In 300 runs of the test, four at a time with the box short of memory, two runs that were running side by side measured 34 s of CPU time for -n on the upper-case tree (0.3 s on the lower-case one), within 0.1 s of each other; the other 298 measured at most 0.6 s, as did 100 more runs of that check alone. CPU time does not grow with a busy CPU, but it can when the kernel has to reclaim memory or reread directorie... closed check-in: 335dc03384 user: grok tags: fsck-perf

Brings in the hard-link index of umsdosfs, the hlink_format_test fix and tests/clockshim.c.

• doc/umsfuse.texi: the only conflict. Both sides added text after the directory cache section: this branch a paragraph saying umsfsck uses the same cache, trunk the Hard-link index section. Kept both, the paragraph first, so it stays in the directory cache section.

• src/ums_dcache.c, src/umsdos.c, src/umsdos.h and tests/run_tests.sh merged without conflicts.

The comparison of each run with the cache off against the same run with the cache on, at 4K and without the stat check masked only the ctime of an EMD record from the start of the runs on. A repair such as -y -s also stamps a record's atime (and some its mtime) with the time it ran, so two runs on either side of a second boundary gave records differing in one byte, ... check-in: d8e1603bcf user: grok tags: fsck-perf

For every host name that is not all lower case (a FAT listing shows them upper case), the copy asked ums_host_alias whether the lower-case spelling was an alias, and ums_host_lookup read the whole host directory to answer; at the end of each directory, every record's host name was looked up the same way, to remember lookup errors for the check. The directory ... check-in: 2e054a3fc2 user: grok tags: fsck-perf

• Info manual, umsfsck: each directory read once and looked up in memory; the 128 MiB default; -n uses the cache only on its scratch copy, which remains the largest part of a dry run; what memory grows with; UMSFSCK_CACHE, UMSFSCK_CACHE_NOCHECK and UMSFSCK_CACHE_STATS.
• Info manual, Directory cache: umsfsck uses the same cache.
• umsfsck.8: UMSFSCK_CACHE under ENVIRONMENT.

UMSFSCK_CACHE_STATS=FILE appends the directory cache's counters to FILE at the end of the run, in the format umsdosfs uses for UMSDOSFS_CACHE_STATS.

New tests/fsck_cache_test.sh, run from run_tests.sh:

• a tree made through the driver (long and short names, gaps, hard links in one and in several directories with three names, a symlink, a fifo, a read-only file), clean and with three sets of host-side damage (untracked and missing files...

find_hid_by_host looked up the host name of every hidden file for each pointer (pointers times hidden files directory opens). The index is made once and again only after a change to the host, counted through ums_mod_before. The pointer, hidden file and collapse candidate lists kept three 4 KiB arrays per entry (some 300 MB for 20,000 smal... check-in: ed3ed6651f user: grok tags: fsck-perf

Without the cache every ums_host_lookup reads the whole host directory, so checking a directory is quadratic in its size and every hard-link resolution rereads each directory on its path. umsfsck now turns the cache on (128 MiB) for the check, -o and the scratch copy of -n (dropped after the copy, whose directories are written past it). The one direct control-file write, -o's compaction, tells the cache through the ne... check-in: 7386c7b262 user: grok tags: fsck-perf

"archives differ by format" (about 6% of runs on an idle machine, 4% pinned to one CPU): the three trees archived by umstar (new format, old format, old format with a link root) are made one after the other by separate mounts, with the clock's times. umstar records each record's mtime (ustar header) and its atime and ctime (pax records) in whole seconds, so when the trees were made on either s... closed check-in: 8ea66b04fc user: grok tags: hlink-format-flake

• umsls_columns_test.sh: onterm now takes the pseudo-terminal width, sets it with stty cols N rows 24 before running the command,...

The reported umscache=0 regression on a flat directory of 20,000 files (ls -l 594 s cold, RSS 68 MiB, against trunk's 173 s and 20 MiB) does not come from the cache code: with umscache=0 every library entry point returns to trunk's code before touching the cache, and a gprof profile of ls -l over 8,000 files shows the same functions with the same call counts as trunk (128 million ums_parse/ums... closed check-in: 0a85145f93 user: grok tags: dircache

• Format: the last of -l, -n, -x, -1 and --format wins, except that -1 after -l stays long (as ls). --format takes vertical, across, horizontal, commas, single-column, long and verbose, or an unambiguous abbreviation. -C keeps its meaning in every UMSFUSE...

"-n changed the tree": only directory access times differed. umsfsck -n did not move them (it reads the tree with O_NOATIME, as trunk does; neither moves any access time even on a strictatime tmpfs). The test itself did: while -n was held, it listed the tree with find, between the two snapshots. On relatime an access time not newer than the change time is updated, and tar leaves the two equal, s... closed check-in: 7e6deded6f user: grok tags: fsck-scratch

Design (src/ums_dcache.c, src/ums_dcache.h):

• One entry per host directory, keyed by (st_dev, st_ino) and held open with...

Scratch copy (-n): every host name, type, mode, mtime and atime is copied as before, and now the owner too where the copy may set it (root, or a group of the caller's own), but data only for:

• control files, whole;
• host files whose record is a hard-link pointer or a symlink: the first 4096 bytes (pointers are read up t...

umsls sorts names with strcoll, as ls does, and the expected listing was written in C (byte) order. Under en_US.UTF-8 the collation ignores the punctuation and orders lower case before upper case, so the same six entries came out as --linuX-.---, --lInux-.---, --Linux-.---, --LiNuX-.---, --LINUX-.---, --LINUX-.--x and the check failed. umsls is unchanged: sorting by locale is what ls does. ... closed check-in: f3008d4950 user: grok tags: linux-emd-name-case

• src/ums_compat.c: ums_name_allowed compares with strc...

• Name fields and mangling: a name cannot contain DEL and cannot be exactly --lin...

The new format is the one used by the UMSDOS driver's internal version 0.85g, which first appears in Linux 2.2.15 and 2.4.10. That dating matches the Linux 2.2 and 2.4 sources, and has been checked on Slackware systems running kernels from before and after the change. References to 2.2.14 for this change now say 2.2.15. The earlier note that the sources dated the format to 2.2.11 and 2.3.2, and the "Unverified" remark on a root-le... check-in: 2a923cd1a2 user: grok tags: umsdos-spec

doc/mkversion already extracts those from the manifest for the Info manual. It now also writes doc/version.man, which supplies the date and the UMSFUSE version used in each page's .TH line, in the same form GNU coreutils uses in the footer. The time and check-in hash are not included.

• The .TH date and the "UMSFUSE" version footer come from version.man, not a hard-coded date.

• make install concatenates versio...

The pages are terse references now that the Info manual holds the full documentation, so they no longer need a generator. Each doc/*.adoc is replaced by an equivalent page written with the man macros (umscat.1, umscp.1, umsdosfs.1, umsls.1, umsstat.1, umstar.1, umsfsck.8), with the same sections and wording.

• doc/Makefile: the pages are sources; the a2x/asciidoctor rules are gone, and "make clean" no longer removes the pages. ...

Replace the TODO in "UMSDOS specification" and fill in its sections: directory layout, the 64-byte slot and record format with an annotated dump from the sample tree, the name mangling algorithm with worked examples, the metadata fields and host permissions, timestamps, hard-link records and the two pointer formats, special files, the PowerPC layout and its detection, and the invariants umsfsck checks.

Behaviour of the Linux 2.x driver... check-in: 492d77dcc2 user: grok tags: umsdos-spec

With write bits passed through and access(2) honoring the shown UMSDOS mode, some files the tests delete are write-protected (the sample tree's exec.exe and named-pipe, control-chars, link2.txt and link3.txt, and the rw_test pointer hard3). GNU rm asks before removing those when stdin is a terminal, so `make test` waited for an answer.

• fsck_test.sh, fsck_dryrun_test.sh, fsck_undo_test.sh: setup uses `rm -f` and `rm -rf`.

• hlink_co...

The view already reported the record's mode in getattr, but u_access only resolved the path and returned 0, so every access(X_OK) succeeded: find -type f -executable listed every file, test -x passed and bash's command completion offered any file (running one was already refused by the kernel from the mode). u_access now checks the mode it sh... closed check-in: cbb84ec91e user: grok tags: host-write-bits