[LU-13528] sanity test_69: read succeeded, expect -ENOENT

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.14.0
• Severity: 3
• Assignee: Sebastien Buisson
• Reporter: Maloo

Description

This issue was created by maloo for S Buisson <[e-mail]>

This issue relates to the following test suite run: https://testing.whamcloud.com/test_sets/2be4adce-b792-4432-8f8b-6caba6148d68

test_69 failed with the following error:

read succeeded, expect -ENOENT

This error occurs only on aarch64 platforms. It seems the fail_loc value set on server side has no effect.

VVVVVVV DO NOT REMOVE LINES BELOW, Added by Maloo for auto-association VVVVVVV
sanity test_69 - read succeeded, expect -ENOENT

Comments

[Peter Jones]

Landed for 2.14

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/38526/
Subject: LU-13528 llite: prevent MAX_DIO_SIZE 32-bit truncation
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 8cfd5be8b04bdaaa4a0c794392cc2e6835e103eb

[Gerrit Updater]

Sebastien Buisson ([e-mail]) uploaded a new patch: https://review.whamcloud.com/38526
Subject: LU-13528 llite: prevent MAX_DIO_SIZE 32-bit truncation
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 0e0984dcdd4a1f0b3c0b619793c31131704519be

[Sebastien Buisson]

After further investigation, it appears that a bug with the Lustre code on aarch64 platforms turns the O_DIRECT IO into a regular IO, hence bypassing the expected fail_loc behavior in sanity test_69.

The bug stems from the fact that the MAX_DIO_SIZE value is truncated to its lower 32 bits.
Its definition is:

#define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_SIZE) & \
		      ~(DT_MAX_BRW_SIZE - 1))

On 4kB PAGE_SIZE platforms, it goes unnoticed as the value fits into 32 bits. But on 64kB PAGE_SIZE platforms, this value can go up to 1365GB.

This problem, in combination with the fact that patch https://review.whamcloud.com/36144 increased the size of struct brw_page to 32 bytes, makes 1024GB (64 bits) truncated to 0 (lower 32 bits). The consequence is that ll_direct_IO_impl() triggers a direct IO of 0 bytes, which in turn makes llite issue a normal, buffered IO.