| Message ID | 20220324170600.16491-1-mike.kravetz@oracle.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [RFC] hugetlbfs: zero partial pages during fallocate hole punch | expand |
On 24.03.22 18:06, Mike Kravetz wrote: > hugetlbfs fallocate support was originally added with commit 70c3547e36f5 > ("hugetlbfs: add hugetlbfs_fallocate()"). Initial support only operated > on whole hugetlb pages. This makes sense for populating files as other > interfaces such as mmap and truncate require hugetlb page size alignment. > Only operating on whole hugetlb pages for the hole punch case was a > simplification and there was no compelling use case to zero partial pages. > > In a recent discussion[1] it was assumed that hugetlbfs hole punch would > zero partial hugetlb pages as that is in line with the man page > description saying 'partial filesystem blocks are zeroed'. However, > the hugetlbfs hole punch code actually does this: > > hole_start = round_up(offset, hpage_size); > hole_end = round_down(offset + len, hpage_size); > > Modify code to zero partial hugetlb pages in hole punch range. It is > possible that application code could note a change in behavior. However, > that would imply the code is passing in an unaligned range and expecting > only whole pages be removed. This is unlikely as the fallocate > documentation states the opposite. Yeah, some weird code could behave differently; an app would have to pass in an unaligned range and expect that partially covered hugetlbfs pages remain unmodified. It's hard to think of reasonable apps that do that, but of course, some buggy code might then be *actually* broken. Like some messed-up align-up implementation that accidentally adds +1 too much. > > The current hugetlbfs fallocate hole punch behavior is tested with the > libhugetlbfs test fallocate_align[2]. This test will be updated to > validate partial page zeroing. This is in line with other fallocate() behavior and documented semantics, so I think that's the right thing to do. I think it's worth to give it a try, it's hard to imagine that this actually breaks something. "After a successful call, subsequent reads from this range will return zeros." will work as expected with your change. > > [1] https://lore.kernel.org/linux-mm/20571829-9d3d-0b48-817c-b6b15565f651@redhat.com/ > [2] https://github.com/libhugetlbfs/libhugetlbfs/blob/master/tests/fallocate_align.c > > Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com> > --- > fs/hugetlbfs/inode.c | 67 ++++++++++++++++++++++++++++++--------- > include/asm-generic/tlb.h | 2 ++ > 2 files changed, 54 insertions(+), 15 deletions(-) > > diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c > index a7c6c7498be0..f62ec4f71132 100644 > --- a/fs/hugetlbfs/inode.c > +++ b/fs/hugetlbfs/inode.c > @@ -587,41 +587,78 @@ static void hugetlb_vmtruncate(struct inode *inode, loff_t offset) > remove_inode_hugepages(inode, offset, LLONG_MAX); > } > > +static void hugetlbfs_zero_partial_page(struct hstate *h, > + struct address_space *mapping, > + unsigned long start, > + unsigned long end) > +{ > + struct page *page; > + pgoff_t idx = start >> huge_page_shift(h); I'm a fan of reverse Christmas trees :) > + > + page = find_lock_page(mapping, idx); > + if (!page) > + return; > + > + start = start & ~huge_page_mask(h); > + end = end & ~huge_page_mask(h); > + if (!end) > + end = huge_page_size(h); > + > + zero_user_segment(page, (unsigned int)start, (unsigned int)end); > + > + unlock_page(page); > + put_page(page); > +} > + > static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len) > { > struct hstate *h = hstate_inode(inode); > loff_t hpage_size = huge_page_size(h); > loff_t hole_start, hole_end; > + struct address_space *mapping = inode->i_mapping; > + struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode); ditto. > > /* > - * For hole punch round up the beginning offset of the hole and > - * round down the end. > + * hole_start and hole_end indicate the full pages within the hole. > */ > hole_start = round_up(offset, hpage_size); > hole_end = round_down(offset + len, hpage_size); > > - if (hole_end > hole_start) { > - struct address_space *mapping = inode->i_mapping; > - struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode); > + inode_lock(inode); > + > + /* protected by i_rwsem */ > + if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) { > + inode_unlock(inode); > + return -EPERM; > + } > > - inode_lock(inode); > + i_mmap_lock_write(mapping); > > - /* protected by i_rwsem */ > - if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) { > - inode_unlock(inode); > - return -EPERM; > - } > + /* If range starts before first full page, zero partial page. */ > + if (offset < hole_start) > + hugetlbfs_zero_partial_page(h, mapping, > + offset, min(offset + len, hole_start)); > > - i_mmap_lock_write(mapping); > + /* Unmap users of full pages in the hole. */ > + if (hole_end > hole_start) { > if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)) > hugetlb_vmdelete_list(&mapping->i_mmap, > hole_start >> PAGE_SHIFT, > hole_end >> PAGE_SHIFT); > - i_mmap_unlock_write(mapping); > - remove_inode_hugepages(inode, hole_start, hole_end); > - inode_unlock(inode); > } > > + /* If range extends beyond last full page, zero partial page. */ > + if ((offset + len) > hole_end && (offset + len) > hole_start) > + hugetlbfs_zero_partial_page(h, mapping, > + hole_end, offset + len); > + > + i_mmap_unlock_write(mapping); > + > + /* Remove full pages from the file. */ > + if (hole_end > hole_start) > + remove_inode_hugepages(inode, hole_start, hole_end); > + inode_unlock(inode); > + > return 0; > } > > diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h > index 2c68a545ffa7..4622ee45f739 100644 > --- a/include/asm-generic/tlb.h > +++ b/include/asm-generic/tlb.h > @@ -562,6 +562,7 @@ static inline void tlb_flush_p4d_range(struct mmu_gather *tlb, > __tlb_remove_tlb_entry(tlb, ptep, address); \ > } while (0) > > +#ifndef tlb_remove_huge_tlb_entry > #define tlb_remove_huge_tlb_entry(h, tlb, ptep, address) \ > do { \ > unsigned long _sz = huge_page_size(h); \ > @@ -571,6 +572,7 @@ static inline void tlb_flush_p4d_range(struct mmu_gather *tlb, > tlb_flush_pud_range(tlb, address, _sz); \ > __tlb_remove_tlb_entry(tlb, ptep, address); \ > } while (0) > +#endif Was this change supposed to be included in this patch?
On 3/30/22 00:58, David Hildenbrand wrote: > On 24.03.22 18:06, Mike Kravetz wrote: >> hugetlbfs fallocate support was originally added with commit 70c3547e36f5 >> ("hugetlbfs: add hugetlbfs_fallocate()"). Initial support only operated >> on whole hugetlb pages. This makes sense for populating files as other >> interfaces such as mmap and truncate require hugetlb page size alignment. >> Only operating on whole hugetlb pages for the hole punch case was a >> simplification and there was no compelling use case to zero partial pages. >> >> In a recent discussion[1] it was assumed that hugetlbfs hole punch would >> zero partial hugetlb pages as that is in line with the man page >> description saying 'partial filesystem blocks are zeroed'. However, >> the hugetlbfs hole punch code actually does this: >> >> hole_start = round_up(offset, hpage_size); >> hole_end = round_down(offset + len, hpage_size); >> >> Modify code to zero partial hugetlb pages in hole punch range. It is >> possible that application code could note a change in behavior. However, >> that would imply the code is passing in an unaligned range and expecting >> only whole pages be removed. This is unlikely as the fallocate >> documentation states the opposite. > > Yeah, some weird code could behave differently; an app would have to > pass in an unaligned range and expect that partially covered hugetlbfs > pages remain unmodified. It's hard to think of reasonable apps that do > that, but of course, some buggy code might then be *actually* broken. > Like some messed-up align-up implementation that accidentally adds +1 > too much. > >> >> The current hugetlbfs fallocate hole punch behavior is tested with the >> libhugetlbfs test fallocate_align[2]. This test will be updated to >> validate partial page zeroing. > > This is in line with other fallocate() behavior and documented > semantics, so I think that's the right thing to do. > > I think it's worth to give it a try, it's hard to imagine that this > actually breaks something. > > > "After a successful call, subsequent reads from this range will return > zeros." will work as expected with your change. > >> >> [1] https://lore.kernel.org/linux-mm/20571829-9d3d-0b48-817c-b6b15565f651@redhat.com/ >> [2] https://github.com/libhugetlbfs/libhugetlbfs/blob/master/tests/fallocate_align.c >> >> Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com> >> --- >> fs/hugetlbfs/inode.c | 67 ++++++++++++++++++++++++++++++--------- >> include/asm-generic/tlb.h | 2 ++ >> 2 files changed, 54 insertions(+), 15 deletions(-) >> >> diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c >> index a7c6c7498be0..f62ec4f71132 100644 >> --- a/fs/hugetlbfs/inode.c >> +++ b/fs/hugetlbfs/inode.c >> @@ -587,41 +587,78 @@ static void hugetlb_vmtruncate(struct inode *inode, loff_t offset) >> remove_inode_hugepages(inode, offset, LLONG_MAX); >> } >> >> +static void hugetlbfs_zero_partial_page(struct hstate *h, >> + struct address_space *mapping, >> + unsigned long start, >> + unsigned long end) >> +{ >> + struct page *page; >> + pgoff_t idx = start >> huge_page_shift(h); > > I'm a fan of reverse Christmas trees :) > Ok, no preference by me so I will change. ... >> diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h >> index 2c68a545ffa7..4622ee45f739 100644 >> --- a/include/asm-generic/tlb.h >> +++ b/include/asm-generic/tlb.h >> @@ -562,6 +562,7 @@ static inline void tlb_flush_p4d_range(struct mmu_gather *tlb, >> __tlb_remove_tlb_entry(tlb, ptep, address); \ >> } while (0) >> >> +#ifndef tlb_remove_huge_tlb_entry >> #define tlb_remove_huge_tlb_entry(h, tlb, ptep, address) \ >> do { \ >> unsigned long _sz = huge_page_size(h); \ >> @@ -571,6 +572,7 @@ static inline void tlb_flush_p4d_range(struct mmu_gather *tlb, >> tlb_flush_pud_range(tlb, address, _sz); \ >> __tlb_remove_tlb_entry(tlb, ptep, address); \ >> } while (0) >> +#endif > > Was this change supposed to be included in this patch? No, this should not have been included. Sorry. Thanks for taking a look.
diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c index a7c6c7498be0..f62ec4f71132 100644 --- a/fs/hugetlbfs/inode.c +++ b/fs/hugetlbfs/inode.c @@ -587,41 +587,78 @@ static void hugetlb_vmtruncate(struct inode *inode, loff_t offset) remove_inode_hugepages(inode, offset, LLONG_MAX); } +static void hugetlbfs_zero_partial_page(struct hstate *h, + struct address_space *mapping, + unsigned long start, + unsigned long end) +{ + struct page *page; + pgoff_t idx = start >> huge_page_shift(h); + + page = find_lock_page(mapping, idx); + if (!page) + return; + + start = start & ~huge_page_mask(h); + end = end & ~huge_page_mask(h); + if (!end) + end = huge_page_size(h); + + zero_user_segment(page, (unsigned int)start, (unsigned int)end); + + unlock_page(page); + put_page(page); +} + static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len) { struct hstate *h = hstate_inode(inode); loff_t hpage_size = huge_page_size(h); loff_t hole_start, hole_end; + struct address_space *mapping = inode->i_mapping; + struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode); /* - * For hole punch round up the beginning offset of the hole and - * round down the end. + * hole_start and hole_end indicate the full pages within the hole. */ hole_start = round_up(offset, hpage_size); hole_end = round_down(offset + len, hpage_size); - if (hole_end > hole_start) { - struct address_space *mapping = inode->i_mapping; - struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode); + inode_lock(inode); + + /* protected by i_rwsem */ + if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) { + inode_unlock(inode); + return -EPERM; + } - inode_lock(inode); + i_mmap_lock_write(mapping); - /* protected by i_rwsem */ - if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) { - inode_unlock(inode); - return -EPERM; - } + /* If range starts before first full page, zero partial page. */ + if (offset < hole_start) + hugetlbfs_zero_partial_page(h, mapping, + offset, min(offset + len, hole_start)); - i_mmap_lock_write(mapping); + /* Unmap users of full pages in the hole. */ + if (hole_end > hole_start) { if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)) hugetlb_vmdelete_list(&mapping->i_mmap, hole_start >> PAGE_SHIFT, hole_end >> PAGE_SHIFT); - i_mmap_unlock_write(mapping); - remove_inode_hugepages(inode, hole_start, hole_end); - inode_unlock(inode); } + /* If range extends beyond last full page, zero partial page. */ + if ((offset + len) > hole_end && (offset + len) > hole_start) + hugetlbfs_zero_partial_page(h, mapping, + hole_end, offset + len); + + i_mmap_unlock_write(mapping); + + /* Remove full pages from the file. */ + if (hole_end > hole_start) + remove_inode_hugepages(inode, hole_start, hole_end); + inode_unlock(inode); + return 0; } diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h index 2c68a545ffa7..4622ee45f739 100644 --- a/include/asm-generic/tlb.h +++ b/include/asm-generic/tlb.h @@ -562,6 +562,7 @@ static inline void tlb_flush_p4d_range(struct mmu_gather *tlb, __tlb_remove_tlb_entry(tlb, ptep, address); \ } while (0) +#ifndef tlb_remove_huge_tlb_entry #define tlb_remove_huge_tlb_entry(h, tlb, ptep, address) \ do { \ unsigned long _sz = huge_page_size(h); \ @@ -571,6 +572,7 @@ static inline void tlb_flush_p4d_range(struct mmu_gather *tlb, tlb_flush_pud_range(tlb, address, _sz); \ __tlb_remove_tlb_entry(tlb, ptep, address); \ } while (0) +#endif /** * tlb_remove_pmd_tlb_entry - remember a pmd mapping for later tlb invalidation
hugetlbfs fallocate support was originally added with commit 70c3547e36f5 ("hugetlbfs: add hugetlbfs_fallocate()"). Initial support only operated on whole hugetlb pages. This makes sense for populating files as other interfaces such as mmap and truncate require hugetlb page size alignment. Only operating on whole hugetlb pages for the hole punch case was a simplification and there was no compelling use case to zero partial pages. In a recent discussion[1] it was assumed that hugetlbfs hole punch would zero partial hugetlb pages as that is in line with the man page description saying 'partial filesystem blocks are zeroed'. However, the hugetlbfs hole punch code actually does this: hole_start = round_up(offset, hpage_size); hole_end = round_down(offset + len, hpage_size); Modify code to zero partial hugetlb pages in hole punch range. It is possible that application code could note a change in behavior. However, that would imply the code is passing in an unaligned range and expecting only whole pages be removed. This is unlikely as the fallocate documentation states the opposite. The current hugetlbfs fallocate hole punch behavior is tested with the libhugetlbfs test fallocate_align[2]. This test will be updated to validate partial page zeroing. [1] https://lore.kernel.org/linux-mm/20571829-9d3d-0b48-817c-b6b15565f651@redhat.com/ [2] https://github.com/libhugetlbfs/libhugetlbfs/blob/master/tests/fallocate_align.c Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com> --- fs/hugetlbfs/inode.c | 67 ++++++++++++++++++++++++++++++--------- include/asm-generic/tlb.h | 2 ++ 2 files changed, 54 insertions(+), 15 deletions(-)