[LU-15837] "lfs find -printf" improvements

• Status: Open
• Project: Lustre
• Affects Version/s: Lustre 2.15.0
• Fix Version/s: Lustre 2.16.0
• Type: Improvement
• Priority: Major
• Reporter: Andreas Dilger
• Assignee: WC Triage
• Resolution: Unresolved
• Labels: easy, lad23dd, lug23dd

Description

Per Rick's LUG presentation, there are still a number of optimizations for "lfs find -printf" that were not included in the original patch https://review.whamcloud.com/45136 "LU-10378 utils: add formatted printf to lfs find" landing that should still be addressed:

• selective fetching of metadata attributes. In particular, fetching the projid for a file requires an extra syscall that is unnecessary if the projid is not being printed, at least until LU-12480 is implemented. Most of the other MDT attributes are "free" once any attribute is read for a file, so it probably doesn't make sense to micro-optimize here.
• pre-parsing of the -printf argument string (assuming this actually shows up in the CPU profile, I'm not sure if it is actually important vs. stat/RPC overhead)
• maybe others, I didn't catch all of them

Issue Links

• is related to: LU-5170 lfs usability
• is related to: LU-12480 add STATX_PROJID to upstream kernel
• is related to: LU-15504 "lfs find" is missing "-ls" support
• is related to: LU-15743 "lfs find" is missing "-xattr" support
• is related to: LU-7495 lfs find is missing "-links" support
• is related to: LU-10378 "lfs find" is missing "-printf" support
• is related to: LU-16798 lfs find: new --jobid option
• is related to: LU-16560 'lfs find -printf %w' does not print ...
• is related to: LU-16808 lfs find --printf fails on FIFOs and ...
• is related to: LU-17219 lfs find: add ability to print extend...
• is related to: LU-16760 "lfs find" support for fscrypt and ot...
• is related to: LU-16798 lfs find: new --jobid option

Comments

[Andreas Dilger]

It looks like a few additional -printf options should be implemented, but nothing overly complex. These would be useful by themselves, and also to implement "lfs find -ls":

• "%i" - inode number
• "%LF" - Lustre FID
• "%k" - file blocks in kilobytes (should be trivial, just "%b/2")
• "%M" - symbolic file access mode (decode file type (-bcdls+/-rwxSt UGO permission bits)
• "%n" - link count (LU-7495)
• "%u" - username (could be mapped to "%U" initially to print numeric UID, and implemented separately)
• "%g" - groupname (use "%g" initially, as with "%u")

[Andreas Dilger]

Looking at the strace output for regular file processing it seems like it is doing far too much for most files:

# strace -ttt -T lfs find /mnt/testfs/dir
:
:
1683657120.708615 getdents(4, /* 102 entries */, 32768) = 6448 <0.000784>
1683657120.709513 ioctl(4, _IOC(_IOC_READ|_IOC_WRITE, 0x69, 0x16, 0x140), 0xea2040) = 0 <0.002730>
1683657120.712309 open("/mnt/testfs/dir/filename_is_long-15", O_RDONLY) = 3 <0.003517>
1683657120.715904 ioctl(3, _IOC(_IOC_READ, 0x66, 0xaf, 0x4), 0x7ffc0a9b3ba0) = 0 <0.000140>
1683657120.716106 ioctl(3, FS_IOC_FSGETXATTR, 0x7ffc0a9b37d0) = 0 <0.000044>
1683657120.716243 newfstatat(4, "filename_is_long-15", {st_mode=S_IFREG|0644, st_size=0, ...}, AT_SYMLINK_NOFOLLOW) = 0 <0.000683>
1683657120.717022 write(1, "/mnt/testfs/dir/filename_is_long-15", 36) = 36 <0.000296> /mnt/testfs/dir/filename_is_long-15
1683657120.717386 close(3) = 0 <0.002073>

The ioctl(4, _IOC(_IOC_READ|_IOC_WRITE, 0x69, 0x16, 0x140) = IOC_MDC_GETFILEINFO_V2) is expected for every file, since it gets the file layout and MDT attributes from the parent directory but we definitely should not be doing open() and ioctl(FS_IOC_FSGETXATTR) (to get projid) and newfstatat() (not sure why?) on every file when we are only printing the pathname. That is a lot of overhead (more than double the original ioctl(IOC_MDC_GETFILEINFO_V2) call).

Running without -printf "%p" shows none of that overhead:

1683657295.265889 getdents(4, /* 102 entries */, 32768) = 6448 <0.000393>
1683657295.266362 write(1, "/mnt/testfs/dir/filename_is_long-15"..., 36) = 36 <0.000009> /mnt/testfs/dir/filename_is_long-15
1683657295.266446 write(1, "/mnt/testfs/dir/filename_is_long-44"..., 36) = 36 <0.000009> /mnt/testfs/dir/filename_is_long-44

[Andreas Dilger]

I'm thinking that "gather_all" should be changed to be a STATX_ bitmask to indicate which fields are needed, and then the code could be changed to check the bitmask when the attributes are being fetched:

#ifndef STATX_PROJID
#define STATX_PROJID 0x40000000 /* only used internally */
#endif

        if (param->fp_check_projid || gather_printf & STATX_PROJID)

The definition of STAX_PROJID would only be used internally until the kernel could itself return projid as part of statx() (LU-12480).