnum)); + + /* Set ACK request on last packet */ + if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK)) + hdr->bth[2] |= cpu_to_be32(1UL << 31); + + /* Set the new offset */ + hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset); + /* Expected packets have to fill in the new TID information */ + if (req_opcode(req->info.ctrl) == EXPECTED) { + tidval = req->tids[req->tididx]; + /* + * If the offset puts us at the end of the current TID, + * advance everything. + */ + if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) * + PAGE_SIZE)) { + req->tidoffset = 0; + /* + * Since we don't copy all the TIDs, all at once, + * we have to check again. + */ + if (++req->tididx > req->n_tids - 1 || + !req->tids[req->tididx]) { + return -EINVAL; + } + tidval = req->tids[req->tididx]; + } + req->omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >= + KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE : KDETH_OM_SMALL; + /* Set KDETH.TIDCtrl based on value for this TID. */ + KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL, + EXP_TID_GET(tidval, CTRL)); + /* Set KDETH.TID based on value for this TID */ + KDETH_SET(hdr->kdeth.ver_tid_offset, TID, + EXP_TID_GET(tidval, IDX)); + /* Clear KDETH.SH when DISABLE_SH flag is set */ + if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) + KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0); + /* + * Set the KDETH.OFFSET and KDETH.OM based on size of + * transfer. + */ + SDMA_DBG(req, "TID offset %ubytes %uunits om%u", + req->tidoffset, req->tidoffset / req->omfactor, + req->omfactor != KDETH_OM_SMALL); + KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET, + req->tidoffset / req->omfactor); + KDETH_SET(hdr->kdeth.ver_tid_offset, OM, + req->omfactor != KDETH_OM_SMALL); + } +done: + // trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt, + // req->info.comp_idx, hdr, tidval); + return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr)); +} + +static int set_txreq_header_ahg(struct user_sdma_request *req, + struct user_sdma_txreq *tx, u32 len) +{ + int diff = 0; + struct hfi1_user_sdma_pkt_q *pq = req->pq; + struct hfi1_pkt_header *hdr = &req->hdr; + u16 pbclen = le16_to_cpu(hdr->pbc[0]); + u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(len)); + + if (PBC2LRH(pbclen) != lrhlen) { + /* PBC.PbcLengthDWs */ + AHG_HEADER_SET(req->ahg, diff, 0, 0, 12, + cpu_to_le16(LRH2PBC(lrhlen))); + /* LRH.PktLen (we need the full 16 bits due to byte swap) */ + AHG_HEADER_SET(req->ahg, diff, 3, 0, 16, + cpu_to_be16(lrhlen >> 2)); + } + + /* + * Do the common updates + */ + /* BTH.PSN and BTH.A */ + val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) & + (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff); + if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK)) + val32 |= 1UL << 31; + AHG_HEADER_SET(req->ahg, diff, 6, 0, 16, cpu_to_be16(val32 >> 16)); + AHG_HEADER_SET(req->ahg, diff, 6, 16, 16, cpu_to_be16(val32 & 0xffff)); + /* KDETH.Offset */ + AHG_HEADER_SET(req->ahg, diff, 15, 0, 16, + cpu_to_le16(req->koffset & 0xffff)); + AHG_HEADER_SET(req->ahg, diff, 15, 16, 16, + cpu_to_le16(req->koffset >> 16)); + if (req_opcode(req->info.ctrl) == EXPECTED) { + __le16 val; + + tidval = req->tids[req->tididx]; + + /* + * If the offset puts us at the end of the current TID, + * advance everything. + */ + if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) * + PAGE_SIZE)) { + req->tidoffset = 0; + /* + * Since we don't copy all the TIDs, all at once, + * we have to check again. + */ + if (++req->tididx > req->n_tids - 1 || + !req->tids[req->tididx]) { + return -EINVAL; + } + tidval = req->tids[req->tididx]; + } + req->omfactor = ((EXP_TID_GET(tidval, LEN) * + PAGE_SIZE) >= + KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE : + KDETH_OM_SMALL; + /* KDETH.OM and KDETH.OFFSET (TID) */ + AHG_HEADER_SET(req->ahg, diff, 7, 0, 16, + ((!!(req->omfactor - KDETH_OM_SMALL)) << 15 | + ((req->tidoffset / req->omfactor) & 0x7fff))); + /* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */ + val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) | + (EXP_TID_GET(tidval, IDX) & 0x3ff)); + + if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) { + val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset, + INTR) << + AHG_KDETH_INTR_SHIFT)); + } else { + val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ? + cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) : + cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset, + INTR) << + AHG_KDETH_INTR_SHIFT)); + } + + AHG_HEADER_SET(req->ahg, diff, 7, 16, 14, val); + } + + // trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt, + // req->info.comp_idx, req->sde->this_idx, + // req->ahg_idx, req->ahg, diff, tidval); + return diff; +} + +/* + * SDMA tx request completion callback. Called when the SDMA progress + * state machine gets notification that the SDMA descriptors for this + * tx request have been processed by the DMA engine. Called in + * interrupt context. + */ +static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status) +{ + struct user_sdma_txreq *tx = + container_of(txreq, struct user_sdma_txreq, txreq); + struct user_sdma_request *req; + struct hfi1_user_sdma_pkt_q *pq; + struct hfi1_user_sdma_comp_q *cq; + u16 idx; + + if (!tx->req) + return; + + req = tx->req; + pq = req->pq; + cq = req->cq; + + if (status != SDMA_TXREQ_S_OK) { + SDMA_DBG(req, "SDMA completion with error %d", + status); + set_bit(SDMA_REQ_HAS_ERROR, &req->flags); + } + + req->seqcomp = tx->seqnum; + //TODO: kmem_cache_free + //kmem_cache_free(pq->txreq_cache, tx); + tx = NULL; + + idx = req->info.comp_idx; + if (req->status == -1 && status == SDMA_TXREQ_S_OK) { + if (req->seqcomp == req->info.npkts - 1) { + req->status = 0; + user_sdma_free_request(req, false); + pq_update(pq); + set_comp_state(pq, cq, idx, COMPLETE, 0); + } + } else { + if (status != SDMA_TXREQ_S_OK) + req->status = status; + if (req->seqcomp == (ACCESS_ONCE(req->seqsubmitted) - 1) && + (test_bit(SDMA_REQ_SEND_DONE, &req->flags) || + test_bit(SDMA_REQ_DONE_ERROR, &req->flags))) { + user_sdma_free_request(req, false); + pq_update(pq); + set_comp_state(pq, cq, idx, ERROR, req->status); + } + } +} + +static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq) +{ + if (atomic_dec_and_test(&pq->n_reqs)) { + xchg(&pq->state, SDMA_PKT_Q_INACTIVE); + //TODO: wake_up + //wake_up(&pq->wait); + } +} + +static void user_sdma_free_request(struct user_sdma_request *req, bool unpin) +{ + hfi1_cdbg(AIOWRITE, "+"); + if (!list_empty(&req->txps)) { + struct sdma_txreq *t, *p; + + list_for_each_entry_safe(t, p, &req->txps, list) { + struct user_sdma_txreq *tx = + container_of(t, struct user_sdma_txreq, txreq); + list_del_init(&t->list); + sdma_txclean(req->pq->dd, t); +#ifdef __HFI1_ORIG__ + kmem_cache_free(req->pq->txreq_cache, tx); +#endif /* __HFI1_ORIG__ */ + } + } + if (req->data_iovs) { + struct sdma_mmu_node *node; + int i; + + for (i = 0; i < req->data_iovs; i++) { + node = req->iovs[i].node; + if (!node) + continue; + +//TODO: +#ifdef __HFI1_ORIG__ + if (unpin) + hfi1_mmu_rb_remove(req->pq->handler, + &node->rb); + else + atomic_dec(&node->refcount); +#endif /* __HFI1_ORIG__ */ + } + } + kfree(req->tids); + clear_bit(req->info.comp_idx, req->pq->req_in_use); + hfi1_cdbg(AIOWRITE, "-"); +} +static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq, + struct hfi1_user_sdma_comp_q *cq, + u16 idx, enum hfi1_sdma_comp_state state, + int ret) +{ + hfi1_cdbg(SDMA, "[%u:%u:%u:%u] Setting completion status %u %d", + pq->dd->unit, pq->ctxt, pq->subctxt, idx, state, ret); + cq->comps[idx].status = state; + if (state == ERROR) + cq->comps[idx].errcode = -ret; + // trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt, + // idx, state, ret); +} +#ifdef __HFI1_ORIG__ + +static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr, + unsigned long len) +{ + return (bool)(node->addr == addr); +} + +static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode) +{ + struct sdma_mmu_node *node = + container_of(mnode, struct sdma_mmu_node, rb); + + atomic_inc(&node->refcount); + return 0; +} + +/* + * Return 1 to remove the node from the rb tree and call the remove op. + * + * Called with the rb tree lock held. + */ +static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode, + void *evict_arg, bool *stop) +{ + struct sdma_mmu_node *node = + container_of(mnode, struct sdma_mmu_node, rb); + struct evict_data *evict_data = evict_arg; + + /* is this node still being used? */ + if (atomic_read(&node->refcount)) + return 0; /* keep this node */ + + /* this node will be evicted, add its pages to our count */ + evict_data->cleared += node->npages; + + /* have enough pages been cleared? */ + if (evict_data->cleared >= evict_data->target) + *stop = true; + + return 1; /* remove this node */ +} + +static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode) +{ + struct sdma_mmu_node *node = + container_of(mnode, struct sdma_mmu_node, rb); + + atomic_sub(node->npages, &node->pq->n_locked); + + unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages); + + kfree(node); +} + +static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode) +{ + struct sdma_mmu_node *node = + container_of(mnode, struct sdma_mmu_node, rb); + + if (!atomic_read(&node->refcount)) + return 1; + return 0; +} + +#endif /* __HFI1_ORIG__ */