[LU-6544] Improve mkfs.lustre --stripe-count-hint inode size calculation for ldiskfs

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.5.3, Lustre 2.8.0
• Fix Version/s: Lustre 2.8.0
• Type: Bug
• Priority: Minor
• Reporter: Andreas Dilger
• Assignee: Emoly Liu
• Resolution: Fixed
• Labels: easy
• Severity: 3

Description

Running mkfs.lustre --stripe-count-hint=N should tune the ldiskfs inode size to allow files with N stripes to fit the lov layout xattr into the inode rather than storing it in an external xattr block (which reduces performance and significantly increases the space used per file on the MDT). However, I think that with newer xattrs such as lma may not be taken into account in the default stripe calculations.

Comments

[Emoly Liu]

Andreas, the current inode_size calculation is shown in the following:

                /* Inode size (for extended attributes).  The LOV EA size is
                 * 32 (EA hdr) + 32 (lov_mds_md) + stripes * 24 (lov_ost_data),
                 * and we want some margin above that for ACLs, other EAs... */
                if (strstr(mop->mo_mkfsopts, "-I") == NULL) {
                        if (IS_MDT(&mop->mo_ldd)) {
                                if (mop->mo_stripe_count > 72)
                                        inode_size = 512; /* bz 7241 */
                                /* see also "-i" below for EA blocks */
                                else if (mop->mo_stripe_count > 32)
                                        inode_size = 2048;
                                else if (mop->mo_stripe_count > 10)
                                        inode_size = 1024;
                                else
                                        inode_size = 512;

If I understand the new requirement right, no matter what stripe_count=N the user sets, we should fit lma and all current xattrs into the inode size. The the calculation looks like

156(ldiskfs_inode_size) + 24(lma) + 32(lov EA hdr) + 32 (lov_mds_md) + 24(lov_ost_data) * N <= 512*2^m {m=0,1,2,3}

And m can't be bigger than 3 (at this situation inode size will be 4096bytes), and the biggest value of N is 160, right?

[Andreas Dilger]

It looks like the previous limits were leaving a small amount of free space at each step (about 12 bytes), but it would be better to reduce the limits to 8, 30, 72 so there is a small amount of space reserved for other xattrs (ACL, etc) in the inode.

[Gerrit Updater]

Emoly Liu ([e-mail]) uploaded a new patch: http://review.whamcloud.com/15643
Subject: LU-6544 mkfs: Reduce mkfs.lustre --stripe-count-hint limits
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 0e2e2ede4eaba5aa66c2436e39c713aa8a95d0c5

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch http://review.whamcloud.com/15643/
Subject: LU-6544 mkfs: Improve MDT inode size calculation
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 71b3cc3315114b2101bc32707fa66740279fff54

[Peter Jones]

Landed for 2.8