[LU-15727] lod_get_default_lov_striping() misinterprets composite striping for append

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.16.0
• Severity: 3
• Assignee: John Hammond
• Reporter: John Hammond

Description

In lod_get_default_lov_striping() if the default striping is composite and dah_append_stripes is non zero then we try to read lmm_pattern from a struct lov_comp_md_v1 which isn't really there and we get confused.

To reproduce:

$LUSTRE/tets/llmount.sh 
POOL=lustre.pool0
lctl pool_new $POOL
lctl pool_add $POOL OST0000
lctl pool_add $POOL OST0001
lfs setstripe -E 1M -c 1 -p $POOL -E 2M -c 2 -p $POOL -E eof -c -1 /mnt/lustre
lctl set_param debug='+layout trace'
lctl clear echo XXX >> /mnt/lustre/f0
lctl dk | grep lod_get_default_lov_striping

Output:

00000004:00000001:1.0:1649345555.830519:0:1905344:0:(lod_object.c:5263:lod_get_default_lov_striping()) Process entered
00000004:00000001:1.0:1649345555.830540:0:1905344:0:(lod_object.c:5347:lod_get_default_lov_striping()) Process leaving (rc=18446744073709551594 : -22 : ffffffffffffffea)
00000004:00000001:1.0:1649345555.830543:0:1905344:0:(lod_object.c:5263:lod_get_default_lov_striping()) Process entered
00000004:00000001:1.0:1649345555.830553:0:1905344:0:(lod_object.c:5347:lod_get_default_lov_striping()) Process leaving rc=18446744073709551594 : -22 : ffffffffffffffea)

                if (!lov_pattern_supported(v1->lmm_pattern) &&
                    !(v1->lmm_pattern & LOV_PATTERN_F_RELEASED)) {
			lod_free_def_comp_entries(lds);
                        RETURN(-EINVAL);
                }

This is because we use composite in two different senses: lti_ea_store contains a composite layout, and we should return a composite layout.

Issue Links

• is related to: LU-16872 sanity: test_27M Error: '(5) stripe count , should be 8 for append'
• is related to: LU-16014 sanity test_27M: crash in lod_qos_prep_create()
• is related to: LU-9341 PFL: append should not instantiate full layout

Comments

[Peter Jones]

green  should we try reverting this patch to confirm?

[Oleg Drokin]

I feel like LU-16014 is a regression introduced by this patch

[Peter Jones]

Landed for 2.16

[Gerrit Updater]

"Oleg Drokin <[e-mail]>" merged in patch https://review.whamcloud.com/47014/
Subject: LU-15727 lod: honor append_pool with default composite layouts
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 03963106926883cf322e085feb8caa3ea64db1d1

[Patrick Farrell]

I don't really have an opinion here - Andreas' comment seems reasonable to me.

[Andreas Dilger]

IMHO, the appended file should still be a plain layout file on an OST using the mdd.*.append_* parameters. Using a DoM file for append could cause grief if the file grows very large, as there wouldn't be any way to limit the size. I think this is reasonable and straight forward until/if there is a fix that avoids instantiating all components of an appended file.

[John Hammond]

adilger, paf0186
When the default striping is PFL with a DoM component then what should happen for O_APPEND:

# lfs setstripe -L mdt -E 1M -E -1 -c 2 .
# echo XXXX >> f0

What should the striping be?

[Gerrit Updater]

"John L. Hammond <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/47014
Subject: LU-15727 lod: honor append_pool with default composite layouts
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 69eb62ad9432dbb15a871f9e8ee6dbf3d3aa8a48