[LU-14423] osd_is_mapped() does not recognize holes

• Status: Resolved
• Project: Lustre
• Fix Version/s: Lustre 2.15.0
• Type: Bug
• Priority: Minor
• Reporter: Alex Zhuravlev
• Assignee: Alex Zhuravlev
• Resolution: Fixed
• Severity: 3

Description

ldiskfs_fiemap() can return a hole in form

{0,0,0}

and then the following construction recognizes this as an allocated space:

        if (start > block) {
                cached_extent->start = block;
                cached_extent->end = start;
                cached_extent->mapped = 0;
        } else {
                cached_extent->start = start;
                cached_extent->end = (fe.fe_logical + fe.fe_length) >>
                                      inode->i_blkbits;
                cached_extent->mapped = 1;
        }

Comments

[Gerrit Updater]

Alex Zhuravlev ([e-mail]) uploaded a new patch: https://review.whamcloud.com/41481
Subject: LU-14423 osd: recognize holes in osd_is_mapped()
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 1ce62777a493b7496005c668677bddaae5a12cf8

[Andreas Dilger]

Alex, AFAIK FIEMAP should not be returning 0,0,0 as a hole. That is just an uninitialized buffer and should be outside the number of returned extents? Holes in the file should just be skipped in the returned extents.

[Alex Zhuravlev]

hmm, this is what I exactly got using linux-3.10.0-957.12.2.el7 - ldiskfs_fiemap() returned 0 and

{.fe_logical=0, fe_length=0}

for non-allocated block.
the file's size was set using truncate.

[Andreas Dilger]

What did it return for the extent count? This may be a bug in ldiskfs.

[Alex Zhuravlev]

What did it return for the extent count? This may be a bug in ldiskfs

fei.fi_extents_mapped was not touched in this case. i.e. it would be 0 for the first call to ldiskfs_fiemap() and 1 if the previous call found an extemt.
fe.fe_* were reset to 0 though.

[Alex Zhuravlev]

AFAICS, fiemap can't return an extent covering a hole to EOF? thus we'd have to lookup each block separately?

[Andreas Dilger]

Correct - FIEMAP does not return anything for a hole. If there are no blocks allocated to a file at all, then it should return no extent in that case. I guess you could use the 0,0 return to detect that, though it is a bit non-standard.

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/41481/
Subject: LU-14423 osd: recognize holes in osd_is_mapped()
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 2eaa49ef0f16798d564883b16cea9e96fad52495

[Peter Jones]

Landed for 2.15

[Gerrit Updater]

"Andreas Dilger <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/45504
Subject: LU-14423 osd: recognize holes in osd_is_mapped()
Project: fs/lustre-release
Branch: b2_14
Current Patch Set: 1
Commit: 543fc7629b06a3cffb8d5ee1830c09a86f91e317

[Gerrit Updater]

"Andreas Dilger <[e-mail]>" merged in patch https://review.whamcloud.com/45504/
Subject: LU-14423 osd: recognize holes in osd_is_mapped()
Project: fs/lustre-release
Branch: b2_14
Current Patch Set:
Commit: b5f71a7e7528831a0a2c7d02b5bd98d68fcb1855