1 <?xml version=
"1.0" encoding=
"UTF-8"?>
2 <!DOCTYPE html PUBLIC
"-//W3C//DTD XHTML 1.1//EN"
3 "http://www.w3.org/TR/xhtml11/DTD/xhtml11.dtd">
4 <html xmlns=
"http://www.w3.org/1999/xhtml" xml:
lang=
"en">
6 <meta http-equiv=
"Content-Type" content=
"application/xhtml+xml; charset=UTF-8" />
7 <meta name=
"generator" content=
"AsciiDoc 10.2.0" />
8 <title>Use of index and Racy Git problem
</title>
9 <style type=
"text/css">
10 /* Shared CSS for AsciiDoc xhtml11 and html5 backends */
14 font-family: Georgia,serif;
18 h1, h2, h3, h4, h5, h6,
19 div.title, caption.title,
20 thead, p.table.header,
22 #author, #revnumber, #revdate, #revremark,
24 font-family: Arial,Helvetica,sans-serif;
28 margin:
1em
5%
1em
5%;
33 text-decoration: underline;
49 h1, h2, h3, h4, h5, h6 {
57 border-bottom:
2px solid silver;
77 border:
1px solid silver;
88 ul
> li { color: #aaa; }
89 ul
> li
> * { color: black; }
91 .monospaced, code, pre {
92 font-family:
"Courier New", Courier, monospace;
99 white-space: pre-wrap;
109 #revnumber, #revdate, #revremark {
114 border-top:
2px solid silver;
120 padding-bottom:
0.5em;
124 padding-bottom:
0.5em;
129 margin-bottom:
1.5em;
131 div.imageblock, div.exampleblock, div.verseblock,
132 div.quoteblock, div.literalblock, div.listingblock, div.sidebarblock,
133 div.admonitionblock {
135 margin-bottom:
1.5em;
137 div.admonitionblock {
139 margin-bottom:
2.0em;
144 div.content { /* Block element content. */
148 /* Block element titles. */
149 div.title, caption.title {
154 margin-bottom:
0.5em;
160 td div.title:first-child {
163 div.content div.title:first-child {
166 div.content + div.title {
170 div.sidebarblock
> div.content {
172 border:
1px solid #dddddd;
173 border-left:
4px solid #f0f0f0;
177 div.listingblock
> div.content {
178 border:
1px solid #dddddd;
179 border-left:
5px solid #f0f0f0;
184 div.quoteblock, div.verseblock {
188 border-left:
5px solid #f0f0f0;
192 div.quoteblock
> div.attribution {
197 div.verseblock
> pre.content {
198 font-family: inherit;
201 div.verseblock
> div.attribution {
205 /* DEPRECATED: Pre version
8.2.7 verse style literal block. */
206 div.verseblock + div.attribution {
210 div.admonitionblock .icon {
214 text-decoration: underline;
216 padding-right:
0.5em;
218 div.admonitionblock td.content {
220 border-left:
3px solid #dddddd;
223 div.exampleblock
> div.content {
224 border-left:
3px solid #dddddd;
228 div.imageblock div.content { padding-left:
0; }
229 span.image img { border-style: none; vertical-align: text-bottom; }
230 a.image:visited { color: white; }
234 margin-bottom:
0.8em;
247 list-style-position: outside;
250 list-style-type: decimal;
253 list-style-type: lower-alpha;
256 list-style-type: upper-alpha;
259 list-style-type: lower-roman;
262 list-style-type: upper-roman;
265 div.compact ul, div.compact ol,
266 div.compact p, div.compact p,
267 div.compact div, div.compact div {
269 margin-bottom:
0.1em;
281 margin-bottom:
0.8em;
284 padding-bottom:
15px;
286 dt.hdlist1.strong, td.hdlist1.strong {
292 padding-right:
0.8em;
298 div.hdlist.compact tr {
307 .footnote, .footnoteref {
311 span.footnote, span.footnoteref {
312 vertical-align: super;
316 margin:
20px
0 20px
0;
320 #footnotes div.footnote {
326 border-top:
1px solid silver;
335 padding-right:
0.5em;
336 padding-bottom:
0.3em;
344 #footer-badges { display: none; }
348 margin-bottom:
2.5em;
356 margin-bottom:
0.1em;
359 div.toclevel0, div.toclevel1, div.toclevel2, div.toclevel3, div.toclevel4 {
376 span.aqua { color: aqua; }
377 span.black { color: black; }
378 span.blue { color: blue; }
379 span.fuchsia { color: fuchsia; }
380 span.gray { color: gray; }
381 span.green { color: green; }
382 span.lime { color: lime; }
383 span.maroon { color: maroon; }
384 span.navy { color: navy; }
385 span.olive { color: olive; }
386 span.purple { color: purple; }
387 span.red { color: red; }
388 span.silver { color: silver; }
389 span.teal { color: teal; }
390 span.white { color: white; }
391 span.yellow { color: yellow; }
393 span.aqua-background { background: aqua; }
394 span.black-background { background: black; }
395 span.blue-background { background: blue; }
396 span.fuchsia-background { background: fuchsia; }
397 span.gray-background { background: gray; }
398 span.green-background { background: green; }
399 span.lime-background { background: lime; }
400 span.maroon-background { background: maroon; }
401 span.navy-background { background: navy; }
402 span.olive-background { background: olive; }
403 span.purple-background { background: purple; }
404 span.red-background { background: red; }
405 span.silver-background { background: silver; }
406 span.teal-background { background: teal; }
407 span.white-background { background: white; }
408 span.yellow-background { background: yellow; }
410 span.big { font-size:
2em; }
411 span.small { font-size:
0.6em; }
413 span.underline { text-decoration: underline; }
414 span.overline { text-decoration: overline; }
415 span.line-through { text-decoration: line-through; }
417 div.unbreakable { page-break-inside: avoid; }
427 margin-bottom:
1.5em;
429 div.tableblock
> table {
430 border:
3px solid #
527bbd;
432 thead, p.table.header {
439 /* Because the table frame attribute is overridden by CSS in most browsers. */
440 div.tableblock
> table[
frame=
"void"] {
443 div.tableblock
> table[
frame=
"hsides"] {
444 border-left-style: none;
445 border-right-style: none;
447 div.tableblock
> table[
frame=
"vsides"] {
448 border-top-style: none;
449 border-bottom-style: none;
460 margin-bottom:
1.5em;
462 thead, p.tableblock.header {
473 border-color: #
527bbd;
474 border-collapse: collapse;
476 th.tableblock, td.tableblock {
480 border-color: #
527bbd;
483 table.tableblock.frame-topbot {
484 border-left-style: hidden;
485 border-right-style: hidden;
487 table.tableblock.frame-sides {
488 border-top-style: hidden;
489 border-bottom-style: hidden;
491 table.tableblock.frame-none {
492 border-style: hidden;
495 th.tableblock.halign-left, td.tableblock.halign-left {
498 th.tableblock.halign-center, td.tableblock.halign-center {
501 th.tableblock.halign-right, td.tableblock.halign-right {
505 th.tableblock.valign-top, td.tableblock.valign-top {
508 th.tableblock.valign-middle, td.tableblock.valign-middle {
509 vertical-align: middle;
511 th.tableblock.valign-bottom, td.tableblock.valign-bottom {
512 vertical-align: bottom;
523 padding-bottom:
0.5em;
524 border-top:
2px solid silver;
525 border-bottom:
2px solid silver;
530 body.manpage div.sectionbody {
535 body.manpage div#toc { display: none; }
540 <script type=
"text/javascript">
542 var asciidoc = { // Namespace.
544 /////////////////////////////////////////////////////////////////////
545 // Table Of Contents generator
546 /////////////////////////////////////////////////////////////////////
548 /* Author: Mihai Bazon, September
2002
549 * http://students.infoiasi.ro/~mishoo
551 * Table Of Content generator
554 * Feel free to use this script under the terms of the GNU General Public
555 * License, as long as you do not remove or alter this notice.
558 /* modified by Troy D. Hanson, September
2006. License: GPL */
559 /* modified by Stuart Rackham,
2006,
2009. License: GPL */
562 toc: function (toclevels) {
564 function getText(el) {
566 for (var i = el.firstChild; i != null; i = i.nextSibling) {
567 if (i.nodeType ==
3 /* Node.TEXT_NODE */) // IE doesn't speak constants.
569 else if (i.firstChild != null)
575 function TocEntry(el, text, toclevel) {
578 this.toclevel = toclevel;
581 function tocEntries(el, toclevels) {
582 var result = new Array;
583 var re = new RegExp('[hH]([
1-'+(toclevels+
1)+'])');
584 // Function that scans the DOM tree for header elements (the DOM2
585 // nodeIterator API would be a better technique but not supported by all
587 var iterate = function (el) {
588 for (var i = el.firstChild; i != null; i = i.nextSibling) {
589 if (i.nodeType ==
1 /* Node.ELEMENT_NODE */) {
590 var mo = re.exec(i.tagName);
591 if (mo && (i.getAttribute(
"class") || i.getAttribute(
"className")) !=
"float") {
592 result[result.length] = new TocEntry(i, getText(i), mo[
1]-
1);
602 var toc = document.getElementById(
"toc");
607 // Delete existing TOC entries in case we're reloading the TOC.
608 var tocEntriesToRemove = [];
610 for (i =
0; i < toc.childNodes.length; i++) {
611 var entry = toc.childNodes[i];
612 if (entry.nodeName.toLowerCase() == 'div'
613 && entry.getAttribute(
"class")
614 && entry.getAttribute(
"class").match(/^toclevel/))
615 tocEntriesToRemove.push(entry);
617 for (i =
0; i < tocEntriesToRemove.length; i++) {
618 toc.removeChild(tocEntriesToRemove[i]);
621 // Rebuild TOC entries.
622 var entries = tocEntries(document.getElementById(
"content"), toclevels);
623 for (var i =
0; i < entries.length; ++i) {
624 var entry = entries[i];
625 if (entry.element.id ==
"")
626 entry.element.id =
"_toc_" + i;
627 var a = document.createElement(
"a");
628 a.href =
"#" + entry.element.id;
629 a.appendChild(document.createTextNode(entry.text));
630 var div = document.createElement(
"div");
632 div.className =
"toclevel" + entry.toclevel;
633 toc.appendChild(div);
635 if (entries.length ==
0)
636 toc.parentNode.removeChild(toc);
640 /////////////////////////////////////////////////////////////////////
641 // Footnotes generator
642 /////////////////////////////////////////////////////////////////////
644 /* Based on footnote generation code from:
645 * http://www.brandspankingnew.net/archive/
2005/
07/format_footnote.html
648 footnotes: function () {
649 // Delete existing footnote entries in case we're reloading the footnodes.
651 var noteholder = document.getElementById(
"footnotes");
655 var entriesToRemove = [];
656 for (i =
0; i < noteholder.childNodes.length; i++) {
657 var entry = noteholder.childNodes[i];
658 if (entry.nodeName.toLowerCase() == 'div' && entry.getAttribute(
"class") ==
"footnote")
659 entriesToRemove.push(entry);
661 for (i =
0; i < entriesToRemove.length; i++) {
662 noteholder.removeChild(entriesToRemove[i]);
665 // Rebuild footnote entries.
666 var cont = document.getElementById(
"content");
667 var spans = cont.getElementsByTagName(
"span");
670 for (i=
0; i
<spans.length; i++) {
671 if (spans[i].className ==
"footnote") {
673 var note = spans[i].getAttribute(
"data-note");
675 // Use [\s\S] in place of . so multi-line matches work.
676 // Because JavaScript has no s (dotall) regex flag.
677 note = spans[i].innerHTML.match(/\s*\[([\s\S]*)]\s*/)[
1];
679 "[<a id='_footnoteref_" + n +
"' href='#_footnote_" + n +
680 "' title='View footnote' class='footnote'>" + n +
"</a>]";
681 spans[i].setAttribute(
"data-note", note);
683 noteholder.innerHTML +=
684 "<div class='footnote' id='_footnote_" + n +
"'>" +
685 "<a href='#_footnoteref_" + n +
"' title='Return to text'>" +
686 n +
"</a>. " + note +
"</div>";
687 var id =spans[i].getAttribute(
"id");
688 if (id != null) refs[
"#"+id] = n;
692 noteholder.parentNode.removeChild(noteholder);
694 // Process footnoterefs.
695 for (i=
0; i
<spans.length; i++) {
696 if (spans[i].className ==
"footnoteref") {
697 var href = spans[i].getElementsByTagName(
"a")[
0].getAttribute(
"href");
698 href = href.match(/#.*/)[
0]; // Because IE return full URL.
701 "[<a href='#_footnote_" + n +
702 "' title='View footnote' class='footnote'>" + n +
"</a>]";
708 install: function(toclevels) {
711 function reinstall() {
712 asciidoc.footnotes();
714 asciidoc.toc(toclevels);
718 function reinstallAndRemoveTimer() {
719 clearInterval(timerId);
723 timerId = setInterval(reinstall,
500);
724 if (document.addEventListener)
725 document.addEventListener(
"DOMContentLoaded", reinstallAndRemoveTimer, false);
727 window.onload = reinstallAndRemoveTimer;
735 <body class=
"article">
737 <h1>Use of index and Racy Git problem
</h1>
738 <span id=
"revdate">2024-
03-
07</span>
742 <h2 id=
"_background">Background
</h2>
743 <div class=
"sectionbody">
744 <div class=
"paragraph"><p>The index is one of the most important data structures in Git.
745 It represents a virtual working tree state by recording list of
746 paths and their object names and serves as a staging area to
747 write out the next tree object to be committed. The state is
748 "virtual" in the sense that it does not necessarily have to, and
749 often does not, match the files in the working tree.
</p></div>
750 <div class=
"paragraph"><p>There are cases where Git needs to examine the differences between the
751 virtual working tree state in the index and the files in the
752 working tree. The most obvious case is when the user asks
<code>git
753 diff
</code> (or its low level implementation,
<code>git diff-files
</code>) or
754 <code>git-ls-files --modified
</code>. In addition, Git internally checks
755 if the files in the working tree are different from what are
756 recorded in the index to avoid stomping on local changes in them
757 during patch application, switching branches, and merging.
</p></div>
758 <div class=
"paragraph"><p>In order to speed up this comparison between the files in the
759 working tree and the index entries, the index entries record the
760 information obtained from the filesystem via
<code>lstat(
2)
</code> system
761 call when they were last updated. When checking if they differ,
762 Git first runs
<code>lstat(
2)
</code> on the files and compares the result
763 with this information (this is what was originally done by the
764 <code>ce_match_stat()
</code> function, but the current code does it in
765 <code>ce_match_stat_basic()
</code> function). If some of these
"cached
766 stat information" fields do not match, Git can tell that the
767 files are modified without even looking at their contents.
</p></div>
768 <div class=
"paragraph"><p>Note: not all members in
<code>struct stat
</code> obtained via
<code>lstat(
2)
</code>
769 are used for this comparison. For example,
<code>st_atime
</code> obviously
770 is not useful. Currently, Git compares the file type (regular
771 files vs symbolic links) and executable bits (only for regular
772 files) from
<code>st_mode
</code> member,
<code>st_mtime
</code> and
<code>st_ctime
</code>
773 timestamps,
<code>st_uid
</code>,
<code>st_gid
</code>,
<code>st_ino
</code>, and
<code>st_size
</code> members.
774 With a
<code>USE_STDEV
</code> compile-time option,
<code>st_dev
</code> is also
775 compared, but this is not enabled by default because this member
776 is not stable on network filesystems. With
<code>USE_NSEC
</code>
777 compile-time option,
<code>st_mtim.tv_nsec
</code> and
<code>st_ctim.tv_nsec
</code>
778 members are also compared. On Linux, this is not enabled by default
779 because in-core timestamps can have finer granularity than
780 on-disk timestamps, resulting in meaningless changes when an
781 inode is evicted from the inode cache. See commit
8ce13b0
782 of git://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git
783 ([PATCH] Sync in core time granularity with filesystems,
784 2005-
01-
04). This patch is included in kernel
2.6.11 and newer, but
785 only fixes the issue for file systems with exactly
1 ns or
1 s
786 resolution. Other file systems are still broken in current Linux
787 kernels (e.g. CEPH, CIFS, NTFS, UDF), see
788 <a href=
"https://lore.kernel.org/lkml/5577240D.7020309@gmail.com/">https://lore.kernel.org/lkml/
5577240D
.7020309@gmail.com/
</a></p></div>
792 <h2 id=
"_racy_git">Racy Git
</h2>
793 <div class=
"sectionbody">
794 <div class=
"paragraph"><p>There is one slight problem with the optimization based on the
795 cached stat information. Consider this sequence:
</p></div>
796 <div class=
"literalblock">
797 <div class=
"content">
798 <pre><code>: modify 'foo'
799 $ git update-index 'foo'
800 : modify 'foo' again, in-place, without changing its size
</code></pre>
802 <div class=
"paragraph"><p>The first
<code>update-index
</code> computes the object name of the
803 contents of file
<code>foo
</code> and updates the index entry for
<code>foo
</code>
804 along with the
<code>struct stat
</code> information. If the modification
805 that follows it happens very fast so that the file
’s
<code>st_mtime
</code>
806 timestamp does not change, after this sequence, the cached stat
807 information the index entry records still exactly match what you
808 would see in the filesystem, even though the file
<code>foo
</code> is now
810 This way, Git can incorrectly think files in the working tree
811 are unmodified even though they actually are. This is called
812 the
"racy Git" problem (discovered by Pasky), and the entries
813 that appear clean when they may not be because of this problem
814 are called
"racily clean".
</p></div>
815 <div class=
"paragraph"><p>To avoid this problem, Git does two things:
</p></div>
816 <div class=
"olist arabic"><ol class=
"arabic">
819 When the cached stat information says the file has not been
820 modified, and the
<code>st_mtime
</code> is the same as (or newer than)
821 the timestamp of the index file itself (which is the time
<code>git
822 update-index foo
</code> finished running in the above example), it
823 also compares the contents with the object registered in the
824 index entry to make sure they match.
829 When the index file is updated that contains racily clean
830 entries, cached
<code>st_size
</code> information is truncated to zero
831 before writing a new version of the index file.
835 <div class=
"paragraph"><p>Because the index file itself is written after collecting all
836 the stat information from updated paths,
<code>st_mtime
</code> timestamp of
837 it is usually the same as or newer than any of the paths the
838 index contains. And no matter how quick the modification that
839 follows
<code>git update-index foo
</code> finishes, the resulting
840 <code>st_mtime
</code> timestamp on
<code>foo
</code> cannot get a value earlier
841 than the index file. Therefore, index entries that can be
842 racily clean are limited to the ones that have the same
843 timestamp as the index file itself.
</p></div>
844 <div class=
"paragraph"><p>The callers that want to check if an index entry matches the
845 corresponding file in the working tree continue to call
846 <code>ce_match_stat()
</code>, but with this change,
<code>ce_match_stat()
</code> uses
847 <code>ce_modified_check_fs()
</code> to see if racily clean ones are
848 actually clean after comparing the cached stat information using
849 <code>ce_match_stat_basic()
</code>.
</p></div>
850 <div class=
"paragraph"><p>The problem the latter solves is this sequence:
</p></div>
851 <div class=
"literalblock">
852 <div class=
"content">
853 <pre><code>$ git update-index 'foo'
854 : modify 'foo' in-place without changing its size
855 : wait for enough time
856 $ git update-index 'bar'
</code></pre>
858 <div class=
"paragraph"><p>Without the latter, the timestamp of the index file gets a newer
859 value, and falsely clean entry
<code>foo
</code> would not be caught by the
860 timestamp comparison check done with the former logic anymore.
861 The latter makes sure that the cached stat information for
<code>foo
</code>
862 would never match with the file in the working tree, so later
863 checks by
<code>ce_match_stat_basic()
</code> would report that the index entry
864 does not match the file and Git does not have to fall back on more
865 expensive
<code>ce_modified_check_fs()
</code>.
</p></div>
869 <h2 id=
"_runtime_penalty">Runtime penalty
</h2>
870 <div class=
"sectionbody">
871 <div class=
"paragraph"><p>The runtime penalty of falling back to
<code>ce_modified_check_fs()
</code>
872 from
<code>ce_match_stat()
</code> can be very expensive when there are many
873 racily clean entries. An obvious way to artificially create
874 this situation is to give the same timestamp to all the files in
875 the working tree in a large project, run
<code>git update-index
</code> on
876 them, and give the same timestamp to the index file:
</p></div>
877 <div class=
"literalblock">
878 <div class=
"content">
879 <pre><code>$ date
>.datestamp
880 $ git ls-files | xargs touch -r .datestamp
881 $ git ls-files | git update-index --stdin
882 $ touch -r .datestamp .git/index
</code></pre>
884 <div class=
"paragraph"><p>This will make all index entries racily clean. The linux project, for
885 example, there are over
20,
000 files in the working tree. On my
886 Athlon
64 X2
3800+, after the above:
</p></div>
887 <div class=
"literalblock">
888 <div class=
"content">
889 <pre><code>$ /usr/bin/time git diff-files
890 1.68user
0.54system
0:
02.22elapsed
100%CPU (
0avgtext+
0avgdata
0maxresident)k
891 0inputs+
0outputs (
0major+
67111minor)pagefaults
0swaps
892 $ git update-index MAINTAINERS
893 $ /usr/bin/time git diff-files
894 0.02user
0.12system
0:
00.14elapsed
100%CPU (
0avgtext+
0avgdata
0maxresident)k
895 0inputs+
0outputs (
0major+
935minor)pagefaults
0swaps
</code></pre>
897 <div class=
"paragraph"><p>Running
<code>git update-index
</code> in the middle checked the racily
898 clean entries, and left the cached
<code>st_mtime
</code> for all the paths
899 intact because they were actually clean (so this step took about
900 the same amount of time as the first
<code>git diff-files
</code>). After
901 that, they are not racily clean anymore but are truly clean, so
902 the second invocation of
<code>git diff-files
</code> fully took advantage
903 of the cached stat information.
</p></div>
907 <h2 id=
"_avoiding_runtime_penalty">Avoiding runtime penalty
</h2>
908 <div class=
"sectionbody">
909 <div class=
"paragraph"><p>In order to avoid the above runtime penalty, post
1.4.2 Git used
910 to have a code that made sure the index file
911 got a timestamp newer than the youngest files in the index when
912 there were many young files with the same timestamp as the
913 resulting index file otherwise would have by waiting
914 before finishing writing the index file out.
</p></div>
915 <div class=
"paragraph"><p>I suspected that in practice the situation where many paths in the
916 index are all racily clean was quite rare. The only code paths
917 that can record recent timestamp for large number of paths are:
</p></div>
918 <div class=
"olist arabic"><ol class=
"arabic">
921 Initial
<code>git add .
</code> of a large project.
926 <code>git checkout
</code> of a large project from an empty index into an
927 unpopulated working tree.
931 <div class=
"paragraph"><p>Note: switching branches with
<code>git checkout
</code> keeps the cached
932 stat information of existing working tree files that are the
933 same between the current branch and the new branch, which are
934 all older than the resulting index file, and they will not
935 become racily clean. Only the files that are actually checked
936 out can become racily clean.
</p></div>
937 <div class=
"paragraph"><p>In a large project where raciness avoidance cost really matters,
938 however, the initial computation of all object names in the
939 index takes more than one second, and the index file is written
940 out after all that happens. Therefore the timestamp of the
941 index file will be more than one second later than the
942 youngest file in the working tree. This means that in these
943 cases there actually will not be any racily clean entry in
944 the resulting index.
</p></div>
945 <div class=
"paragraph"><p>Based on this discussion, the current code does not use the
946 "workaround" to avoid the runtime penalty that does not exist in
947 practice anymore. This was done with commit
0fc82cff on Aug
15,
952 <div id=
"footnotes"><hr /></div>
954 <div id=
"footer-text">
956 2023-
10-
23 14:
43:
46 PDT