I'm working on something else atm.
[<c01187b3>]__might_sleep+0x43/0x47
[<c013b6d4>]__alloc_pages+0x24/0x20c
[<c0133650>]file_read_actor+0x0/0x1b0
[<c01131ed>]pte_alloc_one+0x41/0x104
[<c012d05d>]pte_alloc_map+0x4d/0x210
[<c013bc73>]get_page_cache_size+0xf/0x18
[<c0135f38>]move_one_page+0xe8/0x328
[<c0136061>]move_one_page+0x211/0x328
[<c0130644>]vm_enough_memory+0x34/0xc0
[<c01361a9>]move_page_tables+0x31/0x7c
[<c0136860>]do_mremap+0x66c/0x7ec
[<c0136a30>]sys_mremap+0x50/0x73
[<c010748f>]syscall_call+0x7/0xb
William Lee Irwin III wrote:
>
> I'm working on something else atm.
>
> [<c01187b3>]__might_sleep+0x43/0x47
> [<c013b6d4>]__alloc_pages+0x24/0x20c
> [<c0133650>]file_read_actor+0x0/0x1b0
> [<c01131ed>]pte_alloc_one+0x41/0x104
> [<c012d05d>]pte_alloc_map+0x4d/0x210
> [<c013bc73>]get_page_cache_size+0xf/0x18
> [<c0135f38>]move_one_page+0xe8/0x328
> [<c0136061>]move_one_page+0x211/0x328
> [<c0130644>]vm_enough_memory+0x34/0xc0
> [<c01361a9>]move_page_tables+0x31/0x7c
> [<c0136860>]do_mremap+0x66c/0x7ec
> [<c0136a30>]sys_mremap+0x50/0x73
> [<c010748f>]syscall_call+0x7/0xb
>
ooh, oww, ouch. Look at move_one_page():
src = get_one_pte_map_nested(mm, old_addr);
if (src) {
dst = alloc_one_pte_map(mm, new_addr);
error = copy_one_pte(mm, src, dst);
get_one_pte_map_nested() does a kmap_atomic(), and then we go and
call alloc_one_pte_map->pte_alloc_map->pte_alloc_one->alloc_pages()
inside that kmap_atomic().
I guess that has been there since day one.
A simple fix would be to drop the atomic kmap of the source pte
and take it again after the alloc_one_pte_map() call.
Can you think of a more efficient way?
On Fri, Sep 27, 2002 at 10:54:54PM -0700, Andrew Morton wrote:
> A simple fix would be to drop the atomic kmap of the source pte
> and take it again after the alloc_one_pte_map() call.
> Can you think of a more efficient way?
Not one that isn't highly invasive, no. This is what I had in mind
for the easy fix.
Cheers,
Bill
William Lee Irwin III wrote:
>
> On Fri, Sep 27, 2002 at 10:54:54PM -0700, Andrew Morton wrote:
> > A simple fix would be to drop the atomic kmap of the source pte
> > and take it again after the alloc_one_pte_map() call.
> > Can you think of a more efficient way?
>
> Not one that isn't highly invasive, no. This is what I had in mind
> for the easy fix.
>
OK. kmap_atomics are pretty darn quick, but it might be better
to take a peek to see if the pgd and pmd are present, and only
drop the kmap if not.
Care to eyeball this? I haven't tested it yet.
mm/mremap.c | 25 +++++++++++++++++++++++++
1 files changed, 25 insertions(+)
--- 2.5.39/mm/mremap.c~move_one_page_fix Fri Sep 27 22:59:04 2002
+++ 2.5.39-akpm/mm/mremap.c Fri Sep 27 23:05:16 2002
@@ -53,6 +53,20 @@ end:
return pte;
}
+static inline int page_table_present(struct mm_struct *mm, unsigned long addr)
+{
+ pgd_t *pgd;
+ pmd_t *pmd;
+
+ pgd = pgd_offset(mm, addr);
+ if (pgd_none(*pgd))
+ return 0;
+ pmd = pmd_offset(pgd, addr);
+ if (pmd == NULL)
+ return 0;
+ return 1;
+}
+
static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
{
pmd_t * pmd;
@@ -98,7 +112,18 @@ static int move_one_page(struct vm_area_
spin_lock(&mm->page_table_lock);
src = get_one_pte_map_nested(mm, old_addr);
if (src) {
+ /*
+ * Look to see whether alloc_one_pte_map needs to perform a
+ * memory allocation. If it does then we need to drop the
+ * atomic kmap
+ */
+ if (!page_table_present(mm, new_addr)) {
+ pte_unmap_nested(src);
+ src = NULL;
+ }
dst = alloc_one_pte_map(mm, new_addr);
+ if (src == NULL)
+ src = get_one_pte_map_nested(mm, old_addr);
error = copy_one_pte(mm, src, dst);
pte_unmap_nested(src);
pte_unmap(dst);
.