revlog: automatically read from opened file handles
authorGregory Szorc <gregory.szorc@gmail.com>
Tue, 13 Nov 2018 12:32:05 -0800
changeset 40627 e9293c5f8bb9
parent 40626 87a872555e90
child 40628 29e4a77b5305
revlog: automatically read from opened file handles The revlog reading code commonly opens a new file handle for reading on demand. There is support for passing a file handle to revlog.revision(). But it is marked as an internal argument. When revlogs are written, we write() data as it is available. But we don't flush() data until all revisions are written. Putting these two traits together, it is possible for an in-process revlog reader during active writes to trigger the opening of a new file handle on a file with unflushed writes. The reader won't have access to all "available" revlog data (as it hasn't been flushed). And with the introduction of the previous patch, this can lead to the revlog raising an error due to a partial read. I witnessed this behavior when applying changegroup data (via `hg pull`) before issue6006 was fixed via different means. Having this and the previous patch in play would have helped cause errors earlier rather than manifesting as hash verification failures. While this has been a long-standing issue, I believe the relatively new delta computation code has tickled it into being more common. This is because the new delta computation code will compute deltas in more scenarios. This can lead to revlog reading. While the delta computation code is probably supposed to reuse file handles, it appears it isn't doing so in all circumstances. But the issue runs deeper than that. Theoretically, any code can access revision data during revlog writes. It appears we were just getting lucky that it wasn't. (The "add revision callback" passed to addgroup() provides an avenue to do this.) If I changed the revlog's behavior to not cache the full revision text or to clear caches after revision insertion during addgroup(), I was able to produce crashes 100% of the time when writing changelog revisions. This is because changelog's add revision callback attempts to resolve the revision data to access the changed files list. And without the revision's fulltext being cached, we performed a revlog read, which required opening a new file handle. This attempted to read unflushed data, leading to a partial read and a crash. This commit teaches the revlog to store the file handles used for writing multiple revisions during addgroup(). It also teaches the code for resolving a file handle when reading to use these handles, if available. This ensures that *any* reads (regardless of their source) use the active writing file handles, if available. These file handles have access to the unflushed data because they wrote it. This allows reads to complete without issue. Differential Revision: https://phab.mercurial-scm.org/D5267
mercurial/revlog.py
--- a/mercurial/revlog.py	Tue Nov 13 12:30:59 2018 -0800
+++ b/mercurial/revlog.py	Tue Nov 13 12:32:05 2018 -0800
@@ -375,6 +375,9 @@
         # custom flags.
         self._flagprocessors = dict(_flagprocessors)
 
+        # 2-tuple of file handles being used for active writing.
+        self._writinghandles = None
+
         mmapindexthreshold = None
         v = REVLOG_DEFAULT_VERSION
         opts = getattr(opener, 'options', None)
@@ -505,8 +508,21 @@
     @contextlib.contextmanager
     def _datareadfp(self, existingfp=None):
         """file object suitable to read data"""
+        # Use explicit file handle, if given.
         if existingfp is not None:
             yield existingfp
+
+        # Use a file handle being actively used for writes, if available.
+        # There is some danger to doing this because reads will seek the
+        # file. However, _writeentry() performs a SEEK_END before all writes,
+        # so we should be safe.
+        elif self._writinghandles:
+            if self._inline:
+                yield self._writinghandles[0]
+            else:
+                yield self._writinghandles[1]
+
+        # Otherwise open a new file handle.
         else:
             if self._inline:
                 func = self._indexfp
@@ -1750,6 +1766,9 @@
         if fp:
             fp.flush()
             fp.close()
+            # We can't use the cached file handle after close(). So prevent
+            # its usage.
+            self._writinghandles = None
 
         with self._indexfp('r') as ifh, self._datafp('w') as dfh:
             for r in self:
@@ -1996,7 +2015,9 @@
         # if the file was seeked to before the end. See issue4943 for more.
         #
         # We work around this issue by inserting a seek() before writing.
-        # Note: This is likely not necessary on Python 3.
+        # Note: This is likely not necessary on Python 3. However, because
+        # the file handle is reused for reads and may be seeked there, we need
+        # to be careful before changing this.
         ifh.seek(0, os.SEEK_END)
         if dfh:
             dfh.seek(0, os.SEEK_END)
@@ -2029,6 +2050,9 @@
         this revlog and the node that was added.
         """
 
+        if self._writinghandles:
+            raise error.ProgrammingError('cannot nest addgroup() calls')
+
         nodes = []
 
         r = len(self)
@@ -2048,6 +2072,9 @@
             if dfh:
                 dfh.flush()
             ifh.flush()
+
+        self._writinghandles = (ifh, dfh)
+
         try:
             deltacomputer = deltautil.deltacomputer(self)
             # loop through our set of deltas
@@ -2109,7 +2136,10 @@
                     ifh.close()
                     dfh = self._datafp("a+")
                     ifh = self._indexfp("a+")
+                    self._writinghandles = (ifh, dfh)
         finally:
+            self._writinghandles = None
+
             if dfh:
                 dfh.close()
             ifh.close()