revlog: move revision verification out of verify

File revision verification is performing low-level checks of file
storage, namely that flags are appropriate and revision data can
be resolved.

Since these checks are somewhat revlog-specific and may not
be appropriate for alternate storage backends, this commit moves
those checks from verify.py to revlog.py.

Because we're now emitting warnings/errors that apply to specific
revisions, we taught the iverifyproblem interface to expose the
problematic node and to report this node in verify output. This
was necessary to prevent unwanted test changes.

After this change, revlog.verifyintegrity() and file verify code
in verify.py both iterate over revisions and resolve their fulltext.
But they do so in separate loops. (verify.py needs to resolve
fulltexts as part of calling renamed() - at least when using revlogs.)
This should add overhead.

But on the mozilla-unified repo:

$ hg verify
before: time: real 700.640 secs (user 585.520+0.000 sys 23.480+0.000)
after:  time: real 682.380 secs (user 570.370+0.000 sys 22.240+0.000)

I'm not sure what's going on. Maybe avoiding the filelog attribute
proxies shaved off enough time to offset the losses? Maybe fulltext
resolution has less overhead than I thought?

I've left a comment indicating the potential for optimization. But
because it doesn't produce a performance regression on a large
repository, I'm not going to worry about it.

Differential Revision: https://phab.mercurial-scm.org/D4745

mercurial/filelog.py

@@ -53,7 +53,7 b' class filelog(object):'
     def linkrev(self, rev):
         return self._revlog.linkrev(rev)
 
-    # Used by verify.
+    # Unused.
     def flags(self, rev):
         return self._revlog.flags(rev)
 
@@ -77,7 +77,7 b' class filelog(object):'
     def iscensored(self, rev):
         return self._revlog.iscensored(rev)
 
-    # Used by repo verify.
+    # Unused.
     def rawsize(self, rev):
         return self._revlog.rawsize(rev)
 

mercurial/repository.py

@@ -341,6 +341,12 b' class iverifyproblem(interfaceutil.Inter'
     error = interfaceutil.Attribute(
         """Message indicating a fatal problem.""")
 
+    node = interfaceutil.Attribute(
+        """Revision encountering the problem.
+
+        ``None`` means the problem doesn't apply to a single revision.
+        """)
+
 class irevisiondelta(interfaceutil.Interface):
     """Represents a delta between one revision and another.
 
@@ -790,6 +796,10 b' class ifilestorage(ifileindex, ifiledata'
         used to communicate data between invocations of multiple storage
         primitives.
 
+        If individual revisions cannot have their revision content resolved,
+        the method is expected to set the ``skipread`` key to a set of nodes
+        that encountered problems.
+
         The method yields objects conforming to the ``iverifyproblem``
         interface.
         """

mercurial/revlog.py

@@ -29,6 +29,7 b' from .node import ('
     nullhex,
     nullid,
     nullrev,
+    short,
     wdirfilenodeids,
     wdirhex,
     wdirid,
@@ -260,6 +261,7 b' class revlogrevisiondelta(object):'
 class revlogproblem(object):
     warning = attr.ib(default=None)
     error = attr.ib(default=None)
+    node = attr.ib(default=None)
 
 # index v0:
 #  4 bytes: offset
@@ -2644,6 +2646,89 b' class revlog(object):'
                 warning=_("warning: '%s' uses revlog format %d; expected %d") %
                         (self.indexfile, version, state['expectedversion']))
 
+        state['skipread'] = set()
+
+        for rev in self:
+            node = self.node(rev)
+
+            # Verify contents. 4 cases to care about:
+            #
+            #   common: the most common case
+            #   rename: with a rename
+            #   meta: file content starts with b'\1\n', the metadata
+            #         header defined in filelog.py, but without a rename
+            #   ext: content stored externally
+            #
+            # More formally, their differences are shown below:
+            #
+            #                       | common | rename | meta  | ext
+            #  -------------------------------------------------------
+            #   flags()             | 0      | 0      | 0     | not 0
+            #   renamed()           | False  | True   | False | ?
+            #   rawtext[0:2]=='\1\n'| False  | True   | True  | ?
+            #
+            # "rawtext" means the raw text stored in revlog data, which
+            # could be retrieved by "revision(rev, raw=True)". "text"
+            # mentioned below is "revision(rev, raw=False)".
+            #
+            # There are 3 different lengths stored physically:
+            #  1. L1: rawsize, stored in revlog index
+            #  2. L2: len(rawtext), stored in revlog data
+            #  3. L3: len(text), stored in revlog data if flags==0, or
+            #     possibly somewhere else if flags!=0
+            #
+            # L1 should be equal to L2. L3 could be different from them.
+            # "text" may or may not affect commit hash depending on flag
+            # processors (see revlog.addflagprocessor).
+            #
+            #              | common  | rename | meta  | ext
+            # -------------------------------------------------
+            #    rawsize() | L1      | L1     | L1    | L1
+            #       size() | L1      | L2-LM  | L1(*) | L1 (?)
+            # len(rawtext) | L2      | L2     | L2    | L2
+            #    len(text) | L2      | L2     | L2    | L3
+            #  len(read()) | L2      | L2-LM  | L2-LM | L3 (?)
+            #
+            # LM:  length of metadata, depending on rawtext
+            # (*): not ideal, see comment in filelog.size
+            # (?): could be "- len(meta)" if the resolved content has
+            #      rename metadata
+            #
+            # Checks needed to be done:
+            #  1. length check: L1 == L2, in all cases.
+            #  2. hash check: depending on flag processor, we may need to
+            #     use either "text" (external), or "rawtext" (in revlog).
+
+            try:
+                skipflags = state.get('skipflags', 0)
+                if skipflags:
+                    skipflags &= self.flags(rev)
+
+                if skipflags:
+                    state['skipread'].add(node)
+                else:
+                    # Side-effect: read content and verify hash.
+                    self.revision(node)
+
+                l1 = self.rawsize(rev)
+                l2 = len(self.revision(node, raw=True))
+
+                if l1 != l2:
+                    yield revlogproblem(
+                        error=_('unpacked size is %d, %d expected') % (l2, l1),
+                        node=node)
+
+            except error.CensoredNodeError:
+                if state['erroroncensored']:
+                    yield revlogproblem(error=_('censored file data'),
+                                        node=node)
+                    state['skipread'].add(node)
+            except Exception as e:
+                yield revlogproblem(
+                    error=_('unpacking %s: %s') % (short(node), e),
+                    node=node)
+                state['skipread'].add(node)
+
     def storageinfo(self, exclusivefiles=False, sharedfiles=False,
                     revisionscount=False, trackedsize=False,
                     storedsize=False):

mercurial/verify.py

@@ -343,7 +343,10 b' class verifier(object):'
 
         state = {
             # TODO this assumes revlog storage for changelog.
-            'expectedversion': self.repo.changelog.version & 0xFFFF
+            'expectedversion': self.repo.changelog.version & 0xFFFF,
+            'skipflags': self.skipflags,
+            # experimental config: censor.policy
+            'erroroncensored': ui.config('censor', 'policy') == 'abort',
         }
 
         files = sorted(set(filenodes) | set(filelinkrevs))
@@ -381,18 +384,25 b' class verifier(object):'
             if not len(fl) and (self.havecl or self.havemf):
                 self.err(lr, _("empty or missing %s") % f)
             else:
+                # Guard against implementations not setting this.
+                state['skipread'] = set()
                 for problem in fl.verifyintegrity(state):
+                    if problem.node is not None:
+                        linkrev = fl.linkrev(fl.rev(problem.node))
+                    else:
+                        linkrev = None
+
                     if problem.warning:
                         self.warn(problem.warning)
                     elif problem.error:
-                        self.err(lr, problem.error, f)
+                        self.err(linkrev if linkrev is not None else lr,
+                                 problem.error, f)
                     else:
                         raise error.ProgrammingError(
                             'problem instance does not set warning or error '
                             'attribute: %s' % problem.msg)
 
             seen = {}
-            rp = None
             for i in fl:
                 revisions += 1
                 n = fl.node(i)
@@ -403,75 +413,15 b' class verifier(object):'
                     else:
                         del filenodes[f][n]
 
-                # Verify contents. 4 cases to care about:
-                #
-                #   common: the most common case
-                #   rename: with a rename
-                #   meta: file content starts with b'\1\n', the metadata
-                #         header defined in filelog.py, but without a rename
-                #   ext: content stored externally
-                #
-                # More formally, their differences are shown below:
-                #
-                #                       | common | rename | meta  | ext
-                #  -------------------------------------------------------
-                #   flags()             | 0      | 0      | 0     | not 0
-                #   renamed()           | False  | True   | False | ?
-                #   rawtext[0:2]=='\1\n'| False  | True   | True  | ?
-                #
-                # "rawtext" means the raw text stored in revlog data, which
-                # could be retrieved by "revision(rev, raw=True)". "text"
-                # mentioned below is "revision(rev, raw=False)".
-                #
-                # There are 3 different lengths stored physically:
-                #  1. L1: rawsize, stored in revlog index
-                #  2. L2: len(rawtext), stored in revlog data
-                #  3. L3: len(text), stored in revlog data if flags==0, or
-                #     possibly somewhere else if flags!=0
-                #
-                # L1 should be equal to L2. L3 could be different from them.
-                # "text" may or may not affect commit hash depending on flag
-                # processors (see revlog.addflagprocessor).
-                #
-                #              | common  | rename | meta  | ext
-                # -------------------------------------------------
-                #    rawsize() | L1      | L1     | L1    | L1
-                #       size() | L1      | L2-LM  | L1(*) | L1 (?)
-                # len(rawtext) | L2      | L2     | L2    | L2
-                #    len(text) | L2      | L2     | L2    | L3
-                #  len(read()) | L2      | L2-LM  | L2-LM | L3 (?)
-                #
-                # LM:  length of metadata, depending on rawtext
-                # (*): not ideal, see comment in filelog.size
-                # (?): could be "- len(meta)" if the resolved content has
-                #      rename metadata
-                #
-                # Checks needed to be done:
-                #  1. length check: L1 == L2, in all cases.
-                #  2. hash check: depending on flag processor, we may need to
-                #     use either "text" (external), or "rawtext" (in revlog).
-                try:
-                    skipflags = self.skipflags
-                    if skipflags:
-                        skipflags &= fl.flags(i)
-                    if not skipflags:
-                        fl.read(n) # side effect: read content and do checkhash
-                        rp = fl.renamed(n)
-                    # the "L1 == L2" check
-                    l1 = fl.rawsize(i)
-                    l2 = len(fl.revision(n, raw=True))
-                    if l1 != l2:
-                        self.err(lr, _("unpacked size is %s, %s expected") %
-                                 (l2, l1), f)
-                except error.CensoredNodeError:
-                    # experimental config: censor.policy
-                    if ui.config("censor", "policy") == "abort":
-                        self.err(lr, _("censored file data"), f)
-                except Exception as inst:
-                    self.exc(lr, _("unpacking %s") % short(n), inst, f)
+                if n in state['skipread']:
+                    continue
 
                 # check renames
                 try:
+                    # This requires resolving fulltext (at least on revlogs). We
+                    # may want ``verifyintegrity()`` to pass a set of nodes with
+                    # rename metadata as an optimization.
+                    rp = fl.renamed(n)
                     if rp:
                         if lr is not None and ui.verbose:
                             ctx = lrugetctx(lr)