| Message ID | c91e04ebb792ef7b72966edea8bd6fa2dfa5bfa7.1651216964.git.baolin.wang@linux.alibaba.com (mailing list archive) |
|---|---|
| State | Not Applicable |
| Headers | show |
| Series | Fix CONT-PTE/PMD size hugetlb issue when unmapping or migrating | expand |
On Fri, 29 Apr 2022 16:14:43 +0800 Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > On some architectures (like ARM64), it can support CONT-PTE/PMD size > hugetlb, which means it can support not only PMD/PUD size hugetlb: > 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page > size specified. > > When unmapping a hugetlb page, we will get the relevant page table > entry by huge_pte_offset() only once to nuke it. This is correct > for PMD or PUD size hugetlb, since they always contain only one > pmd entry or pud entry in the page table. > > However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, > since they can contain several continuous pte or pmd entry with > same page table attributes, so we will nuke only one pte or pmd > entry for this CONT-PTE/PMD size hugetlb page. > > And now we only use try_to_unmap() to unmap a poisoned hugetlb page, > which means now we will unmap only one pte entry for a CONT-PTE or > CONT-PMD size poisoned hugetlb page, and we can still access other > subpages of a CONT-PTE or CONT-PMD size poisoned hugetlb page, > which will cause serious issues possibly. > > So we should change to use huge_ptep_clear_flush() to nuke the > hugetlb page table to fix this issue, which already considered > CONT-PTE and CONT-PMD size hugetlb. > > Note we've already used set_huge_swap_pte_at() to set a poisoned > swap entry for a poisoned hugetlb page. > > Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com> > --- > mm/rmap.c | 34 +++++++++++++++++----------------- > 1 file changed, 17 insertions(+), 17 deletions(-) > > diff --git a/mm/rmap.c b/mm/rmap.c > index 7cf2408..1e168d7 100644 > --- a/mm/rmap.c > +++ b/mm/rmap.c > @@ -1564,28 +1564,28 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, > break; > } > } > + pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); Unlike in your patch 2/3, I do not see that this (huge) pteval would later be used again with set_huge_pte_at() instead of set_pte_at(). Not sure if this (huge) pteval could end up at a set_pte_at() later, but if yes, then this would be broken on s390, and you'd need to use set_huge_pte_at() instead of set_pte_at() like in your patch 2/3. Please note that huge_ptep_get functions do not return valid PTEs on s390, and such PTEs must never be set directly with set_pte_at(), but only with set_huge_pte_at(). Background is that, for hugetlb pages, we are of course not really dealing with PTEs at this level, but rather PMDs or PUDs, depending on hugetlb size. On s390, the layout is quite different for PTEs and PMDs / PUDs, and unfortunately the hugetlb code is not properly reflecting this by using PMD or PUD types, like the THP code does. So, as work-around, on s390, the huge_ptep_xxx functions will return only fake PTEs, which must be converted again to a proper PMD or PUD, before writing them to the page table, which is what happens in set_huge_pte_at(), but not in set_pte_at().
On 4/30/2022 4:02 AM, Gerald Schaefer wrote: > On Fri, 29 Apr 2022 16:14:43 +0800 > Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > >> On some architectures (like ARM64), it can support CONT-PTE/PMD size >> hugetlb, which means it can support not only PMD/PUD size hugetlb: >> 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page >> size specified. >> >> When unmapping a hugetlb page, we will get the relevant page table >> entry by huge_pte_offset() only once to nuke it. This is correct >> for PMD or PUD size hugetlb, since they always contain only one >> pmd entry or pud entry in the page table. >> >> However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, >> since they can contain several continuous pte or pmd entry with >> same page table attributes, so we will nuke only one pte or pmd >> entry for this CONT-PTE/PMD size hugetlb page. >> >> And now we only use try_to_unmap() to unmap a poisoned hugetlb page, >> which means now we will unmap only one pte entry for a CONT-PTE or >> CONT-PMD size poisoned hugetlb page, and we can still access other >> subpages of a CONT-PTE or CONT-PMD size poisoned hugetlb page, >> which will cause serious issues possibly. >> >> So we should change to use huge_ptep_clear_flush() to nuke the >> hugetlb page table to fix this issue, which already considered >> CONT-PTE and CONT-PMD size hugetlb. >> >> Note we've already used set_huge_swap_pte_at() to set a poisoned >> swap entry for a poisoned hugetlb page. >> >> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com> >> --- >> mm/rmap.c | 34 +++++++++++++++++----------------- >> 1 file changed, 17 insertions(+), 17 deletions(-) >> >> diff --git a/mm/rmap.c b/mm/rmap.c >> index 7cf2408..1e168d7 100644 >> --- a/mm/rmap.c >> +++ b/mm/rmap.c >> @@ -1564,28 +1564,28 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, >> break; >> } >> } >> + pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); > > Unlike in your patch 2/3, I do not see that this (huge) pteval would later > be used again with set_huge_pte_at() instead of set_pte_at(). Not sure if > this (huge) pteval could end up at a set_pte_at() later, but if yes, then > this would be broken on s390, and you'd need to use set_huge_pte_at() > instead of set_pte_at() like in your patch 2/3. IIUC, As I said in the commit message, we will only unmap a poisoned hugetlb page by try_to_unmap(), and the poisoned hugetlb page will be remapped with a poisoned entry by set_huge_swap_pte_at() in try_to_unmap_one(). So I think no need change to use set_huge_pte_at() instead of set_pte_at() for other cases, since the hugetlb page will not hit other cases. if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) { pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); if (folio_test_hugetlb(folio)) { hugetlb_count_sub(folio_nr_pages(folio), mm); set_huge_swap_pte_at(mm, address, pvmw.pte, pteval, vma_mmu_pagesize(vma)); } else { dec_mm_counter(mm, mm_counter(&folio->page)); set_pte_at(mm, address, pvmw.pte, pteval); } } > > Please note that huge_ptep_get functions do not return valid PTEs on s390, > and such PTEs must never be set directly with set_pte_at(), but only with > set_huge_pte_at(). > > Background is that, for hugetlb pages, we are of course not really dealing > with PTEs at this level, but rather PMDs or PUDs, depending on hugetlb size. > On s390, the layout is quite different for PTEs and PMDs / PUDs, and > unfortunately the hugetlb code is not properly reflecting this by using > PMD or PUD types, like the THP code does. > > So, as work-around, on s390, the huge_ptep_xxx functions will return > only fake PTEs, which must be converted again to a proper PMD or PUD, > before writing them to the page table, which is what happens in > set_huge_pte_at(), but not in set_pte_at(). Thanks for your explanation. As I said as above, I think we've already handled the hugetlb with set_huge_swap_pte_at() in try_to_unmap_one().
On Sat, 30 Apr 2022 11:22:33 +0800 Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > > > On 4/30/2022 4:02 AM, Gerald Schaefer wrote: > > On Fri, 29 Apr 2022 16:14:43 +0800 > > Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > > > >> On some architectures (like ARM64), it can support CONT-PTE/PMD size > >> hugetlb, which means it can support not only PMD/PUD size hugetlb: > >> 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page > >> size specified. > >> > >> When unmapping a hugetlb page, we will get the relevant page table > >> entry by huge_pte_offset() only once to nuke it. This is correct > >> for PMD or PUD size hugetlb, since they always contain only one > >> pmd entry or pud entry in the page table. > >> > >> However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, > >> since they can contain several continuous pte or pmd entry with > >> same page table attributes, so we will nuke only one pte or pmd > >> entry for this CONT-PTE/PMD size hugetlb page. > >> > >> And now we only use try_to_unmap() to unmap a poisoned hugetlb page, > >> which means now we will unmap only one pte entry for a CONT-PTE or > >> CONT-PMD size poisoned hugetlb page, and we can still access other > >> subpages of a CONT-PTE or CONT-PMD size poisoned hugetlb page, > >> which will cause serious issues possibly. > >> > >> So we should change to use huge_ptep_clear_flush() to nuke the > >> hugetlb page table to fix this issue, which already considered > >> CONT-PTE and CONT-PMD size hugetlb. > >> > >> Note we've already used set_huge_swap_pte_at() to set a poisoned > >> swap entry for a poisoned hugetlb page. > >> > >> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com> > >> --- > >> mm/rmap.c | 34 +++++++++++++++++----------------- > >> 1 file changed, 17 insertions(+), 17 deletions(-) > >> > >> diff --git a/mm/rmap.c b/mm/rmap.c > >> index 7cf2408..1e168d7 100644 > >> --- a/mm/rmap.c > >> +++ b/mm/rmap.c > >> @@ -1564,28 +1564,28 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, > >> break; > >> } > >> } > >> + pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); > > > > Unlike in your patch 2/3, I do not see that this (huge) pteval would later > > be used again with set_huge_pte_at() instead of set_pte_at(). Not sure if > > this (huge) pteval could end up at a set_pte_at() later, but if yes, then > > this would be broken on s390, and you'd need to use set_huge_pte_at() > > instead of set_pte_at() like in your patch 2/3. > > IIUC, As I said in the commit message, we will only unmap a poisoned > hugetlb page by try_to_unmap(), and the poisoned hugetlb page will be > remapped with a poisoned entry by set_huge_swap_pte_at() in > try_to_unmap_one(). So I think no need change to use set_huge_pte_at() > instead of set_pte_at() for other cases, since the hugetlb page will not > hit other cases. > > if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) { > pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); > if (folio_test_hugetlb(folio)) { > hugetlb_count_sub(folio_nr_pages(folio), mm); > set_huge_swap_pte_at(mm, address, pvmw.pte, pteval, > vma_mmu_pagesize(vma)); > } else { > dec_mm_counter(mm, mm_counter(&folio->page)); > set_pte_at(mm, address, pvmw.pte, pteval); > } > > } OK, but wouldn't the pteval be overwritten here with pteval = swp_entry_to_pte(make_hwpoison_entry(subpage))? IOW, what sense does it make to save the returned pteval from huge_ptep_clear_flush(), when it is never being used anywhere?
On 5/2/2022 10:02 PM, Gerald Schaefer wrote: > On Sat, 30 Apr 2022 11:22:33 +0800 > Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > >> >> >> On 4/30/2022 4:02 AM, Gerald Schaefer wrote: >>> On Fri, 29 Apr 2022 16:14:43 +0800 >>> Baolin Wang <baolin.wang@linux.alibaba.com> wrote: >>> >>>> On some architectures (like ARM64), it can support CONT-PTE/PMD size >>>> hugetlb, which means it can support not only PMD/PUD size hugetlb: >>>> 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page >>>> size specified. >>>> >>>> When unmapping a hugetlb page, we will get the relevant page table >>>> entry by huge_pte_offset() only once to nuke it. This is correct >>>> for PMD or PUD size hugetlb, since they always contain only one >>>> pmd entry or pud entry in the page table. >>>> >>>> However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, >>>> since they can contain several continuous pte or pmd entry with >>>> same page table attributes, so we will nuke only one pte or pmd >>>> entry for this CONT-PTE/PMD size hugetlb page. >>>> >>>> And now we only use try_to_unmap() to unmap a poisoned hugetlb page, >>>> which means now we will unmap only one pte entry for a CONT-PTE or >>>> CONT-PMD size poisoned hugetlb page, and we can still access other >>>> subpages of a CONT-PTE or CONT-PMD size poisoned hugetlb page, >>>> which will cause serious issues possibly. >>>> >>>> So we should change to use huge_ptep_clear_flush() to nuke the >>>> hugetlb page table to fix this issue, which already considered >>>> CONT-PTE and CONT-PMD size hugetlb. >>>> >>>> Note we've already used set_huge_swap_pte_at() to set a poisoned >>>> swap entry for a poisoned hugetlb page. >>>> >>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com> >>>> --- >>>> mm/rmap.c | 34 +++++++++++++++++----------------- >>>> 1 file changed, 17 insertions(+), 17 deletions(-) >>>> >>>> diff --git a/mm/rmap.c b/mm/rmap.c >>>> index 7cf2408..1e168d7 100644 >>>> --- a/mm/rmap.c >>>> +++ b/mm/rmap.c >>>> @@ -1564,28 +1564,28 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, >>>> break; >>>> } >>>> } >>>> + pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); >>> >>> Unlike in your patch 2/3, I do not see that this (huge) pteval would later >>> be used again with set_huge_pte_at() instead of set_pte_at(). Not sure if >>> this (huge) pteval could end up at a set_pte_at() later, but if yes, then >>> this would be broken on s390, and you'd need to use set_huge_pte_at() >>> instead of set_pte_at() like in your patch 2/3. >> >> IIUC, As I said in the commit message, we will only unmap a poisoned >> hugetlb page by try_to_unmap(), and the poisoned hugetlb page will be >> remapped with a poisoned entry by set_huge_swap_pte_at() in >> try_to_unmap_one(). So I think no need change to use set_huge_pte_at() >> instead of set_pte_at() for other cases, since the hugetlb page will not >> hit other cases. >> >> if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) { >> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); >> if (folio_test_hugetlb(folio)) { >> hugetlb_count_sub(folio_nr_pages(folio), mm); >> set_huge_swap_pte_at(mm, address, pvmw.pte, pteval, >> vma_mmu_pagesize(vma)); >> } else { >> dec_mm_counter(mm, mm_counter(&folio->page)); >> set_pte_at(mm, address, pvmw.pte, pteval); >> } >> >> } > > OK, but wouldn't the pteval be overwritten here with > pteval = swp_entry_to_pte(make_hwpoison_entry(subpage))? > IOW, what sense does it make to save the returned pteval from > huge_ptep_clear_flush(), when it is never being used anywhere? Please see previous code, we'll use the original pte value to check if it is uffd-wp armed, and if need to mark it dirty though the hugetlbfs is set noop_dirty_folio(). pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); /* Set the dirty flag on the folio now the pte is gone. */ if (pte_dirty(pteval)) folio_mark_dirty(folio);
On Tue, 3 May 2022 10:19:46 +0800 Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > > > On 5/2/2022 10:02 PM, Gerald Schaefer wrote: > > On Sat, 30 Apr 2022 11:22:33 +0800 > > Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > > > >> > >> > >> On 4/30/2022 4:02 AM, Gerald Schaefer wrote: > >>> On Fri, 29 Apr 2022 16:14:43 +0800 > >>> Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > >>> > >>>> On some architectures (like ARM64), it can support CONT-PTE/PMD size > >>>> hugetlb, which means it can support not only PMD/PUD size hugetlb: > >>>> 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page > >>>> size specified. > >>>> > >>>> When unmapping a hugetlb page, we will get the relevant page table > >>>> entry by huge_pte_offset() only once to nuke it. This is correct > >>>> for PMD or PUD size hugetlb, since they always contain only one > >>>> pmd entry or pud entry in the page table. > >>>> > >>>> However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, > >>>> since they can contain several continuous pte or pmd entry with > >>>> same page table attributes, so we will nuke only one pte or pmd > >>>> entry for this CONT-PTE/PMD size hugetlb page. > >>>> > >>>> And now we only use try_to_unmap() to unmap a poisoned hugetlb page, > >>>> which means now we will unmap only one pte entry for a CONT-PTE or > >>>> CONT-PMD size poisoned hugetlb page, and we can still access other > >>>> subpages of a CONT-PTE or CONT-PMD size poisoned hugetlb page, > >>>> which will cause serious issues possibly. > >>>> > >>>> So we should change to use huge_ptep_clear_flush() to nuke the > >>>> hugetlb page table to fix this issue, which already considered > >>>> CONT-PTE and CONT-PMD size hugetlb. > >>>> > >>>> Note we've already used set_huge_swap_pte_at() to set a poisoned > >>>> swap entry for a poisoned hugetlb page. > >>>> > >>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com> > >>>> --- > >>>> mm/rmap.c | 34 +++++++++++++++++----------------- > >>>> 1 file changed, 17 insertions(+), 17 deletions(-) > >>>> > >>>> diff --git a/mm/rmap.c b/mm/rmap.c > >>>> index 7cf2408..1e168d7 100644 > >>>> --- a/mm/rmap.c > >>>> +++ b/mm/rmap.c > >>>> @@ -1564,28 +1564,28 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, > >>>> break; > >>>> } > >>>> } > >>>> + pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); > >>> > >>> Unlike in your patch 2/3, I do not see that this (huge) pteval would later > >>> be used again with set_huge_pte_at() instead of set_pte_at(). Not sure if > >>> this (huge) pteval could end up at a set_pte_at() later, but if yes, then > >>> this would be broken on s390, and you'd need to use set_huge_pte_at() > >>> instead of set_pte_at() like in your patch 2/3. > >> > >> IIUC, As I said in the commit message, we will only unmap a poisoned > >> hugetlb page by try_to_unmap(), and the poisoned hugetlb page will be > >> remapped with a poisoned entry by set_huge_swap_pte_at() in > >> try_to_unmap_one(). So I think no need change to use set_huge_pte_at() > >> instead of set_pte_at() for other cases, since the hugetlb page will not > >> hit other cases. > >> > >> if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) { > >> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); > >> if (folio_test_hugetlb(folio)) { > >> hugetlb_count_sub(folio_nr_pages(folio), mm); > >> set_huge_swap_pte_at(mm, address, pvmw.pte, pteval, > >> vma_mmu_pagesize(vma)); > >> } else { > >> dec_mm_counter(mm, mm_counter(&folio->page)); > >> set_pte_at(mm, address, pvmw.pte, pteval); > >> } > >> > >> } > > > > OK, but wouldn't the pteval be overwritten here with > > pteval = swp_entry_to_pte(make_hwpoison_entry(subpage))? > > IOW, what sense does it make to save the returned pteval from > > huge_ptep_clear_flush(), when it is never being used anywhere? > > Please see previous code, we'll use the original pte value to check if > it is uffd-wp armed, and if need to mark it dirty though the hugetlbfs > is set noop_dirty_folio(). > > pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); Uh, ok, that wouldn't work on s390, but we also don't have CONFIG_PTE_MARKER_UFFD_WP / HAVE_ARCH_USERFAULTFD_WP set, so I guess we will be fine (for now). Still, I find it a bit unsettling that pte_install_uffd_wp_if_needed() would work on a potential hugetlb *pte, directly de-referencing it instead of using huge_ptep_get(). The !pte_none(*pte) check at the beginning would be broken in the hugetlb case for s390 (not sure about other archs, but I think s390 might be the only exception strictly requiring huge_ptep_get() for de-referencing hugetlb *pte pointers). > > /* Set the dirty flag on the folio now the pte is gone. */ > if (pte_dirty(pteval)) > folio_mark_dirty(folio); Ok, that should work fine, huge_ptep_clear_flush() will return a pteval properly de-referenced and converted with huge_ptep_get(), and that would contain the hugetlb pmd/pud dirty information.
On 5/3/2022 6:03 PM, Gerald Schaefer wrote: > On Tue, 3 May 2022 10:19:46 +0800 > Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > >> >> >> On 5/2/2022 10:02 PM, Gerald Schaefer wrote: >>> On Sat, 30 Apr 2022 11:22:33 +0800 >>> Baolin Wang <baolin.wang@linux.alibaba.com> wrote: >>> >>>> >>>> >>>> On 4/30/2022 4:02 AM, Gerald Schaefer wrote: >>>>> On Fri, 29 Apr 2022 16:14:43 +0800 >>>>> Baolin Wang <baolin.wang@linux.alibaba.com> wrote: >>>>> >>>>>> On some architectures (like ARM64), it can support CONT-PTE/PMD size >>>>>> hugetlb, which means it can support not only PMD/PUD size hugetlb: >>>>>> 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page >>>>>> size specified. >>>>>> >>>>>> When unmapping a hugetlb page, we will get the relevant page table >>>>>> entry by huge_pte_offset() only once to nuke it. This is correct >>>>>> for PMD or PUD size hugetlb, since they always contain only one >>>>>> pmd entry or pud entry in the page table. >>>>>> >>>>>> However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, >>>>>> since they can contain several continuous pte or pmd entry with >>>>>> same page table attributes, so we will nuke only one pte or pmd >>>>>> entry for this CONT-PTE/PMD size hugetlb page. >>>>>> >>>>>> And now we only use try_to_unmap() to unmap a poisoned hugetlb page, >>>>>> which means now we will unmap only one pte entry for a CONT-PTE or >>>>>> CONT-PMD size poisoned hugetlb page, and we can still access other >>>>>> subpages of a CONT-PTE or CONT-PMD size poisoned hugetlb page, >>>>>> which will cause serious issues possibly. >>>>>> >>>>>> So we should change to use huge_ptep_clear_flush() to nuke the >>>>>> hugetlb page table to fix this issue, which already considered >>>>>> CONT-PTE and CONT-PMD size hugetlb. >>>>>> >>>>>> Note we've already used set_huge_swap_pte_at() to set a poisoned >>>>>> swap entry for a poisoned hugetlb page. >>>>>> >>>>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com> >>>>>> --- >>>>>> mm/rmap.c | 34 +++++++++++++++++----------------- >>>>>> 1 file changed, 17 insertions(+), 17 deletions(-) >>>>>> >>>>>> diff --git a/mm/rmap.c b/mm/rmap.c >>>>>> index 7cf2408..1e168d7 100644 >>>>>> --- a/mm/rmap.c >>>>>> +++ b/mm/rmap.c >>>>>> @@ -1564,28 +1564,28 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, >>>>>> break; >>>>>> } >>>>>> } >>>>>> + pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); >>>>> >>>>> Unlike in your patch 2/3, I do not see that this (huge) pteval would later >>>>> be used again with set_huge_pte_at() instead of set_pte_at(). Not sure if >>>>> this (huge) pteval could end up at a set_pte_at() later, but if yes, then >>>>> this would be broken on s390, and you'd need to use set_huge_pte_at() >>>>> instead of set_pte_at() like in your patch 2/3. >>>> >>>> IIUC, As I said in the commit message, we will only unmap a poisoned >>>> hugetlb page by try_to_unmap(), and the poisoned hugetlb page will be >>>> remapped with a poisoned entry by set_huge_swap_pte_at() in >>>> try_to_unmap_one(). So I think no need change to use set_huge_pte_at() >>>> instead of set_pte_at() for other cases, since the hugetlb page will not >>>> hit other cases. >>>> >>>> if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) { >>>> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); >>>> if (folio_test_hugetlb(folio)) { >>>> hugetlb_count_sub(folio_nr_pages(folio), mm); >>>> set_huge_swap_pte_at(mm, address, pvmw.pte, pteval, >>>> vma_mmu_pagesize(vma)); >>>> } else { >>>> dec_mm_counter(mm, mm_counter(&folio->page)); >>>> set_pte_at(mm, address, pvmw.pte, pteval); >>>> } >>>> >>>> } >>> >>> OK, but wouldn't the pteval be overwritten here with >>> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage))? >>> IOW, what sense does it make to save the returned pteval from >>> huge_ptep_clear_flush(), when it is never being used anywhere? >> >> Please see previous code, we'll use the original pte value to check if >> it is uffd-wp armed, and if need to mark it dirty though the hugetlbfs >> is set noop_dirty_folio(). >> >> pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); > > Uh, ok, that wouldn't work on s390, but we also don't have > CONFIG_PTE_MARKER_UFFD_WP / HAVE_ARCH_USERFAULTFD_WP set, so > I guess we will be fine (for now). OK. > > Still, I find it a bit unsettling that pte_install_uffd_wp_if_needed() > would work on a potential hugetlb *pte, directly de-referencing it > instead of using huge_ptep_get(). > > The !pte_none(*pte) check at the beginning would be broken in the > hugetlb case for s390 (not sure about other archs, but I think s390 > might be the only exception strictly requiring huge_ptep_get() > for de-referencing hugetlb *pte pointers). Right, I think so too. I'll look at the uffd code in detail, seems need another patch to fix the hugetlb for uffd. Thanks for your comments.
On 4/29/22 01:14, Baolin Wang wrote: > On some architectures (like ARM64), it can support CONT-PTE/PMD size > hugetlb, which means it can support not only PMD/PUD size hugetlb: > 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page > size specified. > > When unmapping a hugetlb page, we will get the relevant page table > entry by huge_pte_offset() only once to nuke it. This is correct > for PMD or PUD size hugetlb, since they always contain only one > pmd entry or pud entry in the page table. > > However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, > since they can contain several continuous pte or pmd entry with > same page table attributes, so we will nuke only one pte or pmd > entry for this CONT-PTE/PMD size hugetlb page. > > And now we only use try_to_unmap() to unmap a poisoned hugetlb page, Since try_to_unmap can be called for non-hugetlb pages, perhaps the following is more accurate? try_to_unmap is only passed a hugetlb page in the case where the hugetlb page is poisoned. It does concern me that this assumption is built into the code as pointed out in your discussion with Gerald. Should we perhaps add a VM_BUG_ON() to make sure the passed huge page is poisoned? This would be in the same 'if block' where we call adjust_range_if_pmd_sharing_possible.
On 5/3/22 03:03, Gerald Schaefer wrote: > On Tue, 3 May 2022 10:19:46 +0800 > Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > >> >> >> On 5/2/2022 10:02 PM, Gerald Schaefer wrote: >>> On Sat, 30 Apr 2022 11:22:33 +0800 >>> Baolin Wang <baolin.wang@linux.alibaba.com> wrote: >>> >>>> >>>> >>>> On 4/30/2022 4:02 AM, Gerald Schaefer wrote: >>>>> On Fri, 29 Apr 2022 16:14:43 +0800 >>>>> Baolin Wang <baolin.wang@linux.alibaba.com> wrote: >>>>> >>>>>> On some architectures (like ARM64), it can support CONT-PTE/PMD size >>>>>> hugetlb, which means it can support not only PMD/PUD size hugetlb: >>>>>> 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page >>>>>> size specified. >>>>>> >>>>>> When unmapping a hugetlb page, we will get the relevant page table >>>>>> entry by huge_pte_offset() only once to nuke it. This is correct >>>>>> for PMD or PUD size hugetlb, since they always contain only one >>>>>> pmd entry or pud entry in the page table. >>>>>> >>>>>> However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, >>>>>> since they can contain several continuous pte or pmd entry with >>>>>> same page table attributes, so we will nuke only one pte or pmd >>>>>> entry for this CONT-PTE/PMD size hugetlb page. >>>>>> >>>>>> And now we only use try_to_unmap() to unmap a poisoned hugetlb page, >>>>>> which means now we will unmap only one pte entry for a CONT-PTE or >>>>>> CONT-PMD size poisoned hugetlb page, and we can still access other >>>>>> subpages of a CONT-PTE or CONT-PMD size poisoned hugetlb page, >>>>>> which will cause serious issues possibly. >>>>>> >>>>>> So we should change to use huge_ptep_clear_flush() to nuke the >>>>>> hugetlb page table to fix this issue, which already considered >>>>>> CONT-PTE and CONT-PMD size hugetlb. >>>>>> >>>>>> Note we've already used set_huge_swap_pte_at() to set a poisoned >>>>>> swap entry for a poisoned hugetlb page. >>>>>> >>>>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com> >>>>>> --- >>>>>> mm/rmap.c | 34 +++++++++++++++++----------------- >>>>>> 1 file changed, 17 insertions(+), 17 deletions(-) >>>>>> >>>>>> diff --git a/mm/rmap.c b/mm/rmap.c >>>>>> index 7cf2408..1e168d7 100644 >>>>>> --- a/mm/rmap.c >>>>>> +++ b/mm/rmap.c >>>>>> @@ -1564,28 +1564,28 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, >>>>>> break; >>>>>> } >>>>>> } >>>>>> + pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); >>>>> >>>>> Unlike in your patch 2/3, I do not see that this (huge) pteval would later >>>>> be used again with set_huge_pte_at() instead of set_pte_at(). Not sure if >>>>> this (huge) pteval could end up at a set_pte_at() later, but if yes, then >>>>> this would be broken on s390, and you'd need to use set_huge_pte_at() >>>>> instead of set_pte_at() like in your patch 2/3. >>>> >>>> IIUC, As I said in the commit message, we will only unmap a poisoned >>>> hugetlb page by try_to_unmap(), and the poisoned hugetlb page will be >>>> remapped with a poisoned entry by set_huge_swap_pte_at() in >>>> try_to_unmap_one(). So I think no need change to use set_huge_pte_at() >>>> instead of set_pte_at() for other cases, since the hugetlb page will not >>>> hit other cases. >>>> >>>> if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) { >>>> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); >>>> if (folio_test_hugetlb(folio)) { >>>> hugetlb_count_sub(folio_nr_pages(folio), mm); >>>> set_huge_swap_pte_at(mm, address, pvmw.pte, pteval, >>>> vma_mmu_pagesize(vma)); >>>> } else { >>>> dec_mm_counter(mm, mm_counter(&folio->page)); >>>> set_pte_at(mm, address, pvmw.pte, pteval); >>>> } >>>> >>>> } >>> >>> OK, but wouldn't the pteval be overwritten here with >>> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage))? >>> IOW, what sense does it make to save the returned pteval from >>> huge_ptep_clear_flush(), when it is never being used anywhere? >> >> Please see previous code, we'll use the original pte value to check if >> it is uffd-wp armed, and if need to mark it dirty though the hugetlbfs >> is set noop_dirty_folio(). >> >> pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); > > Uh, ok, that wouldn't work on s390, but we also don't have > CONFIG_PTE_MARKER_UFFD_WP / HAVE_ARCH_USERFAULTFD_WP set, so > I guess we will be fine (for now). > > Still, I find it a bit unsettling that pte_install_uffd_wp_if_needed() > would work on a potential hugetlb *pte, directly de-referencing it > instead of using huge_ptep_get(). > > The !pte_none(*pte) check at the beginning would be broken in the > hugetlb case for s390 (not sure about other archs, but I think s390 > might be the only exception strictly requiring huge_ptep_get() > for de-referencing hugetlb *pte pointers). > Adding Peter Wu mostly for above as he is working uffd_wp. >> >> /* Set the dirty flag on the folio now the pte is gone. */ >> if (pte_dirty(pteval)) >> folio_mark_dirty(folio); > > Ok, that should work fine, huge_ptep_clear_flush() will return > a pteval properly de-referenced and converted with huge_ptep_get(), > and that would contain the hugetlb pmd/pud dirty information. >
On 5/7/2022 2:55 AM, Mike Kravetz wrote: > On 4/29/22 01:14, Baolin Wang wrote: >> On some architectures (like ARM64), it can support CONT-PTE/PMD size >> hugetlb, which means it can support not only PMD/PUD size hugetlb: >> 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page >> size specified. >> >> When unmapping a hugetlb page, we will get the relevant page table >> entry by huge_pte_offset() only once to nuke it. This is correct >> for PMD or PUD size hugetlb, since they always contain only one >> pmd entry or pud entry in the page table. >> >> However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, >> since they can contain several continuous pte or pmd entry with >> same page table attributes, so we will nuke only one pte or pmd >> entry for this CONT-PTE/PMD size hugetlb page. >> >> And now we only use try_to_unmap() to unmap a poisoned hugetlb page, > > Since try_to_unmap can be called for non-hugetlb pages, perhaps the following > is more accurate? > > try_to_unmap is only passed a hugetlb page in the case where the > hugetlb page is poisoned. Yes, will update in next version. > It does concern me that this assumption is built into the code as > pointed out in your discussion with Gerald. Should we perhaps add > a VM_BUG_ON() to make sure the passed huge page is poisoned? This > would be in the same 'if block' where we call > adjust_range_if_pmd_sharing_possible. Good point. Will do in next version. Thanks.
On Fri, May 06, 2022 at 12:07:13PM -0700, Mike Kravetz wrote: > On 5/3/22 03:03, Gerald Schaefer wrote: > > On Tue, 3 May 2022 10:19:46 +0800 > > Baolin Wang <baolin.wang@linux.alibaba.com> wrote: > >> On 5/2/2022 10:02 PM, Gerald Schaefer wrote: [...] > >> Please see previous code, we'll use the original pte value to check if > >> it is uffd-wp armed, and if need to mark it dirty though the hugetlbfs > >> is set noop_dirty_folio(). > >> > >> pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); > > > > Uh, ok, that wouldn't work on s390, but we also don't have > > CONFIG_PTE_MARKER_UFFD_WP / HAVE_ARCH_USERFAULTFD_WP set, so > > I guess we will be fine (for now). > > > > Still, I find it a bit unsettling that pte_install_uffd_wp_if_needed() > > would work on a potential hugetlb *pte, directly de-referencing it > > instead of using huge_ptep_get(). > > > > The !pte_none(*pte) check at the beginning would be broken in the > > hugetlb case for s390 (not sure about other archs, but I think s390 > > might be the only exception strictly requiring huge_ptep_get() > > for de-referencing hugetlb *pte pointers). We could have used is_vm_hugetlb_page(vma) within the helper so as to properly use either generic pte or hugetlb version of pte fetching. We may want to conditionally do set_[huge_]pte_at() too at the end. I could prepare a patch for that even if it's not really anything urgently needed. I assume that won't need to block this patchset since we need the pteval for pte_dirty() check anyway and uffd-wp definitely needs it too. Thanks,
On 5/10/2022 12:41 AM, Peter Xu wrote: > On Fri, May 06, 2022 at 12:07:13PM -0700, Mike Kravetz wrote: >> On 5/3/22 03:03, Gerald Schaefer wrote: >>> On Tue, 3 May 2022 10:19:46 +0800 >>> Baolin Wang <baolin.wang@linux.alibaba.com> wrote: >>>> On 5/2/2022 10:02 PM, Gerald Schaefer wrote: > > [...] > >>>> Please see previous code, we'll use the original pte value to check if >>>> it is uffd-wp armed, and if need to mark it dirty though the hugetlbfs >>>> is set noop_dirty_folio(). >>>> >>>> pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); >>> >>> Uh, ok, that wouldn't work on s390, but we also don't have >>> CONFIG_PTE_MARKER_UFFD_WP / HAVE_ARCH_USERFAULTFD_WP set, so >>> I guess we will be fine (for now). >>> >>> Still, I find it a bit unsettling that pte_install_uffd_wp_if_needed() >>> would work on a potential hugetlb *pte, directly de-referencing it >>> instead of using huge_ptep_get(). >>> >>> The !pte_none(*pte) check at the beginning would be broken in the >>> hugetlb case for s390 (not sure about other archs, but I think s390 >>> might be the only exception strictly requiring huge_ptep_get() >>> for de-referencing hugetlb *pte pointers). > > We could have used is_vm_hugetlb_page(vma) within the helper so as to > properly use either generic pte or hugetlb version of pte fetching. We may > want to conditionally do set_[huge_]pte_at() too at the end. > > I could prepare a patch for that even if it's not really anything urgently > needed. I assume that won't need to block this patchset since we need the > pteval for pte_dirty() check anyway and uffd-wp definitely needs it too. OK. Thanks Peter.
diff --git a/mm/rmap.c b/mm/rmap.c index 7cf2408..1e168d7 100644 --- a/mm/rmap.c +++ b/mm/rmap.c @@ -1564,28 +1564,28 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, break; } } + pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); } else { flush_cache_page(vma, address, pte_pfn(*pvmw.pte)); - } - - /* - * Nuke the page table entry. When having to clear - * PageAnonExclusive(), we always have to flush. - */ - if (should_defer_flush(mm, flags) && !anon_exclusive) { /* - * We clear the PTE but do not flush so potentially - * a remote CPU could still be writing to the folio. - * If the entry was previously clean then the - * architecture must guarantee that a clear->dirty - * transition on a cached TLB entry is written through - * and traps if the PTE is unmapped. + * Nuke the page table entry. When having to clear + * PageAnonExclusive(), we always have to flush. */ - pteval = ptep_get_and_clear(mm, address, pvmw.pte); + if (should_defer_flush(mm, flags) && !anon_exclusive) { + /* + * We clear the PTE but do not flush so potentially + * a remote CPU could still be writing to the folio. + * If the entry was previously clean then the + * architecture must guarantee that a clear->dirty + * transition on a cached TLB entry is written through + * and traps if the PTE is unmapped. + */ + pteval = ptep_get_and_clear(mm, address, pvmw.pte); - set_tlb_ubc_flush_pending(mm, pte_dirty(pteval)); - } else { - pteval = ptep_clear_flush(vma, address, pvmw.pte); + set_tlb_ubc_flush_pending(mm, pte_dirty(pteval)); + } else { + pteval = ptep_clear_flush(vma, address, pvmw.pte); + } } /*
On some architectures (like ARM64), it can support CONT-PTE/PMD size hugetlb, which means it can support not only PMD/PUD size hugetlb: 2M and 1G, but also CONT-PTE/PMD size: 64K and 32M if a 4K page size specified. When unmapping a hugetlb page, we will get the relevant page table entry by huge_pte_offset() only once to nuke it. This is correct for PMD or PUD size hugetlb, since they always contain only one pmd entry or pud entry in the page table. However this is incorrect for CONT-PTE and CONT-PMD size hugetlb, since they can contain several continuous pte or pmd entry with same page table attributes, so we will nuke only one pte or pmd entry for this CONT-PTE/PMD size hugetlb page. And now we only use try_to_unmap() to unmap a poisoned hugetlb page, which means now we will unmap only one pte entry for a CONT-PTE or CONT-PMD size poisoned hugetlb page, and we can still access other subpages of a CONT-PTE or CONT-PMD size poisoned hugetlb page, which will cause serious issues possibly. So we should change to use huge_ptep_clear_flush() to nuke the hugetlb page table to fix this issue, which already considered CONT-PTE and CONT-PMD size hugetlb. Note we've already used set_huge_swap_pte_at() to set a poisoned swap entry for a poisoned hugetlb page. Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com> --- mm/rmap.c | 34 +++++++++++++++++----------------- 1 file changed, 17 insertions(+), 17 deletions(-)