gh0s7/mckernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

hfi1/user_exp_rcv/setup: do not skip over pages

Browse Source 
If the vaddr we consider is not at the start of a page, we could skip
over (smaller, not contigous) areas.

For example consider this segment of virtual memory:
[ 2MB | 4k | 4k | ... ]
Starting at 1MB offset, we would get a pgsize of 2MB so would skip
straight over 1MB worth of 4k pages.
This commit is contained in:
Dominique Martinet
2017-09-21 16:50:34 +09:00
committed by Balazs Gerofi
parent 50fafa6d71
commit 04434320fc
1 changed files with 29 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
kernel/user_exp_rcv.c
View File

@@ -83,11 +83,11 @@ int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd, struct hfi1_tid_info *tinf
{ {
int ret = -EFAULT; int ret = -EFAULT;
struct hfi1_ctxtdata *uctxt = fd->uctxt; struct hfi1_ctxtdata *uctxt = fd->uctxt;
uintptr_t vaddr_prev, vaddr, vaddr_end; uintptr_t vaddr_prev, vaddr, vaddr_end, base_vaddr;
u32 *tidlist; u32 *tidlist;
u16 tididx = 0; u16 tididx = 0;
struct process_vm *vm = cpu_local_var(current)->vm; struct process_vm *vm = cpu_local_var(current)->vm;
size_t base_pgsize, len = 0; size_t base_pgsize, pgsize, len = 0;
pte_t *ptep; pte_t *ptep;
u64 phys_start = 0, phys_prev = 0, phys; u64 phys_start = 0, phys_prev = 0, phys;
@@ -111,10 +111,11 @@ int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd, struct hfi1_tid_info *tinf
vaddr_end = tinfo->vaddr + tinfo->length; vaddr_end = tinfo->vaddr + tinfo->length;
dkprintf("setup start: 0x%llx, length: %zu\n", tinfo->vaddr, dkprintf("setup start: 0x%llx, length: %zu\n", tinfo->vaddr,
tinfo->length); tinfo->length);
for (vaddr = tinfo->vaddr; vaddr < vaddr_end; vaddr += base_pgsize) { for (vaddr = tinfo->vaddr; vaddr < vaddr_end; vaddr += pgsize) {
ptep = ihk_mc_pt_lookup_pte(vm->address_space->page_table, ptep = ihk_mc_pt_lookup_pte(vm->address_space->page_table,
(void*)vaddr, 0, 0, &base_pgsize, (void*)vaddr, 0,
0); (void**)&base_vaddr,
&base_pgsize, 0);
if (unlikely(!ptep || !pte_is_present(ptep))) { if (unlikely(!ptep || !pte_is_present(ptep))) {
kprintf("%s: ERRROR: no valid PTE for 0x%lx\n", kprintf("%s: ERRROR: no valid PTE for 0x%lx\n",
__FUNCTION__, vaddr); __FUNCTION__, vaddr);
@@ -148,26 +149,39 @@ int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd, struct hfi1_tid_info *tinf
len = 0; len = 0;
} }
/* Corner case #1: base_pgsize is too big (requested length) */ pgsize = base_pgsize;
if (vaddr + base_pgsize > vaddr_end) { /* Corner case #1: pgsize is too big wrt. requested length */
base_pgsize = vaddr_end - vaddr; if (vaddr + pgsize > vaddr_end) {
pgsize = vaddr_end - vaddr;
} }
/* Corner case #2: base_pgsize is too big (max request size). /* Corner case #2: pgsize is too big wrt. max request size.
* This will result in an extra virt to phys lookup, * This will result in an extra virt to phys lookup,
* but should be rare in practice * but should be rare in practice
*/ */
if (len + base_pgsize > MAX_EXPECTED_BUFFER) { if (len + pgsize > MAX_EXPECTED_BUFFER) {
size_t extra = len + base_pgsize - MAX_EXPECTED_BUFFER; size_t extra = len + pgsize - MAX_EXPECTED_BUFFER;
phys -= extra; phys -= extra;
vaddr -= extra; vaddr -= extra;
base_pgsize -= extra; pgsize -= extra;
}
/* Corner case #3: don't let us go further than start of
* next page.
* If we're already towards the end of a big page, and the
* next ones are smaller, we could fly over some pages without
* noticing. No skipping!
*/
if (vaddr + pgsize > base_vaddr + base_pgsize) {
dkprintf("truncating down %zu to %zu, vaddr 0x%llx, base 0x%llx\n",
pgsize, base_vaddr - vaddr + base_pgsize, vaddr, base_vaddr);
pgsize = base_vaddr - vaddr + base_pgsize;
} }
phys_prev = phys; phys_prev = phys;
vaddr_prev = vaddr; vaddr_prev = vaddr;
len += base_pgsize; len += pgsize;
dkprintf("phys 0x%llx, vaddr 0x%llx, len %zu, base_pgsize %zu\n", dkprintf("phys 0x%llx, vaddr 0x%llx, len %zu, pgsize %zu\n",
phys, vaddr, len, base_pgsize); phys, vaddr, len, pgsize);
} }
/* Register whatever is left */ /* Register whatever is left */