[PATCH]mm: use correct address for pte_unmap_unlock in zap_pte_range

[PATCH]mm: use correct address for pte_unmap_unlock in zap_pte_range

[Shaohua Li]

Boot an i386 kernel, I got:
[    3.766613] WARNING: at /workshop/kernel/git/my/linux/arch/x86/mm/highmem_32.c:81 __kunmap_atomic+0x6f/0x113()
[    3.766615] Hardware name: Studio XPS 8000
[    3.766617] Modules linked in:
[    3.766619] Pid: 214, comm: blkid Not tainted 3.0.0-rc3+ #529
[    3.766621] Call Trace:
[    3.766625]  [<c1063803>] warn_slowpath_common+0x6a/0x7f
[    3.766628]  [<c104f494>] ? __kunmap_atomic+0x6f/0x113
[    3.766631]  [<c106382c>] warn_slowpath_null+0x14/0x18
[    3.766633]  [<c104f494>] __kunmap_atomic+0x6f/0x113
[    3.766637]  [<c10e6033>] zap_pte_range+0x291/0x2b4
[    3.766641]  [<c105a37d>] ? get_parent_ip+0xb/0x31
[    3.766644]  [<c10e6180>] unmap_page_range+0x12a/0x147
[    3.766647]  [<c10e6243>] unmap_vmas+0xa6/0xe2
[    3.766650]  [<c10e7c26>] exit_mmap+0x78/0xd3
[    3.766654]  [<c106207a>] mmput+0x39/0x99
[    3.766657]  [<c106582f>] exit_mm+0x101/0x109
[    3.766660]  [<c1066d83>] do_exit+0x1e6/0x312
[    3.766664]  [<c10fff66>] ? fput+0x18/0x1a
[    3.766666]  [<c1066f16>] do_group_exit+0x67/0x8a
[    3.766669]  [<c1066f51>] sys_exit_group+0x18/0x1c
[    3.766673]  [<c1620154>] syscall_call+0x7/0xb

the zap_pte_range does:
	pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
	do {
		...
	             if (force_flush)
                     	break;
		...
        } while (pte++, addr += PAGE_SIZE, addr != end);

        pte_unmap_unlock(pte - 1, ptl);

the pte_unmap_unlock(pte - 1) only works if pte++ does once.

this is a regression introduced by d16dfc550f5326a4000f3(mm: mmu_gather
rework)

Signed-off-by: Shaohua Li <shaohua.li@intel.com>

diff --git a/mm/memory.c b/mm/memory.c
index 6953d39..f624945 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1112,11 +1112,11 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,
 	int force_flush = 0;
 	int rss[NR_MM_COUNTERS];
 	spinlock_t *ptl;
-	pte_t *pte;
+	pte_t *pte, *orig_pte;
 
 again:
 	init_rss_vec(rss);
-	pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
+	orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 	arch_enter_lazy_mmu_mode();
 	do {
 		pte_t ptent = *pte;
@@ -1196,7 +1196,7 @@ again:
 
 	add_mm_rss_vec(mm, rss);
 	arch_leave_lazy_mmu_mode();
-	pte_unmap_unlock(pte - 1, ptl);
+	pte_unmap_unlock(orig_pte, ptl);
 
 	/*
 	 * mmu_gather ran out of room to batch pages, we break out of

Re: [PATCH]mm: use correct address for pte_unmap_unlock in zap_pte_range

[Peter Zijlstra]

On Wed, 2011-06-15 at 18:29 +0800, Shaohua Li wrote:
> Boot an i386 kernel, I got:
> [    3.766613] WARNING: at /workshop/kernel/git/my/linux/arch/x86/mm/highmem_32.c:81 __kunmap_atomic+0x6f/0x113()
> [    3.766615] Hardware name: Studio XPS 8000
> [    3.766617] Modules linked in:
> [    3.766619] Pid: 214, comm: blkid Not tainted 3.0.0-rc3+ #529
> [    3.766621] Call Trace:
> [    3.766625]  [<c1063803>] warn_slowpath_common+0x6a/0x7f
> [    3.766628]  [<c104f494>] ? __kunmap_atomic+0x6f/0x113
> [    3.766631]  [<c106382c>] warn_slowpath_null+0x14/0x18
> [    3.766633]  [<c104f494>] __kunmap_atomic+0x6f/0x113
> [    3.766637]  [<c10e6033>] zap_pte_range+0x291/0x2b4
> [    3.766641]  [<c105a37d>] ? get_parent_ip+0xb/0x31
> [    3.766644]  [<c10e6180>] unmap_page_range+0x12a/0x147
> [    3.766647]  [<c10e6243>] unmap_vmas+0xa6/0xe2
> [    3.766650]  [<c10e7c26>] exit_mmap+0x78/0xd3
> [    3.766654]  [<c106207a>] mmput+0x39/0x99
> [    3.766657]  [<c106582f>] exit_mm+0x101/0x109
> [    3.766660]  [<c1066d83>] do_exit+0x1e6/0x312
> [    3.766664]  [<c10fff66>] ? fput+0x18/0x1a
> [    3.766666]  [<c1066f16>] do_group_exit+0x67/0x8a
> [    3.766669]  [<c1066f51>] sys_exit_group+0x18/0x1c
> [    3.766673]  [<c1620154>] syscall_call+0x7/0xb
> 
> the zap_pte_range does:
> 	pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
> 	do {
> 		...
> 	             if (force_flush)
>                      	break;
> 		...
>         } while (pte++, addr += PAGE_SIZE, addr != end);
> 
>         pte_unmap_unlock(pte - 1, ptl);
> 
> the pte_unmap_unlock(pte - 1) only works if pte++ does once.
> 
> this is a regression introduced by d16dfc550f5326a4000f3(mm: mmu_gather
> rework)

Already send by Steven and picked up by Andrew:

lkml.kernel.org/r/1307158665.3667.86.camel@gandalf.stny.rr.com
lkml.kernel.org/r/201106040413.p544D9ox027032@imap1.linux-foundation.org

Apparently it hasn't made its way to Linus though.

Re: [PATCH]mm: use correct address for pte_unmap_unlock in zap_pte_range

[Linus Torvalds]

On Wed, Jun 15, 2011 at 3:43 AM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
>
> Apparently it hasn't made its way to Linus though.

Andrew sent his queue (well, at least part of it) to me today, so now
it's commit 5f1a19070b16 ("mm: fix wrong kunmap_atomic() pointer") in
the current -git tree.

                      Linus