add fileoff-type PTE

fileoff-type PTE holds a file offset which the page data should be
loaded from.

refs #21
This commit is contained in:
NAKAMURA Gou
2014-06-23 18:37:19 +09:00
parent 04f1b3f401
commit bb3756dc74
3 changed files with 70 additions and 17 deletions

View File

@@ -78,6 +78,8 @@
#define PFL3_DIRTY ((pte_t)0x40) #define PFL3_DIRTY ((pte_t)0x40)
#define PFL3_SIZE ((pte_t)0x80) /* Used in 1G page */ #define PFL3_SIZE ((pte_t)0x80) /* Used in 1G page */
#define PFL3_GLOBAL ((pte_t)0x100) #define PFL3_GLOBAL ((pte_t)0x100)
#define PFL3_IGNORED_11 ((pte_t)1 << 11)
#define PFL3_FILEOFF PFL3_IGNORED_11
#define PFL2_PRESENT ((pte_t)0x01) #define PFL2_PRESENT ((pte_t)0x01)
#define PFL2_WRITABLE ((pte_t)0x02) #define PFL2_WRITABLE ((pte_t)0x02)
@@ -88,6 +90,8 @@
#define PFL2_DIRTY ((pte_t)0x40) #define PFL2_DIRTY ((pte_t)0x40)
#define PFL2_SIZE ((pte_t)0x80) /* Used in 2M page */ #define PFL2_SIZE ((pte_t)0x80) /* Used in 2M page */
#define PFL2_GLOBAL ((pte_t)0x100) #define PFL2_GLOBAL ((pte_t)0x100)
#define PFL2_IGNORED_11 ((pte_t)1 << 11)
#define PFL2_FILEOFF PFL2_IGNORED_11
#define PFL1_PRESENT ((pte_t)0x01) #define PFL1_PRESENT ((pte_t)0x01)
#define PFL1_WRITABLE ((pte_t)0x02) #define PFL1_WRITABLE ((pte_t)0x02)
@@ -96,6 +100,8 @@
#define PFL1_PCD ((pte_t)0x10) #define PFL1_PCD ((pte_t)0x10)
#define PFL1_ACCESSED ((pte_t)0x20) #define PFL1_ACCESSED ((pte_t)0x20)
#define PFL1_DIRTY ((pte_t)0x40) #define PFL1_DIRTY ((pte_t)0x40)
#define PFL1_IGNORED_11 ((pte_t)1 << 11)
#define PFL1_FILEOFF PFL1_IGNORED_11
/* We allow user programs to access all the memory */ /* We allow user programs to access all the memory */
#define PFL4_KERN_ATTR (PFL4_PRESENT | PFL4_WRITABLE) #define PFL4_KERN_ATTR (PFL4_PRESENT | PFL4_WRITABLE)
@@ -108,6 +114,9 @@
#define PFL3_PDIR_ATTR (PFL3_PRESENT | PFL3_WRITABLE | PFL3_USER) #define PFL3_PDIR_ATTR (PFL3_PRESENT | PFL3_WRITABLE | PFL3_USER)
#define PFL2_PDIR_ATTR (PFL2_PRESENT | PFL2_WRITABLE | PFL2_USER) #define PFL2_PDIR_ATTR (PFL2_PRESENT | PFL2_WRITABLE | PFL2_USER)
#define PTE_NULL ((pte_t)0)
typedef unsigned long pte_t;
/* For easy conversion, it is better to be the same as architecture's ones */ /* For easy conversion, it is better to be the same as architecture's ones */
enum ihk_mc_pt_attribute { enum ihk_mc_pt_attribute {
PTATTR_ACTIVE = 0x01, PTATTR_ACTIVE = 0x01,
@@ -115,14 +124,12 @@ enum ihk_mc_pt_attribute {
PTATTR_USER = 0x04, PTATTR_USER = 0x04,
PTATTR_DIRTY = 0x40, PTATTR_DIRTY = 0x40,
PTATTR_LARGEPAGE = 0x80, PTATTR_LARGEPAGE = 0x80,
PTATTR_FILEOFF = PFL2_FILEOFF,
PTATTR_NO_EXECUTE = 0x8000000000000000, PTATTR_NO_EXECUTE = 0x8000000000000000,
PTATTR_UNCACHABLE = 0x10000, PTATTR_UNCACHABLE = 0x10000,
PTATTR_FOR_USER = 0x20000, PTATTR_FOR_USER = 0x20000,
}; };
#define PTE_NULL ((pte_t)0)
typedef unsigned long pte_t;
static inline int pte_is_null(pte_t *ptep) static inline int pte_is_null(pte_t *ptep)
{ {
return (*ptep == PTE_NULL); return (*ptep == PTE_NULL);
@@ -153,11 +160,31 @@ static inline int pte_is_dirty(pte_t *ptep, size_t pgsize)
} }
} }
static inline int pte_is_fileoff(pte_t *ptep, size_t pgsize)
{
switch (pgsize) {
case PTL1_SIZE: return !!(*ptep & PFL1_FILEOFF);
case PTL2_SIZE: return !!(*ptep & PFL2_FILEOFF);
case PTL3_SIZE: return !!(*ptep & PFL3_FILEOFF);
default:
#if 0 /* XXX: workaround. cannot use panic() here */
panic("pte_is_fileoff");
#else
return !!(*ptep & PTATTR_FILEOFF);
#endif
}
}
static inline uintptr_t pte_get_phys(pte_t *ptep) static inline uintptr_t pte_get_phys(pte_t *ptep)
{ {
return (*ptep & PT_PHYSMASK); return (*ptep & PT_PHYSMASK);
} }
static inline off_t pte_get_off(pte_t *ptep, size_t pgsize)
{
return (off_t)(*ptep & PAGE_MASK);
}
#if 0 /* XXX: workaround. cannot use panic() here */ #if 0 /* XXX: workaround. cannot use panic() here */
static inline void pte_xchg(pte_t *ptep, pte_t *valp) static inline void pte_xchg(pte_t *ptep, pte_t *valp)
{ {

View File

@@ -219,7 +219,13 @@ static struct page_table *__alloc_new_pt(enum ihk_mc_ap_flag ap_flag)
* but L2 and L1 do not! * but L2 and L1 do not!
*/ */
static enum ihk_mc_pt_attribute attr_mask = PTATTR_WRITABLE | PTATTR_USER | PTATTR_ACTIVE; static enum ihk_mc_pt_attribute attr_mask
= 0
| PTATTR_FILEOFF
| PTATTR_WRITABLE
| PTATTR_USER
| PTATTR_ACTIVE
| 0;
#define ATTR_MASK attr_mask #define ATTR_MASK attr_mask
void enable_ptattr_no_execute(void) void enable_ptattr_no_execute(void)
@@ -523,6 +529,7 @@ int ihk_mc_pt_print_pte(struct page_table *pt, void *virt)
if (!(pt->entry[l4idx] & PFL4_PRESENT)) { if (!(pt->entry[l4idx] & PFL4_PRESENT)) {
__kprintf("0x%lX l4idx not present! \n", (unsigned long)virt); __kprintf("0x%lX l4idx not present! \n", (unsigned long)virt);
__kprintf("l4 entry: 0x%lX\n", pt->entry[l4idx]);
return -EFAULT; return -EFAULT;
} }
pt = phys_to_virt(pt->entry[l4idx] & PAGE_MASK); pt = phys_to_virt(pt->entry[l4idx] & PAGE_MASK);
@@ -530,6 +537,7 @@ int ihk_mc_pt_print_pte(struct page_table *pt, void *virt)
__kprintf("l3 table: 0x%lX l3idx: %d \n", virt_to_phys(pt), l3idx); __kprintf("l3 table: 0x%lX l3idx: %d \n", virt_to_phys(pt), l3idx);
if (!(pt->entry[l3idx] & PFL3_PRESENT)) { if (!(pt->entry[l3idx] & PFL3_PRESENT)) {
__kprintf("0x%lX l3idx not present! \n", (unsigned long)virt); __kprintf("0x%lX l3idx not present! \n", (unsigned long)virt);
__kprintf("l3 entry: 0x%lX\n", pt->entry[l3idx]);
return -EFAULT; return -EFAULT;
} }
pt = phys_to_virt(pt->entry[l3idx] & PAGE_MASK); pt = phys_to_virt(pt->entry[l3idx] & PAGE_MASK);
@@ -537,6 +545,7 @@ int ihk_mc_pt_print_pte(struct page_table *pt, void *virt)
__kprintf("l2 table: 0x%lX l2idx: %d \n", virt_to_phys(pt), l2idx); __kprintf("l2 table: 0x%lX l2idx: %d \n", virt_to_phys(pt), l2idx);
if (!(pt->entry[l2idx] & PFL2_PRESENT)) { if (!(pt->entry[l2idx] & PFL2_PRESENT)) {
__kprintf("0x%lX l2idx not present! \n", (unsigned long)virt); __kprintf("0x%lX l2idx not present! \n", (unsigned long)virt);
__kprintf("l2 entry: 0x%lX\n", pt->entry[l2idx]);
return -EFAULT; return -EFAULT;
} }
if ((pt->entry[l2idx] & PFL2_SIZE)) { if ((pt->entry[l2idx] & PFL2_SIZE)) {
@@ -546,11 +555,12 @@ int ihk_mc_pt_print_pte(struct page_table *pt, void *virt)
__kprintf("l1 table: 0x%lX l1idx: %d \n", virt_to_phys(pt), l1idx); __kprintf("l1 table: 0x%lX l1idx: %d \n", virt_to_phys(pt), l1idx);
if (!(pt->entry[l1idx] & PFL1_PRESENT)) { if (!(pt->entry[l1idx] & PFL1_PRESENT)) {
__kprintf("0x%lX PTE (l1) not present! entry: 0x%lX\n", __kprintf("0x%lX l1idx not present! \n", (unsigned long)virt);
(unsigned long)virt, pt->entry[l1idx]); __kprintf("l1 entry: 0x%lX\n", pt->entry[l1idx]);
return -EFAULT; return -EFAULT;
} }
__kprintf("l1 entry: 0x%lX\n", pt->entry[l1idx]);
return 0; return 0;
} }
@@ -822,8 +832,16 @@ static int split_large_page(pte_t *ptep)
return -ENOMEM; return -ENOMEM;
} }
phys = *ptep & PT_PHYSMASK; if (!(*ptep & PFL2_FILEOFF)) {
attr = *ptep & ~PFL2_SIZE; phys = *ptep & PT_PHYSMASK;
attr = *ptep & ~PT_PHYSMASK;
attr &= ~PFL2_SIZE;
}
else {
phys = *ptep & PAGE_MASK; /* file offset */
attr = *ptep & ~PAGE_MASK;
attr &= ~PFL2_SIZE;
}
for (i = 0; i < PT_ENTRIES; ++i) { for (i = 0; i < PT_ENTRIES; ++i) {
pt->entry[i] = (phys + (i * PTL1_SIZE)) | attr; pt->entry[i] = (phys + (i * PTL1_SIZE)) | attr;
@@ -1008,7 +1026,7 @@ static int clear_range_l1(void *args0, pte_t *ptep, uint64_t base,
memobj_flush_page(args->memobj, phys, PTL1_SIZE); memobj_flush_page(args->memobj, phys, PTL1_SIZE);
} }
if (args->free_physical) { if (!(old & PFL1_FILEOFF) && args->free_physical) {
page = phys_to_page(phys); page = phys_to_page(phys);
if (page && page_unmap(page)) { if (page && page_unmap(page)) {
ihk_mc_free_pages(phys_to_virt(phys), 1); ihk_mc_free_pages(phys_to_virt(phys), 1);
@@ -1054,7 +1072,7 @@ static int clear_range_l2(void *args0, pte_t *ptep, uint64_t base,
memobj_flush_page(args->memobj, phys, PTL2_SIZE); memobj_flush_page(args->memobj, phys, PTL2_SIZE);
} }
if (args->free_physical) { if (!(old & PFL2_FILEOFF) && args->free_physical) {
page = phys_to_page(phys); page = phys_to_page(phys);
if (page && page_unmap(page)) { if (page && page_unmap(page)) {
ihk_mc_free_pages(phys_to_virt(phys), PTL2_SIZE/PTL1_SIZE); ihk_mc_free_pages(phys_to_virt(phys), PTL2_SIZE/PTL1_SIZE);
@@ -1148,7 +1166,7 @@ static int change_attr_range_l1(void *arg0, pte_t *ptep, uint64_t base,
{ {
struct change_attr_args *args = arg0; struct change_attr_args *args = arg0;
if (*ptep == PTE_NULL) { if ((*ptep == PTE_NULL) || (*ptep & PFL1_FILEOFF)) {
return -ENOENT; return -ENOENT;
} }
@@ -1163,7 +1181,7 @@ static int change_attr_range_l2(void *arg0, pte_t *ptep, uint64_t base,
int error; int error;
struct page_table *pt; struct page_table *pt;
if (*ptep == PTE_NULL) { if ((*ptep == PTE_NULL) || (*ptep & PFL2_FILEOFF)) {
return -ENOENT; return -ENOENT;
} }
@@ -1182,7 +1200,9 @@ static int change_attr_range_l2(void *arg0, pte_t *ptep, uint64_t base,
} }
if (*ptep & PFL2_SIZE) { if (*ptep & PFL2_SIZE) {
*ptep = (*ptep & ~args->clrpte) | args->setpte; if (!(*ptep & PFL2_FILEOFF)) {
*ptep = (*ptep & ~args->clrpte) | args->setpte;
}
return 0; return 0;
} }
@@ -1195,7 +1215,7 @@ static int change_attr_range_l3(void *arg0, pte_t *ptep, uint64_t base,
{ {
struct page_table *pt; struct page_table *pt;
if (*ptep == PTE_NULL) { if ((*ptep == PTE_NULL) || (*ptep & PFL3_FILEOFF)) {
return -ENOENT; return -ENOENT;
} }

View File

@@ -1012,7 +1012,8 @@ static int page_fault_process_memory_range(struct process_vm *vm, struct vm_rang
ihk_mc_spinlock_lock_noirq(&vm->page_table_lock); ihk_mc_spinlock_lock_noirq(&vm->page_table_lock);
/*****/ /*****/
ptep = ihk_mc_pt_lookup_pte(vm->page_table, (void *)fault_addr, &pgaddr, &pgsize, &p2align); ptep = ihk_mc_pt_lookup_pte(vm->page_table, (void *)fault_addr, &pgaddr, &pgsize, &p2align);
if (!(reason & PF_PROT) && ptep && !pte_is_null(ptep)) { if (!(reason & PF_PROT) && ptep && !pte_is_null(ptep)
&& !pte_is_fileoff(ptep, pgsize)) {
if (!pte_is_present(ptep)) { if (!pte_is_present(ptep)) {
error = -EFAULT; error = -EFAULT;
kprintf("page_fault_process_memory_range(%p,%lx-%lx %lx,%lx,%lx):PROT_NONE. %d\n", vm, range->start, range->end, range->flag, fault_addr, reason, error); kprintf("page_fault_process_memory_range(%p,%lx-%lx %lx,%lx,%lx):PROT_NONE. %d\n", vm, range->start, range->end, range->flag, fault_addr, reason, error);
@@ -1034,11 +1035,16 @@ static int page_fault_process_memory_range(struct process_vm *vm, struct vm_rang
} }
attr = arch_vrflag_to_ptattr(range->flag, reason, ptep); attr = arch_vrflag_to_ptattr(range->flag, reason, ptep);
pgaddr = (void *)(fault_addr & ~(pgsize - 1)); pgaddr = (void *)(fault_addr & ~(pgsize - 1));
if (!ptep || pte_is_null(ptep)) { if (!ptep || pte_is_null(ptep) || pte_is_fileoff(ptep, pgsize)) {
if (range->memobj) { if (range->memobj) {
off_t off; off_t off;
off = range->objoff + ((uintptr_t)pgaddr - range->start); if (!ptep || !pte_is_fileoff(ptep, pgsize)) {
off = range->objoff + ((uintptr_t)pgaddr - range->start);
}
else {
off = pte_get_off(ptep, pgsize);
}
error = memobj_get_page(range->memobj, off, p2align, &phys); error = memobj_get_page(range->memobj, off, p2align, &phys);
if (error) { if (error) {
if (error != -ERESTART) { if (error != -ERESTART) {