| Message ID | 20220120042513.1648831-3-pasha.tatashin@soleen.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | page table check fixes and cleanups | expand |
On Wed, Jan 19, 2022 at 8:25 PM Pasha Tatashin <pasha.tatashin@soleen.com> wrote: > > syzbot detected a case where the page table counters were not properly > updated. > > syzkaller login: ------------[ cut here ]------------ > kernel BUG at mm/page_table_check.c:162! > invalid opcode: 0000 [#1] PREEMPT SMP KASAN > CPU: 0 PID: 3099 Comm: pasha Not tainted 5.16.0+ #48 > Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIO4 > RIP: 0010:__page_table_check_zero+0x159/0x1a0 > Code: 7d 3a b2 ff 45 39 f5 74 2a e8 43 38 b2 ff 4d 85 e4 01 > RSP: 0018:ffff888010667418 EFLAGS: 00010293 > RAX: 0000000000000000 RBX: 0000000000000001 RCX: 0000000000 > RDX: ffff88800cea8680 RSI: ffffffff81becaf9 RDI: 0000000003 > RBP: ffff888010667450 R08: 0000000000000001 R09: 0000000000 > R10: ffffffff81becaab R11: 0000000000000001 R12: ffff888008 > R13: 0000000000000001 R14: 0000000000000200 R15: dffffc0000 > FS: 0000000000000000(0000) GS:ffff888035e00000(0000) knlG0 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 00007ffd875cad00 CR3: 00000000094ce000 CR4: 0000000000 > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000 > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000 > Call Trace: > <TASK> > free_pcp_prepare+0x3be/0xaa0 > free_unref_page+0x1c/0x650 > ? trace_hardirqs_on+0x6a/0x1d0 > free_compound_page+0xec/0x130 > free_transhuge_page+0x1be/0x260 > __put_compound_page+0x90/0xd0 > release_pages+0x54c/0x1060 > ? filemap_remove_folio+0x161/0x210 > ? lock_downgrade+0x720/0x720 > ? __put_page+0x150/0x150 > ? filemap_free_folio+0x164/0x350 > __pagevec_release+0x7c/0x110 > shmem_undo_range+0x85e/0x1250 > ... > > The repro involved having a huge page that is split due to uprobe event > temporarily replacing one of the pages in the huge page. Later the huge > page was combined again, but the counters were off, as the PTE level > was not properly updated. > > Make sure that not only huge page but also small pages are updated when > a new entry is set or cleared. > > Fixes: df4e817b7108 ("mm: page table check") > > Signed-off-by: Pasha Tatashin <pasha.tatashin@soleen.com> > --- > mm/page_table_check.c | 52 +++++++++++++++++++++++++++++++++++++++++++ > 1 file changed, 52 insertions(+) > > diff --git a/mm/page_table_check.c b/mm/page_table_check.c > index 7504e7caa2a1..2341ac382cd5 100644 > --- a/mm/page_table_check.c > +++ b/mm/page_table_check.c > @@ -145,6 +145,40 @@ static void page_table_check_set(struct mm_struct *mm, unsigned long addr, > } > } > > +static void pte_clear_level(struct mm_struct *mm, unsigned long addr, > + pte_t *ptep) > +{ > + unsigned long i; > + > + for (i = 0; i < PTRS_PER_PTE; i++) { > + __page_table_check_pte_clear(mm, addr, *ptep); > + addr += PAGE_SIZE; > + ptep++; > + } > +} > + > +static void pmd_clear_level(struct mm_struct *mm, unsigned long addr, > + pmd_t *pmdp) > +{ > + unsigned long i; > + > + for (i = 0; i < PTRS_PER_PMD; i++) { > + pmd_t old_pmd = *pmdp; > + > + if (pmd_user_accessible_page(old_pmd)) { > + page_table_check_clear(mm, addr, pmd_pfn(old_pmd), > + PMD_PAGE_SIZE >> PAGE_SHIFT); > + } else if (!pmd_bad(old_pmd) && !pmd_leaf(old_pmd)) { > + pte_t *ptep = pte_offset_map(&old_pmd, addr); > + > + pte_clear_level(mm, addr, ptep); > + pte_unmap(ptep); > + } You can call __page_table_check_pmd_clear(mm, addr, old_pmd, addr) instead to share the new code. > > + addr += PMD_PAGE_SIZE; > + pmdp++; > + } > +} > + > /* > * page is on free list, or is being allocated, verify that counters are zeroes > * crash if they are not. > @@ -186,6 +220,11 @@ void __page_table_check_pmd_clear(struct mm_struct *mm, unsigned long addr, > if (pmd_user_accessible_page(pmd)) { > page_table_check_clear(mm, addr, pmd_pfn(pmd), > PMD_PAGE_SIZE >> PAGE_SHIFT); > + } else if (!pmd_bad(pmd) && !pmd_leaf(pmd)) { > + pte_t *ptep = pte_offset_map(&pmd, addr); > + > + pte_clear_level(mm, addr, ptep); > + pte_unmap(ptep); > } > } > EXPORT_SYMBOL(__page_table_check_pmd_clear); > @@ -199,6 +238,10 @@ void __page_table_check_pud_clear(struct mm_struct *mm, unsigned long addr, > if (pud_user_accessible_page(pud)) { > page_table_check_clear(mm, addr, pud_pfn(pud), > PUD_PAGE_SIZE >> PAGE_SHIFT); > + } else if (!pud_bad(pud) && !pud_leaf(pud)) { > + pmd_t *pmdp = pmd_offset(&pud, addr); > + > + pmd_clear_level(mm, addr, pmdp); > } > } > EXPORT_SYMBOL(__page_table_check_pud_clear); > @@ -237,6 +280,11 @@ void __page_table_check_pmd_set(struct mm_struct *mm, unsigned long addr, > if (pmd_user_accessible_page(old_pmd)) { > page_table_check_clear(mm, addr, pmd_pfn(old_pmd), > PMD_PAGE_SIZE >> PAGE_SHIFT); > + } else if (!pmd_bad(old_pmd) && !pmd_leaf(old_pmd)) { > + pte_t *ptep = pte_offset_map(&old_pmd, addr); > + > + pte_clear_level(mm, addr, ptep); > + pte_unmap(ptep); > } > How about replacing the above code with __page_table_check_pmd_clear(mm, addr, old_pmd)? > if (pmd_user_accessible_page(pmd)) { > @@ -259,6 +307,10 @@ void __page_table_check_pud_set(struct mm_struct *mm, unsigned long addr, > if (pud_user_accessible_page(old_pud)) { > page_table_check_clear(mm, addr, pud_pfn(old_pud), > PUD_PAGE_SIZE >> PAGE_SHIFT); > + } else if (!pud_bad(old_pud) && !pud_leaf(old_pud)) { > + pmd_t *pmdp = pmd_offset(&old_pud, addr); > + > + pmd_clear_level(mm, addr, pmdp); > } Replacing with __page_table_check_pud_clear(mm, addr, old_pud)? > > > > if (pud_user_accessible_page(pud)) { > -- > 2.34.1.703.g22d0c6ccf7-goog >
Hi Wei, Thank you for looking at this patch. > > +static void pmd_clear_level(struct mm_struct *mm, unsigned long addr, > > + pmd_t *pmdp) > > +{ > > + unsigned long i; > > + > > + for (i = 0; i < PTRS_PER_PMD; i++) { > > + pmd_t old_pmd = *pmdp; > > + > > + if (pmd_user_accessible_page(old_pmd)) { > > + page_table_check_clear(mm, addr, pmd_pfn(old_pmd), > > + PMD_PAGE_SIZE >> PAGE_SHIFT); > > + } else if (!pmd_bad(old_pmd) && !pmd_leaf(old_pmd)) { > > + pte_t *ptep = pte_offset_map(&old_pmd, addr); > > + > > + pte_clear_level(mm, addr, ptep); > > + pte_unmap(ptep); > > + } > > You can call __page_table_check_pmd_clear(mm, addr, old_pmd, addr) > instead to share the new code. Yeap. > > > > > + addr += PMD_PAGE_SIZE; > > + pmdp++; > > + } > > +} > > + > > /* > > * page is on free list, or is being allocated, verify that counters are zeroes > > * crash if they are not. > > @@ -186,6 +220,11 @@ void __page_table_check_pmd_clear(struct mm_struct *mm, unsigned long addr, > > if (pmd_user_accessible_page(pmd)) { > > page_table_check_clear(mm, addr, pmd_pfn(pmd), > > PMD_PAGE_SIZE >> PAGE_SHIFT); > > + } else if (!pmd_bad(pmd) && !pmd_leaf(pmd)) { > > + pte_t *ptep = pte_offset_map(&pmd, addr); > > + > > + pte_clear_level(mm, addr, ptep); > > + pte_unmap(ptep); > > } > > } > > EXPORT_SYMBOL(__page_table_check_pmd_clear); > > @@ -199,6 +238,10 @@ void __page_table_check_pud_clear(struct mm_struct *mm, unsigned long addr, > > if (pud_user_accessible_page(pud)) { > > page_table_check_clear(mm, addr, pud_pfn(pud), > > PUD_PAGE_SIZE >> PAGE_SHIFT); > > + } else if (!pud_bad(pud) && !pud_leaf(pud)) { > > + pmd_t *pmdp = pmd_offset(&pud, addr); > > + > > + pmd_clear_level(mm, addr, pmdp); > > } > > } > > EXPORT_SYMBOL(__page_table_check_pud_clear); > > @@ -237,6 +280,11 @@ void __page_table_check_pmd_set(struct mm_struct *mm, unsigned long addr, > > if (pmd_user_accessible_page(old_pmd)) { > > page_table_check_clear(mm, addr, pmd_pfn(old_pmd), > > PMD_PAGE_SIZE >> PAGE_SHIFT); > > + } else if (!pmd_bad(old_pmd) && !pmd_leaf(old_pmd)) { > > + pte_t *ptep = pte_offset_map(&old_pmd, addr); > > + > > + pte_clear_level(mm, addr, ptep); > > + pte_unmap(ptep); > > } > > > > How about replacing the above code with > __page_table_check_pmd_clear(mm, addr, old_pmd)? OK > > > if (pmd_user_accessible_page(pmd)) { > > @@ -259,6 +307,10 @@ void __page_table_check_pud_set(struct mm_struct *mm, unsigned long addr, > > if (pud_user_accessible_page(old_pud)) { > > page_table_check_clear(mm, addr, pud_pfn(old_pud), > > PUD_PAGE_SIZE >> PAGE_SHIFT); > > + } else if (!pud_bad(old_pud) && !pud_leaf(old_pud)) { > > + pmd_t *pmdp = pmd_offset(&old_pud, addr); > > + > > + pmd_clear_level(mm, addr, pmdp); > > } > > Replacing with __page_table_check_pud_clear(mm, addr, old_pud)? OK Good suggestions, I will simplify the code with your suggestions. Pasha
diff --git a/mm/page_table_check.c b/mm/page_table_check.c index 7504e7caa2a1..2341ac382cd5 100644 --- a/mm/page_table_check.c +++ b/mm/page_table_check.c @@ -145,6 +145,40 @@ static void page_table_check_set(struct mm_struct *mm, unsigned long addr, } } +static void pte_clear_level(struct mm_struct *mm, unsigned long addr, + pte_t *ptep) +{ + unsigned long i; + + for (i = 0; i < PTRS_PER_PTE; i++) { + __page_table_check_pte_clear(mm, addr, *ptep); + addr += PAGE_SIZE; + ptep++; + } +} + +static void pmd_clear_level(struct mm_struct *mm, unsigned long addr, + pmd_t *pmdp) +{ + unsigned long i; + + for (i = 0; i < PTRS_PER_PMD; i++) { + pmd_t old_pmd = *pmdp; + + if (pmd_user_accessible_page(old_pmd)) { + page_table_check_clear(mm, addr, pmd_pfn(old_pmd), + PMD_PAGE_SIZE >> PAGE_SHIFT); + } else if (!pmd_bad(old_pmd) && !pmd_leaf(old_pmd)) { + pte_t *ptep = pte_offset_map(&old_pmd, addr); + + pte_clear_level(mm, addr, ptep); + pte_unmap(ptep); + } + addr += PMD_PAGE_SIZE; + pmdp++; + } +} + /* * page is on free list, or is being allocated, verify that counters are zeroes * crash if they are not. @@ -186,6 +220,11 @@ void __page_table_check_pmd_clear(struct mm_struct *mm, unsigned long addr, if (pmd_user_accessible_page(pmd)) { page_table_check_clear(mm, addr, pmd_pfn(pmd), PMD_PAGE_SIZE >> PAGE_SHIFT); + } else if (!pmd_bad(pmd) && !pmd_leaf(pmd)) { + pte_t *ptep = pte_offset_map(&pmd, addr); + + pte_clear_level(mm, addr, ptep); + pte_unmap(ptep); } } EXPORT_SYMBOL(__page_table_check_pmd_clear); @@ -199,6 +238,10 @@ void __page_table_check_pud_clear(struct mm_struct *mm, unsigned long addr, if (pud_user_accessible_page(pud)) { page_table_check_clear(mm, addr, pud_pfn(pud), PUD_PAGE_SIZE >> PAGE_SHIFT); + } else if (!pud_bad(pud) && !pud_leaf(pud)) { + pmd_t *pmdp = pmd_offset(&pud, addr); + + pmd_clear_level(mm, addr, pmdp); } } EXPORT_SYMBOL(__page_table_check_pud_clear); @@ -237,6 +280,11 @@ void __page_table_check_pmd_set(struct mm_struct *mm, unsigned long addr, if (pmd_user_accessible_page(old_pmd)) { page_table_check_clear(mm, addr, pmd_pfn(old_pmd), PMD_PAGE_SIZE >> PAGE_SHIFT); + } else if (!pmd_bad(old_pmd) && !pmd_leaf(old_pmd)) { + pte_t *ptep = pte_offset_map(&old_pmd, addr); + + pte_clear_level(mm, addr, ptep); + pte_unmap(ptep); } if (pmd_user_accessible_page(pmd)) { @@ -259,6 +307,10 @@ void __page_table_check_pud_set(struct mm_struct *mm, unsigned long addr, if (pud_user_accessible_page(old_pud)) { page_table_check_clear(mm, addr, pud_pfn(old_pud), PUD_PAGE_SIZE >> PAGE_SHIFT); + } else if (!pud_bad(old_pud) && !pud_leaf(old_pud)) { + pmd_t *pmdp = pmd_offset(&old_pud, addr); + + pmd_clear_level(mm, addr, pmdp); } if (pud_user_accessible_page(pud)) {
syzbot detected a case where the page table counters were not properly updated. syzkaller login: ------------[ cut here ]------------ kernel BUG at mm/page_table_check.c:162! invalid opcode: 0000 [#1] PREEMPT SMP KASAN CPU: 0 PID: 3099 Comm: pasha Not tainted 5.16.0+ #48 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIO4 RIP: 0010:__page_table_check_zero+0x159/0x1a0 Code: 7d 3a b2 ff 45 39 f5 74 2a e8 43 38 b2 ff 4d 85 e4 01 RSP: 0018:ffff888010667418 EFLAGS: 00010293 RAX: 0000000000000000 RBX: 0000000000000001 RCX: 0000000000 RDX: ffff88800cea8680 RSI: ffffffff81becaf9 RDI: 0000000003 RBP: ffff888010667450 R08: 0000000000000001 R09: 0000000000 R10: ffffffff81becaab R11: 0000000000000001 R12: ffff888008 R13: 0000000000000001 R14: 0000000000000200 R15: dffffc0000 FS: 0000000000000000(0000) GS:ffff888035e00000(0000) knlG0 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007ffd875cad00 CR3: 00000000094ce000 CR4: 0000000000 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000 Call Trace: <TASK> free_pcp_prepare+0x3be/0xaa0 free_unref_page+0x1c/0x650 ? trace_hardirqs_on+0x6a/0x1d0 free_compound_page+0xec/0x130 free_transhuge_page+0x1be/0x260 __put_compound_page+0x90/0xd0 release_pages+0x54c/0x1060 ? filemap_remove_folio+0x161/0x210 ? lock_downgrade+0x720/0x720 ? __put_page+0x150/0x150 ? filemap_free_folio+0x164/0x350 __pagevec_release+0x7c/0x110 shmem_undo_range+0x85e/0x1250 ... The repro involved having a huge page that is split due to uprobe event temporarily replacing one of the pages in the huge page. Later the huge page was combined again, but the counters were off, as the PTE level was not properly updated. Make sure that not only huge page but also small pages are updated when a new entry is set or cleared. Fixes: df4e817b7108 ("mm: page table check") Signed-off-by: Pasha Tatashin <pasha.tatashin@soleen.com> --- mm/page_table_check.c | 52 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 52 insertions(+)