| Message ID | 20190130211443.16678-1-mike.kravetz@oracle.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | huegtlbfs: fix page leak during migration of file pages | expand |
Hi,
[This is an automated email]
This commit has been processed because it contains a "Fixes:" tag,
fixing commit: 290408d4a250 hugetlb: hugepage migration core.
The bot has tested the following trees: v4.20.5, v4.19.18, v4.14.96, v4.9.153, v4.4.172, v3.18.133.
v4.20.5: Build OK!
v4.19.18: Build OK!
v4.14.96: Build OK!
v4.9.153: Failed to apply! Possible dependencies:
2916ecc0f9d4 ("mm/migrate: new migrate mode MIGRATE_SYNC_NO_COPY")
v4.4.172: Failed to apply! Possible dependencies:
09cbfeaf1a5a ("mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros")
0e749e54244e ("dax: increase granularity of dax_clear_blocks() operations")
2916ecc0f9d4 ("mm/migrate: new migrate mode MIGRATE_SYNC_NO_COPY")
2a28900be206 ("udf: Export superblock magic to userspace")
4420cfd3f51c ("staging: lustre: format properly all comment blocks for LNet core")
48b4800a1c6a ("zsmalloc: page migration support")
5057dcd0f1aa ("virtio_balloon: export 'available' memory to balloon statistics")
52db400fcd50 ("pmem, dax: clean up clear_pmem()")
5b7a487cf32d ("f2fs: add customized migrate_page callback")
5fd88337d209 ("staging: lustre: fix all conditional comparison to zero in LNet layer")
a188222b6ed2 ("net: Rename NETIF_F_ALL_CSUM to NETIF_F_CSUM_MASK")
b1123ea6d3b3 ("mm: balloon: use general non-lru movable page feature")
b2e0d1625e19 ("dax: fix lifetime of in-kernel dax mappings with dax_map_atomic()")
bda807d44454 ("mm: migrate: support non-lru movable page migration")
c8b8e32d700f ("direct-io: eliminate the offset argument to ->direct_IO")
d1a5f2b4d8a1 ("block: use DAX for partition table reads")
e10624f8c097 ("pmem: fail io-requests to known bad blocks")
v3.18.133: Failed to apply! Possible dependencies:
0722b1011a5f ("f2fs: set page private for inmemory pages for truncation")
1601839e9e5b ("f2fs: fix to release count of meta page in ->invalidatepage")
2916ecc0f9d4 ("mm/migrate: new migrate mode MIGRATE_SYNC_NO_COPY")
31a3268839c1 ("f2fs: cleanup if-statement of phase in gc_data_segment")
34ba94bac938 ("f2fs: do not make dirty any inmemory pages")
34d67debe02b ("f2fs: add infra struct and helper for inline dir")
4634d71ed190 ("f2fs: fix missing kmem_cache_free")
487261f39bcd ("f2fs: merge {invalidate,release}page for meta/node/data pages")
5b7a487cf32d ("f2fs: add customized migrate_page callback")
67298804f344 ("f2fs: introduce struct inode_management to wrap inner fields")
769ec6e5b7d4 ("f2fs: call radix_tree_preload before radix_tree_insert")
7dda2af83b2b ("f2fs: more fast lookup for gc_inode list")
8b26ef98da33 ("f2fs: use rw_semaphore for nat entry lock")
8c402946f074 ("f2fs: introduce the number of inode entries")
9be32d72becc ("f2fs: do retry operations with cond_resched")
9e4ded3f309e ("f2fs: activate f2fs_trace_pid")
d5053a34a9cc ("f2fs: introduce -o fastboot for reducing booting time only")
e5e7ea3c86e5 ("f2fs: control the memory footprint used by ino entries")
f68daeebba5a ("f2fs: keep PagePrivate during releasepage")
How should we proceed with this patch?
--
Thanks,
Sasha
On 1/31/19 6:12 AM, Sasha Levin wrote: > Hi, > > [This is an automated email] > > This commit has been processed because it contains a "Fixes:" tag, > fixing commit: 290408d4a250 hugetlb: hugepage migration core. > > The bot has tested the following trees: v4.20.5, v4.19.18, v4.14.96, v4.9.153, v4.4.172, v3.18.133. > > v4.20.5: Build OK! > v4.19.18: Build OK! > v4.14.96: Build OK! > v4.9.153: Failed to apply! Possible dependencies: > 2916ecc0f9d4 ("mm/migrate: new migrate mode MIGRATE_SYNC_NO_COPY") > > v4.4.172: Failed to apply! Possible dependencies: > 09cbfeaf1a5a ("mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros") > 0e749e54244e ("dax: increase granularity of dax_clear_blocks() operations") > 2916ecc0f9d4 ("mm/migrate: new migrate mode MIGRATE_SYNC_NO_COPY") > 2a28900be206 ("udf: Export superblock magic to userspace") > 4420cfd3f51c ("staging: lustre: format properly all comment blocks for LNet core") > 48b4800a1c6a ("zsmalloc: page migration support") > 5057dcd0f1aa ("virtio_balloon: export 'available' memory to balloon statistics") > 52db400fcd50 ("pmem, dax: clean up clear_pmem()") > 5b7a487cf32d ("f2fs: add customized migrate_page callback") > 5fd88337d209 ("staging: lustre: fix all conditional comparison to zero in LNet layer") > a188222b6ed2 ("net: Rename NETIF_F_ALL_CSUM to NETIF_F_CSUM_MASK") > b1123ea6d3b3 ("mm: balloon: use general non-lru movable page feature") > b2e0d1625e19 ("dax: fix lifetime of in-kernel dax mappings with dax_map_atomic()") > bda807d44454 ("mm: migrate: support non-lru movable page migration") > c8b8e32d700f ("direct-io: eliminate the offset argument to ->direct_IO") > d1a5f2b4d8a1 ("block: use DAX for partition table reads") > e10624f8c097 ("pmem: fail io-requests to known bad blocks") > > v3.18.133: Failed to apply! Possible dependencies: > 0722b1011a5f ("f2fs: set page private for inmemory pages for truncation") > 1601839e9e5b ("f2fs: fix to release count of meta page in ->invalidatepage") > 2916ecc0f9d4 ("mm/migrate: new migrate mode MIGRATE_SYNC_NO_COPY") > 31a3268839c1 ("f2fs: cleanup if-statement of phase in gc_data_segment") > 34ba94bac938 ("f2fs: do not make dirty any inmemory pages") > 34d67debe02b ("f2fs: add infra struct and helper for inline dir") > 4634d71ed190 ("f2fs: fix missing kmem_cache_free") > 487261f39bcd ("f2fs: merge {invalidate,release}page for meta/node/data pages") > 5b7a487cf32d ("f2fs: add customized migrate_page callback") > 67298804f344 ("f2fs: introduce struct inode_management to wrap inner fields") > 769ec6e5b7d4 ("f2fs: call radix_tree_preload before radix_tree_insert") > 7dda2af83b2b ("f2fs: more fast lookup for gc_inode list") > 8b26ef98da33 ("f2fs: use rw_semaphore for nat entry lock") > 8c402946f074 ("f2fs: introduce the number of inode entries") > 9be32d72becc ("f2fs: do retry operations with cond_resched") > 9e4ded3f309e ("f2fs: activate f2fs_trace_pid") > d5053a34a9cc ("f2fs: introduce -o fastboot for reducing booting time only") > e5e7ea3c86e5 ("f2fs: control the memory footprint used by ino entries") > f68daeebba5a ("f2fs: keep PagePrivate during releasepage") > > > How should we proceed with this patch? Hello automated Sasha, First, let's wait for review/ack. However, the patch does not strictly 'depend' on the functionality of the commits in the lists above. If/when this goes upstream I can provide backports for 4.9, 4.4 and 3.18.
On 1/30/19 1:14 PM, Mike Kravetz wrote: > Files can be created and mapped in an explicitly mounted hugetlbfs > filesystem. If pages in such files are migrated, the filesystem > usage will not be decremented for the associated pages. This can > result in mmap or page allocation failures as it appears there are > fewer pages in the filesystem than there should be. Does anyone have a little time to take a look at this? While migration of hugetlb pages 'should' not be a common issue, we have seen it happen via soft memory errors/page poisoning in production environments. Didn't see a leak in that case as it was with pages in a Sys V shared mem segment. However, our DB code is starting to make use of files in explicitly mounted hugetlbfs filesystems. Therefore, we are more likely to hit this bug in the field.
On Thu, Feb 07, 2019 at 10:50:55AM -0800, Mike Kravetz wrote: > On 1/30/19 1:14 PM, Mike Kravetz wrote: > > Files can be created and mapped in an explicitly mounted hugetlbfs > > filesystem. If pages in such files are migrated, the filesystem > > usage will not be decremented for the associated pages. This can > > result in mmap or page allocation failures as it appears there are > > fewer pages in the filesystem than there should be. > > Does anyone have a little time to take a look at this? > > While migration of hugetlb pages 'should' not be a common issue, we > have seen it happen via soft memory errors/page poisoning in production > environments. Didn't see a leak in that case as it was with pages in a > Sys V shared mem segment. However, our DB code is starting to make use > of files in explicitly mounted hugetlbfs filesystems. Therefore, we are > more likely to hit this bug in the field. Hi Mike, Thank you for finding/reporting the problem. # sorry for my late response. > > > > > For example, a test program which hole punches, faults and migrates > > pages in such a file (1G in size) will eventually fail because it > > can not allocate a page. Reported counts and usage at time of failure: > > > > node0 > > 537 free_hugepages > > 1024 nr_hugepages > > 0 surplus_hugepages > > node1 > > 1000 free_hugepages > > 1024 nr_hugepages > > 0 surplus_hugepages > > > > Filesystem Size Used Avail Use% Mounted on > > nodev 4.0G 4.0G 0 100% /var/opt/hugepool > > > > Note that the filesystem shows 4G of pages used, while actual usage is > > 511 pages (just under 1G). Failed trying to allocate page 512. > > > > If a hugetlb page is associated with an explicitly mounted filesystem, > > this information in contained in the page_private field. At migration > > time, this information is not preserved. To fix, simply transfer > > page_private from old to new page at migration time if necessary. Also, > > migrate_page_states() unconditionally clears page_private and PagePrivate > > of the old page. It is unlikely, but possible that these fields could > > be non-NULL and are needed at hugetlb free page time. So, do not touch > > these fields for hugetlb pages. > > > > Cc: <stable@vger.kernel.org> > > Fixes: 290408d4a250 ("hugetlb: hugepage migration core") > > Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com> > > --- > > fs/hugetlbfs/inode.c | 10 ++++++++++ > > mm/migrate.c | 10 ++++++++-- > > 2 files changed, 18 insertions(+), 2 deletions(-) > > > > diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c > > index 32920a10100e..fb6de1db8806 100644 > > --- a/fs/hugetlbfs/inode.c > > +++ b/fs/hugetlbfs/inode.c > > @@ -859,6 +859,16 @@ static int hugetlbfs_migrate_page(struct address_space *mapping, > > rc = migrate_huge_page_move_mapping(mapping, newpage, page); > > if (rc != MIGRATEPAGE_SUCCESS) > > return rc; > > + > > + /* > > + * page_private is subpool pointer in hugetlb pages, transfer > > + * if needed. > > + */ > > + if (page_private(page) && !page_private(newpage)) { > > + set_page_private(newpage, page_private(page)); > > + set_page_private(page, 0); You don't have to copy PagePrivate flag? > > + } > > + > > if (mode != MIGRATE_SYNC_NO_COPY) > > migrate_page_copy(newpage, page); > > else > > diff --git a/mm/migrate.c b/mm/migrate.c > > index f7e4bfdc13b7..0d9708803553 100644 > > --- a/mm/migrate.c > > +++ b/mm/migrate.c > > @@ -703,8 +703,14 @@ void migrate_page_states(struct page *newpage, struct page *page) > > */ > > if (PageSwapCache(page)) > > ClearPageSwapCache(page); > > - ClearPagePrivate(page); > > - set_page_private(page, 0); > > + /* > > + * Unlikely, but PagePrivate and page_private could potentially > > + * contain information needed at hugetlb free page time. > > + */ > > + if (!PageHuge(page)) { > > + ClearPagePrivate(page); > > + set_page_private(page, 0); > > + } # This argument is mainly for existing code... According to the comment on migrate_page(): /* * Common logic to directly migrate a single LRU page suitable for * pages that do not use PagePrivate/PagePrivate2. * * Pages are locked upon entry and exit. */ int migrate_page(struct address_space *mapping, ... So this common logic assumes that page_private is not used, so why do we explicitly clear page_private in migrate_page_states()? buffer_migrate_page(), which is commonly used for the case when page_private is used, does that clearing outside migrate_page_states(). So I thought that hugetlbfs_migrate_page() could do in the similar manner. IOW, migrate_page_states() should not do anything on PagePrivate. But there're a few other .migratepage callbacks, and I'm not sure all of them are safe for the change, so this approach might not fit for a small fix. # BTW, there seems a typo in $SUBJECT. Thanks, Naoya Horiguchi
On 2/7/19 6:31 PM, Naoya Horiguchi wrote: > On Thu, Feb 07, 2019 at 10:50:55AM -0800, Mike Kravetz wrote: >> On 1/30/19 1:14 PM, Mike Kravetz wrote: >>> +++ b/fs/hugetlbfs/inode.c >>> @@ -859,6 +859,16 @@ static int hugetlbfs_migrate_page(struct address_space *mapping, >>> rc = migrate_huge_page_move_mapping(mapping, newpage, page); >>> if (rc != MIGRATEPAGE_SUCCESS) >>> return rc; >>> + >>> + /* >>> + * page_private is subpool pointer in hugetlb pages, transfer >>> + * if needed. >>> + */ >>> + if (page_private(page) && !page_private(newpage)) { >>> + set_page_private(newpage, page_private(page)); >>> + set_page_private(page, 0); > > You don't have to copy PagePrivate flag? > Well my original thought was no. For hugetlb pages, PagePrivate is not associated with page_private. It indicates a reservation was consumed. It is set when a hugetlb page is newly allocated and the allocation is associated with a reservation and the global reservation count is decremented. When the page is added to the page cache or rmap, PagePrivate is cleared. If the page is free'ed before being added to page cache or rmap, PagePrivate tells free_huge_page to restore (increment) the reserve count as we did not 'instantiate' the page. So, PagePrivate is only set from the time a huge page is allocated until it is added to page cache or rmap. My original thought was that the page could not be migrated during this time. However, I am not sure if that reasoning is correct. The page is not locked, so it would appear that it could be migrated? But, if it can be migrated at this time then perhaps there are bigger issues for the (hugetlb) page fault code? >>> + >>> + } >>> + >>> if (mode != MIGRATE_SYNC_NO_COPY) >>> migrate_page_copy(newpage, page); >>> else >>> diff --git a/mm/migrate.c b/mm/migrate.c >>> index f7e4bfdc13b7..0d9708803553 100644 >>> --- a/mm/migrate.c >>> +++ b/mm/migrate.c >>> @@ -703,8 +703,14 @@ void migrate_page_states(struct page *newpage, struct page *page) >>> */ >>> if (PageSwapCache(page)) >>> ClearPageSwapCache(page); >>> - ClearPagePrivate(page); >>> - set_page_private(page, 0); >>> + /* >>> + * Unlikely, but PagePrivate and page_private could potentially >>> + * contain information needed at hugetlb free page time. >>> + */ >>> + if (!PageHuge(page)) { >>> + ClearPagePrivate(page); >>> + set_page_private(page, 0); >>> + } > > # This argument is mainly for existing code... > > According to the comment on migrate_page(): > > /* > * Common logic to directly migrate a single LRU page suitable for > * pages that do not use PagePrivate/PagePrivate2. > * > * Pages are locked upon entry and exit. > */ > int migrate_page(struct address_space *mapping, ... > > So this common logic assumes that page_private is not used, so why do > we explicitly clear page_private in migrate_page_states()? Perhaps someone else knows. If not, I can do some git research and try to find out why. > buffer_migrate_page(), which is commonly used for the case when > page_private is used, does that clearing outside migrate_page_states(). > So I thought that hugetlbfs_migrate_page() could do in the similar manner. > IOW, migrate_page_states() should not do anything on PagePrivate. > But there're a few other .migratepage callbacks, and I'm not sure all of > them are safe for the change, so this approach might not fit for a small fix. I will look at those as well unless someone knows without researching. > > # BTW, there seems a typo in $SUBJECT. Thanks!
On Thu, Feb 07, 2019 at 09:50:30PM -0800, Mike Kravetz wrote: > On 2/7/19 6:31 PM, Naoya Horiguchi wrote: > > On Thu, Feb 07, 2019 at 10:50:55AM -0800, Mike Kravetz wrote: > >> On 1/30/19 1:14 PM, Mike Kravetz wrote: > >>> +++ b/fs/hugetlbfs/inode.c > >>> @@ -859,6 +859,16 @@ static int hugetlbfs_migrate_page(struct address_space *mapping, > >>> rc = migrate_huge_page_move_mapping(mapping, newpage, page); > >>> if (rc != MIGRATEPAGE_SUCCESS) > >>> return rc; > >>> + > >>> + /* > >>> + * page_private is subpool pointer in hugetlb pages, transfer > >>> + * if needed. > >>> + */ > >>> + if (page_private(page) && !page_private(newpage)) { > >>> + set_page_private(newpage, page_private(page)); > >>> + set_page_private(page, 0); > > > > You don't have to copy PagePrivate flag? > > > > Well my original thought was no. For hugetlb pages, PagePrivate is not > associated with page_private. It indicates a reservation was consumed. > It is set when a hugetlb page is newly allocated and the allocation is > associated with a reservation and the global reservation count is > decremented. When the page is added to the page cache or rmap, > PagePrivate is cleared. If the page is free'ed before being added to page > cache or rmap, PagePrivate tells free_huge_page to restore (increment) the > reserve count as we did not 'instantiate' the page. > > So, PagePrivate is only set from the time a huge page is allocated until > it is added to page cache or rmap. My original thought was that the page > could not be migrated during this time. However, I am not sure if that > reasoning is correct. The page is not locked, so it would appear that it > could be migrated? But, if it can be migrated at this time then perhaps > there are bigger issues for the (hugetlb) page fault code? In my understanding, free hugetlb pages are not expected to be passed to migrate_pages(), and currently that's ensured by each migration caller which checks and avoids free hugetlb pages on its own. migrate_pages() and its internal code are probably not aware of handling free hugetlb pages, so if they are accidentally passed to migration code, that's a big problem as you are concerned. So the above reasoning should work at least this assumption is correct. Most of migration callers are not intersted in moving free hugepages. The one I'm not sure of is the code path from alloc_contig_range(). If someone think it's worthwhile to migrate free hugepage to get bigger contiguous memory, he/she tries to enable that code path and the assumption will be broken. Thanks, Naoya Horiguchi > > >>> + > >>> + } > >>> + > >>> if (mode != MIGRATE_SYNC_NO_COPY) > >>> migrate_page_copy(newpage, page); > >>> else > >>> diff --git a/mm/migrate.c b/mm/migrate.c > >>> index f7e4bfdc13b7..0d9708803553 100644 > >>> --- a/mm/migrate.c > >>> +++ b/mm/migrate.c > >>> @@ -703,8 +703,14 @@ void migrate_page_states(struct page *newpage, struct page *page) > >>> */ > >>> if (PageSwapCache(page)) > >>> ClearPageSwapCache(page); > >>> - ClearPagePrivate(page); > >>> - set_page_private(page, 0); > >>> + /* > >>> + * Unlikely, but PagePrivate and page_private could potentially > >>> + * contain information needed at hugetlb free page time. > >>> + */ > >>> + if (!PageHuge(page)) { > >>> + ClearPagePrivate(page); > >>> + set_page_private(page, 0); > >>> + } > > > > # This argument is mainly for existing code... > > > > According to the comment on migrate_page(): > > > > /* > > * Common logic to directly migrate a single LRU page suitable for > > * pages that do not use PagePrivate/PagePrivate2. > > * > > * Pages are locked upon entry and exit. > > */ > > int migrate_page(struct address_space *mapping, ... > > > > So this common logic assumes that page_private is not used, so why do > > we explicitly clear page_private in migrate_page_states()? > > Perhaps someone else knows. If not, I can do some git research and > try to find out why. > > > buffer_migrate_page(), which is commonly used for the case when > > page_private is used, does that clearing outside migrate_page_states(). > > So I thought that hugetlbfs_migrate_page() could do in the similar manner. > > IOW, migrate_page_states() should not do anything on PagePrivate. > > But there're a few other .migratepage callbacks, and I'm not sure all of > > them are safe for the change, so this approach might not fit for a small fix. > > I will look at those as well unless someone knows without researching. > > > > > # BTW, there seems a typo in $SUBJECT. > > Thanks! > > -- > Mike Kravetz >
On 2/7/19 11:31 PM, Naoya Horiguchi wrote: > On Thu, Feb 07, 2019 at 09:50:30PM -0800, Mike Kravetz wrote: >> On 2/7/19 6:31 PM, Naoya Horiguchi wrote: >>> On Thu, Feb 07, 2019 at 10:50:55AM -0800, Mike Kravetz wrote: >>>> On 1/30/19 1:14 PM, Mike Kravetz wrote: >>>>> +++ b/fs/hugetlbfs/inode.c >>>>> @@ -859,6 +859,16 @@ static int hugetlbfs_migrate_page(struct address_space *mapping, >>>>> rc = migrate_huge_page_move_mapping(mapping, newpage, page); >>>>> if (rc != MIGRATEPAGE_SUCCESS) >>>>> return rc; >>>>> + >>>>> + /* >>>>> + * page_private is subpool pointer in hugetlb pages, transfer >>>>> + * if needed. >>>>> + */ >>>>> + if (page_private(page) && !page_private(newpage)) { >>>>> + set_page_private(newpage, page_private(page)); >>>>> + set_page_private(page, 0); >>> >>> You don't have to copy PagePrivate flag? >>> >> >> Well my original thought was no. For hugetlb pages, PagePrivate is not >> associated with page_private. It indicates a reservation was consumed. >> It is set when a hugetlb page is newly allocated and the allocation is >> associated with a reservation and the global reservation count is >> decremented. When the page is added to the page cache or rmap, >> PagePrivate is cleared. If the page is free'ed before being added to page >> cache or rmap, PagePrivate tells free_huge_page to restore (increment) the >> reserve count as we did not 'instantiate' the page. >> >> So, PagePrivate is only set from the time a huge page is allocated until >> it is added to page cache or rmap. My original thought was that the page >> could not be migrated during this time. However, I am not sure if that >> reasoning is correct. The page is not locked, so it would appear that it >> could be migrated? But, if it can be migrated at this time then perhaps >> there are bigger issues for the (hugetlb) page fault code? > > In my understanding, free hugetlb pages are not expected to be passed to > migrate_pages(), and currently that's ensured by each migration caller > which checks and avoids free hugetlb pages on its own. > migrate_pages() and its internal code are probably not aware of handling > free hugetlb pages, so if they are accidentally passed to migration code, > that's a big problem as you are concerned. > So the above reasoning should work at least this assumption is correct. > > Most of migration callers are not intersted in moving free hugepages. > The one I'm not sure of is the code path from alloc_contig_range(). > If someone think it's worthwhile to migrate free hugepage to get bigger > contiguous memory, he/she tries to enable that code path and the assumption > will be broken. You are correct. We do not migrate free huge pages. I was thinking more about problems if we migrate a page while it is being added to a task's page table as in hugetlb_no_page. Commit bcc54222309c ("mm: hugetlb: introduce page_huge_active") addresses this issue, but I believe there is a bug in the implementation. isolate_huge_page contains this test: if (!page_huge_active(page) || !get_page_unless_zero(page)) { ret = false; goto unlock; } If the condition is not met, then the huge page can be isolated and migrated. In hugetlb_no_page, there is this block of code: page = alloc_huge_page(vma, haddr, 0); if (IS_ERR(page)) { ret = vmf_error(PTR_ERR(page)); goto out; } clear_huge_page(page, address, pages_per_huge_page(h)); __SetPageUptodate(page); set_page_huge_active(page); if (vma->vm_flags & VM_MAYSHARE) { int err = huge_add_to_page_cache(page, mapping, idx); if (err) { put_page(page); if (err == -EEXIST) goto retry; goto out; } } else { lock_page(page); if (unlikely(anon_vma_prepare(vma))) { ret = VM_FAULT_OOM; goto backout_unlocked; } anon_rmap = 1; } } else { Note that we call set_page_huge_active BEFORE locking the page. This means that we can isolate the page and have migration take place while we continue to add the page to page tables. I was able to make this happen by adding a udelay() after set_page_huge_active to simulate worst case scheduling behavior. It resulted in VM_BUG_ON while unlocking page. My test had several threads faulting in huge pages. Another thread was offlining the memory blocks forcing migration. To fix this, we need to delay the set_page_huge_active call until after the page is locked. I am testing a patch with this change. Perhaps we should even delay calling set_page_huge_active until we know there are no errors and we know the page is actually in page tables? While looking at this, I think there is another issue. When a hugetlb page is migrated, we do not migrate the 'page_huge_active' state of the page. That should be moved as the page is migrated. Correct?
On Mon, Feb 11, 2019 at 03:06:27PM -0800, Mike Kravetz wrote: > On 2/7/19 11:31 PM, Naoya Horiguchi wrote: > > On Thu, Feb 07, 2019 at 09:50:30PM -0800, Mike Kravetz wrote: > >> On 2/7/19 6:31 PM, Naoya Horiguchi wrote: > >>> On Thu, Feb 07, 2019 at 10:50:55AM -0800, Mike Kravetz wrote: > >>>> On 1/30/19 1:14 PM, Mike Kravetz wrote: > >>>>> +++ b/fs/hugetlbfs/inode.c > >>>>> @@ -859,6 +859,16 @@ static int hugetlbfs_migrate_page(struct address_space *mapping, > >>>>> rc = migrate_huge_page_move_mapping(mapping, newpage, page); > >>>>> if (rc != MIGRATEPAGE_SUCCESS) > >>>>> return rc; > >>>>> + > >>>>> + /* > >>>>> + * page_private is subpool pointer in hugetlb pages, transfer > >>>>> + * if needed. > >>>>> + */ > >>>>> + if (page_private(page) && !page_private(newpage)) { > >>>>> + set_page_private(newpage, page_private(page)); > >>>>> + set_page_private(page, 0); > >>> > >>> You don't have to copy PagePrivate flag? > >>> > >> > >> Well my original thought was no. For hugetlb pages, PagePrivate is not > >> associated with page_private. It indicates a reservation was consumed. > >> It is set when a hugetlb page is newly allocated and the allocation is > >> associated with a reservation and the global reservation count is > >> decremented. When the page is added to the page cache or rmap, > >> PagePrivate is cleared. If the page is free'ed before being added to page > >> cache or rmap, PagePrivate tells free_huge_page to restore (increment) the > >> reserve count as we did not 'instantiate' the page. > >> > >> So, PagePrivate is only set from the time a huge page is allocated until > >> it is added to page cache or rmap. My original thought was that the page > >> could not be migrated during this time. However, I am not sure if that > >> reasoning is correct. The page is not locked, so it would appear that it > >> could be migrated? But, if it can be migrated at this time then perhaps > >> there are bigger issues for the (hugetlb) page fault code? > > > > In my understanding, free hugetlb pages are not expected to be passed to > > migrate_pages(), and currently that's ensured by each migration caller > > which checks and avoids free hugetlb pages on its own. > > migrate_pages() and its internal code are probably not aware of handling > > free hugetlb pages, so if they are accidentally passed to migration code, > > that's a big problem as you are concerned. > > So the above reasoning should work at least this assumption is correct. > > > > Most of migration callers are not intersted in moving free hugepages. > > The one I'm not sure of is the code path from alloc_contig_range(). > > If someone think it's worthwhile to migrate free hugepage to get bigger > > contiguous memory, he/she tries to enable that code path and the assumption > > will be broken. > > You are correct. We do not migrate free huge pages. I was thinking more > about problems if we migrate a page while it is being added to a task's page > table as in hugetlb_no_page. > > Commit bcc54222309c ("mm: hugetlb: introduce page_huge_active") addresses > this issue, but I believe there is a bug in the implementation. > isolate_huge_page contains this test: > > if (!page_huge_active(page) || !get_page_unless_zero(page)) { > ret = false; > goto unlock; > } > > If the condition is not met, then the huge page can be isolated and migrated. > > In hugetlb_no_page, there is this block of code: > > page = alloc_huge_page(vma, haddr, 0); > if (IS_ERR(page)) { > ret = vmf_error(PTR_ERR(page)); > goto out; > } > clear_huge_page(page, address, pages_per_huge_page(h)); > __SetPageUptodate(page); > set_page_huge_active(page); > > if (vma->vm_flags & VM_MAYSHARE) { > int err = huge_add_to_page_cache(page, mapping, idx); > if (err) { > put_page(page); > if (err == -EEXIST) > goto retry; > goto out; > } > } else { > lock_page(page); > if (unlikely(anon_vma_prepare(vma))) { > ret = VM_FAULT_OOM; > goto backout_unlocked; > } > anon_rmap = 1; > } > } else { > > Note that we call set_page_huge_active BEFORE locking the page. This > means that we can isolate the page and have migration take place while > we continue to add the page to page tables. I was able to make this > happen by adding a udelay() after set_page_huge_active to simulate worst > case scheduling behavior. It resulted in VM_BUG_ON while unlocking page. > My test had several threads faulting in huge pages. Another thread was > offlining the memory blocks forcing migration. This shows another problem, so I agree we need a fix. > > To fix this, we need to delay the set_page_huge_active call until after > the page is locked. I am testing a patch with this change. Perhaps we > should even delay calling set_page_huge_active until we know there are > no errors and we know the page is actually in page tables? Yes, calling set_page_huge_active after page table is set up sounds nice to me. > > While looking at this, I think there is another issue. When a hugetlb > page is migrated, we do not migrate the 'page_huge_active' state of the > page. That should be moved as the page is migrated. Correct? Yes, and I think that putback_active_hugepage(new_hpage) at the last step of migration sequence handles the copying of 'page_huge_active' state. Thanks, Naoya Horiguchi
On 2/11/19 6:24 PM, Naoya Horiguchi wrote: > On Mon, Feb 11, 2019 at 03:06:27PM -0800, Mike Kravetz wrote: >> While looking at this, I think there is another issue. When a hugetlb >> page is migrated, we do not migrate the 'page_huge_active' state of the >> page. That should be moved as the page is migrated. Correct? > > Yes, and I think that putback_active_hugepage(new_hpage) at the last step > of migration sequence handles the copying of 'page_huge_active' state. > Thanks! I missed the putback_active_hugepage that takes care of making the target migration page active.
diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c index 32920a10100e..fb6de1db8806 100644 --- a/fs/hugetlbfs/inode.c +++ b/fs/hugetlbfs/inode.c @@ -859,6 +859,16 @@ static int hugetlbfs_migrate_page(struct address_space *mapping, rc = migrate_huge_page_move_mapping(mapping, newpage, page); if (rc != MIGRATEPAGE_SUCCESS) return rc; + + /* + * page_private is subpool pointer in hugetlb pages, transfer + * if needed. + */ + if (page_private(page) && !page_private(newpage)) { + set_page_private(newpage, page_private(page)); + set_page_private(page, 0); + } + if (mode != MIGRATE_SYNC_NO_COPY) migrate_page_copy(newpage, page); else diff --git a/mm/migrate.c b/mm/migrate.c index f7e4bfdc13b7..0d9708803553 100644 --- a/mm/migrate.c +++ b/mm/migrate.c @@ -703,8 +703,14 @@ void migrate_page_states(struct page *newpage, struct page *page) */ if (PageSwapCache(page)) ClearPageSwapCache(page); - ClearPagePrivate(page); - set_page_private(page, 0); + /* + * Unlikely, but PagePrivate and page_private could potentially + * contain information needed at hugetlb free page time. + */ + if (!PageHuge(page)) { + ClearPagePrivate(page); + set_page_private(page, 0); + } /* * If any waiters have accumulated on the new page then
Files can be created and mapped in an explicitly mounted hugetlbfs filesystem. If pages in such files are migrated, the filesystem usage will not be decremented for the associated pages. This can result in mmap or page allocation failures as it appears there are fewer pages in the filesystem than there should be. For example, a test program which hole punches, faults and migrates pages in such a file (1G in size) will eventually fail because it can not allocate a page. Reported counts and usage at time of failure: node0 537 free_hugepages 1024 nr_hugepages 0 surplus_hugepages node1 1000 free_hugepages 1024 nr_hugepages 0 surplus_hugepages Filesystem Size Used Avail Use% Mounted on nodev 4.0G 4.0G 0 100% /var/opt/hugepool Note that the filesystem shows 4G of pages used, while actual usage is 511 pages (just under 1G). Failed trying to allocate page 512. If a hugetlb page is associated with an explicitly mounted filesystem, this information in contained in the page_private field. At migration time, this information is not preserved. To fix, simply transfer page_private from old to new page at migration time if necessary. Also, migrate_page_states() unconditionally clears page_private and PagePrivate of the old page. It is unlikely, but possible that these fields could be non-NULL and are needed at hugetlb free page time. So, do not touch these fields for hugetlb pages. Cc: <stable@vger.kernel.org> Fixes: 290408d4a250 ("hugetlb: hugepage migration core") Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com> --- fs/hugetlbfs/inode.c | 10 ++++++++++ mm/migrate.c | 10 ++++++++-- 2 files changed, 18 insertions(+), 2 deletions(-)