[LU-12587] DIO fallback to Buffer IO unexpectedly

• Status: Reopened
• Project: Lustre
• Type: Bug
• Priority: Minor
• Reporter: Wang Shilong (Inactive)
• Assignee: Patrick Farrell
• Resolution: Unresolved
• Labels: always_except
• Severity: 3

Description

It is possible that DirectIO might fallback to Buffer IO in some cases, but following
DIO fallback to Buffer IO very easily:

dd if=/dev/zero of=data bs=1M count=1
dd iff=/dev/zero of=data bs=1M count=1 oflag=direct conv=notrunc

Issue Links

• is related to: LU-14541 Memory reclaim caused a stale data read
• is related to: LU-15815 fast_read/stale data/reclaim workroun...

Comments

[Gerrit Updater]

Wang Shilong ([e-mail]) uploaded a new patch: https://review.whamcloud.com/35610
Subject: LU-12587 llite: don't check vmpage refcount in ll_releasepage()
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: eeb3f3884e0db092742daf88fc54422e207e5270

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/35610/
Subject: LU-12587 llite: don't check vmpage refcount in ll_releasepage()
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: e59f0c9a245f91ad929fd31dcbcbee88bf7a11df

[Peter Jones]

Landed for 2.13

[Patrick Farrell]

This change introduced LU-14541.  The original check here is a bit off, so fixing this is not just a revert, but mostly.  See https://review.whamcloud.com/47262

We will have to see if there's another way to handle this issue.  It might be possible to do the flushing here (see https://review.whamcloud.com/35610/ ) inside Lustre rather than relying on the kernel to do it.

eg, if we flushed pages before generic_file_direct_write:

->generic_file_direct_write()

->filemap_write_and_wait_range()

->invalidate_inode_pages2_range()

->invalidate_complete_page2() If a page can not be invalidated, return 0 to fall back to buffered write.

->try_to_release_page()

->ll_releasepage() return 0 becuase of vmpage count is 4 > 3

->generic_file_buffered_write

It may just be that Lustre can't release pages in the desired manner...  This will require some more investigation.

[Andreas Dilger]

sanity test_277 is being skipped on master because of LU-14541, but it should really be referencing this ticket.