[Bug 218622] libc/gen/telldir [hack-n-PATCH] performance limited to O(n) vs file count, O(n^2) against samba ls workload

bugzilla-noreply at freebsd.org bugzilla-noreply at freebsd.org
Wed Jan 17 21:54:35 UTC 2018


Alan Somers <asomers at FreeBSD.org> changed:

           What    |Removed                     |Added
                 CC|                            |asomers at FreeBSD.org

--- Comment #7 from Alan Somers <asomers at FreeBSD.org> ---
FWIW I have further improved upon mav's solution with r326640.  Now, telldir is
O(1) in both time and space on 64-bit systems.

r326640 | asomers | 2017-12-06 15:06:48 -0700 (Wed, 06 Dec 2017) | 22 lines

Optimize telldir(3)

Currently each call to telldir() requires a malloc and adds an entry to a
linked list which must be traversed on future telldir(), seekdir(),
closedir(), and readdir() calls. Applications that call telldir() for every
directory entry incur O(n^2) behavior in readdir() and O(n) in telldir() and

This optimization eliminates the malloc() and linked list in most cases by
packing the relevant information into a single long. On 64-bit architectures
msdosfs, NFS, tmpfs, UFS, and ZFS can all use the packed representation.  On
32-bit architectures msdosfs, NFS, and UFS can use the packed
representation, but ZFS and tmpfs can only use it for about the first 128
files per directory.  Memory savings is about 50 bytes per telldir(3) call.
Speedup for telldir()-heavy directory traversals is about 20-30x for one
million files per directory.

Reviewed by:    kib, mav, mckusick
MFC after:      3 weeks
Sponsored by:   Spectra Logic Corp
Differential Revision:  https://reviews.freebsd.org/D13385

You are receiving this mail because:
You are the assignee for the bug.

More information about the freebsd-standards mailing list