Hi there,
I think I found a few nasty problems with racy detection, as well as
performance issues when using git implementations with different file
time resolutions on the same repository (e.g. git compiled with and
without USE_NSEC, libgit2 compiled with and without USE_NSEC, JGit
executed in different Java implementations...).
Let me start by listing relevant file time gotchas (skip this if it
sounds too familiar) before diving into problem descriptions. Some
ideas for potential solutions are at the end.
Notable file time facts:
========================
The st_ctime discrepancy:
* stat.st_ctime means "change time" (of file metadata) on POSIX
systems and "creation time" on Windows
* While some file systems may track all four time stamps (mtime,
atime, change time and creation time), there are no public OS APIs
to obtain creation time on POSIX / change time on Windows.
Linux:
* In-core file times may not be properly rounded to on-disk
precision, causing spurious file time changes when the cache is
refreshed from disk. This was fixed for typical Unix file systems
in kernel 2.6.11. The fix for CEPH, CIFS, NTFS, UFS and FUSE will
be in kernel 4.3. There's no fix for FAT-based file systems yet.
* Maximum file time precision is 1 ns (or 1 s with really old glibc).
Windows:
* Maximum file time precision is 100 ns.
Java <= 6:
* Only exposes mtime in milliseconds (via File.getLastModifiedTime).
Java >= 7:
* Only exposes mtime, atime and creation time, no change time (see
java.nio.file.attribute.BasicFileAttributes).
* Maximum file time precision is implementation specific (OpenJDK:
1 microsecond on both Unix [1] and Windows [2]).
* On platforms or file systems that don't support creation time,
BasicFileAttribtes.creationTime() is implementation specific
(OpenJDK returns mtime instead). There's no public API to detect
whether creation time is supported or "emulated" in some way.
Git Options:
* NO_NSEC (git only): compile-time option that disables recording of
nanoseconds in the index, implies USE_NSEC=false.
* USE_NSEC (git and libgit2 with [3]): compile-time option that
enables nanosecond comparison in both up-to-date and racy checks.
* core.checkStat=minimal (git, libgit2, JGit): config-option that
disables nanosecond comparison in up-to-date checks, but not in
racy checks.
JGit:
* Only uses mtime, rounded to milliseconds. While there is a
DirCacheEntry.setCreationTime() [4] to set the index entry's ctime
field, AFAICT its not used anywhere.
* Does not compare nanoseconds if the cached value recorded in the
index is 0, to prevent performance issues with NO_NSEC git
implementations [5].
Problem 1: Failure to detect racy files (without USE_NSEC)
==========================================================
Git may not detect racy changes when 'update-index' runs in parallel
to work tree updates.
Consider this (where timestamps are t<seconds>.<nanoseconds>):
t0.0$ echo "foo" > file1
t0.1$ git update-index file1 & # runs in background
t0.2$ # update-index records stats and sha1 of file1 in new index
t0.3$ echo "bar" > file1
....$ # update-index writes other index entries
t1.0$ # update-index finishes (sets mtime of the new index to t1.0!)
t1.1$ git status # doesn't detect that file1 has changed
The problem here is that racy checks in 'git status' compare against
the new index file's mtime (t1.0), which may be newer than the last
change of file1.
Problem 2: Failure to detect racy files (mixed USE_NSEC)
========================================================
Git may fail to detect racy conditions if file times in .git/index
have been recorded by another git implementation with better file
time resolution.
Consider the following sequence:
t0.0$ echo "foo" > file1
t0.1$ use-nsec-git update-index file1
t0.2$ echo "bar" > file1
....$ sleep 1
t1.0$ touch file2
t1.1$ use-nsec-git status # rewrites index, to store file2 change
t1.2$ git status # doesn't detect that file1 has changed
The problem here is that the first, nsec-enabled 'git status' does
not consider file1 racy (with nanosecond precision, the file is dirty
already (t0.0 != t0.2), so no racy-checks are performed). Thus, it
will not squash the size field (as a second-precision-git would).
However, it will rewrite the index to capture the status change of
file2, and thus create a new index file with mtime = t1.1. Similar
to problem 1, subsequent 'git status' with second-precision has no
way to detect that file1 has changed.
This problem would not be limited to USE_NSEC-enabled/disabled git,
it occurs whenever different file time resolutions are at play, e.g.:
* second-based git vs. millisecond-based JGit
* millisecond-based JGit vs. nanosecond-enabled git
* GIT_WORK_TREE on ext2 (1 s) and GIT_DIR on ext4 (1 ns)
* JGit executed by different Java implementations (with different
file time resolutions)
Problem 3: Failure to detect racy files with core.checkStat=minimal
===================================================================
Consider the example above (problem 2). With core.checkStat=minimal,
the nanosecond-enabled git also fails to detect that file1 has
changed.
This is because racy checks are still done with nanosecond precision
(despite checkStat=minimal), and against the *cached* mtime, not the
real one. I.e.:
* in match_stat_data(), nanoseconds are ignored, and file1 is
considered unchanged (as t0[.0] == t0[.2]).
* in ie_match_stat(), we pass the cache entry to is_racy_timestamp()
(which has mtime == t0.0), even though we know the current mtime
at this point (t0.2)
* in is_racy_stat(), file1 is not considered racy, because the index
file's mtime (t0.1) is newer than the cached mtime (t0.0)
Problem 4: Performance issues with mixed file time resolutions
==============================================================
A git implementation will consider files dirty (i.e. triggering a
content check) if the index entry has been recorded by another git
implementation with lower file time resolution.
Examples:
Git compiled with NO_NSEC writes index entries with nanosecond
fields == 0. A USE_NSEC-enabled git will consider these files dirty
(except in the rare case that on-disk nanoseconds of the file time
are really 0).
JGit writes index entries with mtime nanosecond fields rounded to
milliseconds. Again, a USE_NSEC-enabled git will consider the files
dirty.
JGit writes index entries with ctime seconds and nanoseconds == 0.
All other git implementations will consider such files dirty.
Ideas for potential solutions:
==============================
Performance issues:
-------------------
1. Compare file times in minimum supported precision
When comparing file times, use the minimum precision supported by
both the writing and reading git implementations.
1a. Simplest variant: Don't compare nanoseconds if the field in the
cached index entry is 0. JGit already does this [5], but at the
same time it is very unfriendly to USE_NSEC-enabled git by storing
only milliseconds in the nanosecond field. This "simple" solution
implies that git implementations that cannot provide full
nanosecond precision must leave the nanosecond field empty.
1b. More involved: Store the precision in the index entry.
We only need 30 bits to encode nanoseconds, so the high 2 bits of
the nanosecond field could be used as follows:
00: second precision (i.e. ignore, for backward compatibility)
01: millisecond precision
10: microsecond precision
11: nanosecond precision
When reading the index, USE-NSEC-enabled git implementations would
do dirty checks with the minimum precision supported by themselves
and the creator of the index entry.
2. Don't use ctime in dirty checks if ctime.sec == 0.
Racy detection:
---------------
3. Minimal racy solution
* Do all racy checks with second-precision only.
* When committing an index.lock file, reset mtime to the time
before git started reading the old index (i.e. time(null) when
calling read_cache()).
I believe this should fix all three racy problems described above,
although restraining ourselves to second-precision somewhat
thwarts the ability to track nanoseconds in the first place.
The problem with this solution is that files changed by git itself
will appear racy to the next git process, thus increasing the
performance penalty after e.g. a large checkout. Although I think
that re-reading the file after the file's mtime is the only way to
be really sure it hasn't been changed.
4. More ideas to solve the racy problem
Conceptually, any changes that happen at the same time or after we
start capturing information about a file may be missed by the
recording process. Thus, a "safe" way to use file times for racy /
dirty checks would be as follows:
start_capture = filesystem(file).now()
oid = read_sha1(file)
mtime1 = lstat(file).mtime
racy = mtime1 >= start_capture
...
mtime2 = lstat(file).mtime
check_content = mtime1 != mtime2 || racy
Whereas Git currently does something like this:
mtime1 = lstat(file).mtime
oid = read_sha1(file)
...
end_capture = lstat(index).mtime
mtime2 = lstat(file).mtime
check_content = mtime1 != mtime2 || mtime1 >= end_capture
One problem with this is that end_capture is only known after
closing the index file, which is why currently, racy checks can
only be done by the next git process that reads the index.
Additionally, rewriting the index file changes its mtime and thus
deprives subsequent git processes from doing racy checks. This is
currently solved by squashing the size field of racy entries.
Which means that a third git process needs to fill the size back
in, rewriting the index again...
I suspect that we could get away with fewer index rewrites if we
did racy checks in the git process that initially updates the
index entry. I.e.:
* get start_capture from index.lock immediately after creating it
(this ignores that index.lock may be on another file system
with different file time precision than the work tree)
* do racy checks immediately and store the results in the entry
* to accommodate different file time precisions, the racy "flag"
could indicate at which file time precision the entry would
have to be considered racy. E.g.
if (mtime1.sec < start_capture.sec)
return NOT_RACY;
else if (mtime1.sec > start_capture.sec ||
mtime1.nsec >= start_capture.nsec)
return ALWAYS_RACY;
else if (mtime1.nsec / 1000 == start_capture.nsec / 1000)
return RACY_AT_USEC_MSEC_SEC;
else if (mitme1.nsec / 1000000 == start_capture.nsec / 1000000)
return RACY_AT_MSEC_SEC;
else
return RACY_AT_SEC;
* for backward compatibility, we could still squash the size and
store the original size + racy info in an index extension - on
the other hand, reliable change detection is so fundamental to
an SCM that we may want to keep racy info in the core index
entry structure, probably even if it means a format change
An advantage of this would be that when rewriting the index, git
is no longer required to treat racy entries in any special way -
if the git command is not interested in the racy entry, it can
simply copy it to the next index file, without checking file
content or squashing the size field. Commands like git status
would only need to rewrite the index if the racy info changes
(i.e. enough time has passed).
Please let me know what you think of this...maybe I've completely
screwed up and can no longer see the forest for all the trees.
TIA,
Karsten
--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html