Skip to content
Snippets Groups Projects
  1. Jan 22, 2018
  2. Jan 21, 2018
    • Gregory Szorc's avatar
      bundle2: increase payload part chunk size to 32kb · 4fb2bb61597c
      Gregory Szorc authored
      Bundle2 payload parts are framed chunks. Esentially, we obtain
      data in equal size chunks of size `preferedchunksize` and emit those
      to a generator. That generator is fed into a compressor (which can
      be the no-op compressor, which just re-emits the generator). And
      the output from the compressor likely goes to a file descriptor
      or socket.
      
      What this means is that small chunk sizes create more Python objects
      and Python function calls than larger chunk sizes. And as we know,
      Python object and function call overhead in performance sensitive
      code matters (at least with CPython).
      
      This commit increases the bundle2 part payload chunk size from 4k
      to 32k. Practically speaking, this means that the chunks we feed
      into a compressor (implemented in C code) or feed directly into a
      file handle or socket write() are larger. It's possible the chunks
      might be larger than what the receiver can handle in one logical
      operation. But at that point, we're in C code, which is much more
      efficient at dealing with splitting up the chunk and making multiple
      function calls than Python is.
      
      A downside to larger chunks is that the receiver has to wait for that
      much data to arrive (either raw or from a decompressor) before it
      can process the chunk. But 32kb still feels like a small buffer to
      have to wait for. And in many cases, the client will convert from
      8 read(4096) to 1 read(32768). That's happening in Python land. So
      we cut down on the number of Python objects and function calls,
      making the client faster as well. I don't think there are any
      significant concerns to increasing the payload chunk size to 32kb.
      
      The impact of this change on performance significant. Using `curl`
      to obtain a stream clone bundle2 payload from a server on localhost
      serving the mozilla-unified repository:
      
      before: 20.78 user; 7.71 system;  80.5 MB/s
      after:  13.90 user; 3.51 system; 132 MB/s
      legacy:  9.72 user; 8.16 system; 132 MB/s
      
      bundle2 stream clone generation is still more resource intensive than
      legacy stream clone (that's likely because of the use of a
      util.chunkbuffer). But the throughput is the same. We might
      be in territory we're this is effectively a benchmark of the
      networking stack or Python's syscall throughput.
      
      From the client perspective, `hg clone -U --stream`:
      
      before: 33.50 user; 7.95 system; 53.3 MB/s
      after:  22.82 user; 7.33 system; 72.7 MB/s
      legacy: 29.96 user; 7.94 system; 58.0 MB/s
      
      And for `hg clone --stream` with a working directory update of
      ~230k files:
      
      after:  119.55 user; 26.47 system; 0:57.08 wall
      legacy: 126.98 user; 26.94 system; 1:05.56 wall
      
      So, it appears that bundle2's stream clone is now definitively faster
      than legacy stream clone!
      
      Differential Revision: https://phab.mercurial-scm.org/D1932
      4fb2bb61597c
  3. Jan 22, 2018
    • Gregory Szorc's avatar
      bundle2: always advertise client support for stream parts · 1d118f9f4f57
      Gregory Szorc authored
      Previously, enabling of stream clone over bundle2 was a server-side
      only change. And clients would only advertise bundle2 support for
      stream clones if an experimental config option was set.
      
      This situation wasn't forward compatible because in the future
      (when the feature is enabled on servers by default), an old client
      would send a request to the server but it wouldn't send its own
      bundle2 capability support for stream parts. Servers would have to
      infer that clients not sending this capability were old Mercurial
      clients that only sent the capability if the feature was
      explicitly enabled. Implicit and inferred behavior makes implementing
      servers hard. It's much better to be explicit about client features.
      
      We should either make the config option for bundle2 stream clones
      disable the feature client-side as well (so a server doesn't see
      a request from a client not advertising stream support). Or we
      should always advertise stream support if a client is willing
      to accept stream parts.
      
      Since I anticipating stabilizing stream clone support in bundle2
      quickly, I think it's safe to always advertise client support
      in the bundle2 capabilities. So this commit changes things to
      do that.
      
      Because capabilities now vary between client and server, we had
      to create variations of the variable substitutions for those
      strings.
      
      Differential Revision: https://phab.mercurial-scm.org/D1931
      1d118f9f4f57
    • Gregory Szorc's avatar
      exchange: don't send stream data when server.uncompressed is set · 9adae6a20e1f
      Gregory Szorc authored
      Previously, bundle2 stream support would send out data even though
      the streaming clone feature was disabled. This commit changes
      the part handler to respect the server config.
      
      Differential Revision: https://phab.mercurial-scm.org/D1930
      9adae6a20e1f
    • Gregory Szorc's avatar
      bundle2: don't advertise stream bundle2 capability when feature disabled · c1b9eb15a51c
      Gregory Szorc authored
      The server.uncompressed config option can be used to disable streaming
      clones. While the top-level capability to advertise streaming clone
      support isn't advertised when this option is set, we were still sending
      the bundle2-level capabilities advertising support for stream parts.
      It makes sense to not advertise that support when streaming clones
      are globally disabled.
      
      If the structure of the new code seems a bit odd, it is because we'll
      have to further tweak the behavior in commit(s) ahead.
      
      Differential Revision: https://phab.mercurial-scm.org/D1929
      c1b9eb15a51c
    • Gregory Szorc's avatar
      tests: add more testing around server.uncompressed · 8ed5f7609728
      Gregory Szorc authored
      We already have testing for server.uncompressed in test-http*.t.
      However, it doesn't cover the new bundle2 use case. And, we don't
      have comprehensive testing of advertised capabilities.
      
      We add tests to test-clone-uncompressed.t that demonstrate
      behavior for both legacy and bundle2 configurations.
      
      If you look closely, the bundle2 capabilities are advertising
      stream support when it isn't enabled. That's a bug.
      
      In addition, while the client is smart enough to not request
      a stream clone when the server doesn't have the feature enabled,
      the getbundle wire protocol command is still sending stream
      clone data. This doesn't match the behavior of the legacy
      stream_out wire protocol command. That's also a bug. Tests
      have been added.
      
      While I was here, I also changed how the PID is recorded in
      $DAEMON_PIDS. If we kill a process, the PID formerly in
      $DAEMON_PIDS no longer exists. So we should replace that file
      instead of appending to it.
      
      Differential Revision: https://phab.mercurial-scm.org/D1928
      8ed5f7609728
    • Gregory Szorc's avatar
      bundle2: move version of stream clone into part name · b116a66bcc44
      Gregory Szorc authored
      I don't like having version numbers as part parameters. It means
      that parts can theoretically vary wildly in their generation and
      processing semantics. I think that a named part should have consistent
      behavior over time. In other words, if you need to introduce new
      functionality or behavior, that should be expressed by inventing
      a new bundle2 part, not adding functionality to an existing part.
      
      This commit applies this advice to the just-introduced stream clone
      via bundle2 feature.
      
      The "version" part parameter is removed. The name of the bundle2 part
      is now "stream2" instead of "stream" with "version=v2".
      
      Differential Revision: https://phab.mercurial-scm.org/D1927
      b116a66bcc44
    • Gregory Szorc's avatar
      exchange: send bundle2 stream clones uncompressed · a84dbc87dae9
      Gregory Szorc authored
      Stream clones don't compress well. And compression undermines
      a point of stream clones which is to trade significant CPU
      reductions by increasing size.
      
      Building upon our introduction of metadata to communicate bundle
      information back to callers of exchange.getbundlechunks(), we add
      an attribute to the bundler that communicates whether the bundle is
      best left uncompressed. We return this attribute as part of the bundle
      metadata. And the wire protocol honors it when determining whether
      to compress the wire protocol response.
      
      The added test demonstrates that the raw result from the wire
      protocol is not compressed. It also demonstrates that the server
      will serve stream responses when the feature isn't enabled. We'll
      address that in another commit.
      
      The effect of this change is that server-side CPU usage for bundle2
      stream clones is significantly reduced by removing zstd compression.
      For the mozilla-unified repository:
      
      before: 37.69 user 8.01 system
      after:  27.38 user 7.34 system
      
      Assuming things are CPU bound, that ~10s reduction would translate to
      faster clones on the client. zstd can decompress at >1 GB/s. So the
      overhead from decompression on the client is small in the grand scheme
      of things. But if zlib compression were being used, the overhead would
      be much greater.
      
      Differential Revision: https://phab.mercurial-scm.org/D1926
      a84dbc87dae9
    • Gregory Szorc's avatar
      tests: update test to work with Git 2.16 · 8144f1b07e21
      Gregory Szorc authored
      It looks like Git 2.16 removed the "..." from some strings.
      Glob over those characters in the test output.
      
      Differential Revision: https://phab.mercurial-scm.org/D1935
      8144f1b07e21
  4. Jan 20, 2018
    • Gregory Szorc's avatar
      exchange: return bundle info from getbundlechunks() (API) · ba15580e53d5
      Gregory Szorc authored
      We generally want a mechanism to pass information about the
      generated bundle back to callers (in addition to the byte stream).
      Ideally we would return a bundler from this function and have the
      caller code to an interface. But the bundling APIs are not great
      and getbundlechunks() is the best API we have for obtaining bundle
      contents in a unified manner.
      
      We change getbundlechunks() to return a dict that we can use to
      communicate metadata.
      
      We populate that dict with the bundle version number to demonstrate
      some value.
      
      .. api::
      
         exchange.getbundlechunks() now returns a 2-tuple instead of just
         an iterator.
      
      Differential Revision: https://phab.mercurial-scm.org/D1925
      ba15580e53d5
    • Gregory Szorc's avatar
      exchange: make stream bundle part deterministic · 08cc94dd3d3c
      Gregory Szorc authored
      repo.requirements is a set. We need to sort it so the part
      content is deterministic.
      
      Differential Revision: https://phab.mercurial-scm.org/D1924
      08cc94dd3d3c
    • Gregory Szorc's avatar
      bundle2: specify what capabilities will be used for · c97639ad6874
      Gregory Szorc authored
      We currently assume there is a symmetric relationship of bundle2
      capabilities between client and server. However, this may not always be
      the case.
      
      We need a bundle2 capability to advertise bundle2 streaming clone support
      on servers to differentiate it from the existing, legacy streaming clone
      support.
      
      However, servers may wish to disable streaming clone support. If bundle2
      capabilities were the same between client and server, a client (which
      may also be a server) that has disabled streaming clone support would
      not be able to perform a streaming clone itself!
      
      This commit introduces a "role" argument to bundle2.getrepocaps() that
      explicitly defines the role being performed. This will allow us (and
      extensions) to alter bundle2 capabilities depending on the operation
      being performed.
      
      Differential Revision: https://phab.mercurial-scm.org/D1923
      c97639ad6874
    • Gregory Szorc's avatar
      wireproto: don't compress errors from getbundle() · 9d249f3de730
      Gregory Szorc authored
      Errors should be small. There's no real need to compress them.
      Truth be told, there's no good reason to not compress them either.
      But leaving them uncompressed makes it easier to test failures
      by looking at the raw HTTP response. This makes it easier for us
      to write tests. It may make it easier for people writing their
      own clients.
      
      Differential Revision: https://phab.mercurial-scm.org/D1922
      9d249f3de730
  5. Jan 21, 2018
  6. Jan 20, 2018
  7. Jan 22, 2018
    • Gregory Szorc's avatar
      convert: use a collections.deque · eefabd9ed3e1
      Gregory Szorc authored
      This function was doing a list.pop(0) on a list whose length
      was the number of revisions to convert. Popping an early element
      from long lists is not an efficient operation.
      
      collections.deque supports efficient inserts and pops at both
      ends. So we switch to that data structure.
      
      When converting the mozilla-unified repository, which has 445,748
      revisions, this change makes the "sorting..." step of
      `hg convert --sourcesort` significantly faster:
      
      before: ~59.2s
      after:   ~1.3s
      
      Differential Revision: https://phab.mercurial-scm.org/D1934
      eefabd9ed3e1
  8. Jan 21, 2018
    • Martin von Zweigbergk's avatar
      repair: invalidate volatile sets after stripping · 128dd940bedc
      Martin von Zweigbergk authored
      Matt Harbison reported that some tests were broken on Windows after
      1a09dad8b85a (evolution: report new unstable changesets,
      2018-01-14). The failures were exactly as seen in this patch. The
      failures actually seemed correct, which made me wonder why they didn't
      fail the same way on Linux. It turned out to be a cache invalidation
      problem.
      
      The new orphan mentioned in the test case actually does get created
      when we're re-applying the temporary bundle that's created while
      stripping. However, without the invalidation, it appears that there
      was already an orphan before applying the temporary bundle.
      
      The warnings about unknown working parent appear because the
      aformentioned changeset means that we're now accessing the dirstate
      while it's invalid.
      
      We may want to suppress these messages that happen in the intermediate
      strip state, but they're technically correct (although confusing to
      the user), so I think just fixing the cache invalidation is fine for
      now.
      
      I haven't figured out why the caches seemed to get correctly
      invalidated on Windows.
      
      Differential Revision: https://phab.mercurial-scm.org/D1933
      128dd940bedc
    • Matt Harbison's avatar
      subrepo: handle 'C:' style paths on the command line (issue5770) · 0c0689a7565e
      Matt Harbison authored
      If you think 'C:' and 'C:\' are equivalent paths, see the inline comment before
      proceeding.
      
      The problem here was that several commands that take a URL argument (incoming,
      outgoing, pull, and push) will use that value to set 'repo._subtoppath' on the
      repository object after command specific manipulation of it, but before
      converting it to an absolute path.  When an operation is performed on a relative
      subrepo, subrepo._abssource() will posixpath.join() this value with the relative
      subrepo path.  That adds a '/' after the drive letter, changing how it is
      evaluated by abspath()/realpath() in vfsmod.vfs(..., realpath=True) as the
      subrepo is instantiated.
      
      I initially tried sanitizing the path in url.localpath(), because url.isabs()
      only checks that it starts with a drive letter.  By the sample behavior, this is
      clearly not an absolute path.  (Though the comment in isabs() is weasely- this
      style path can't be joined either.)  But not everything funnels through there,
      and it required explicitly calling localpath() in hg.parseurl() and assigning to
      url.path to fix.  But then tests failed with urls like 'a#0'.
      
      Next up was sanitizing the path in the url constructor.  That caused doctest
      failures, because there are drive letter tests, so those got expanded in system
      specific ways.  Yuya correctly pointed out that util.url is a parser, and
      shouldn't be substituting the path too.
      
      Rather than fixing every command call site, just convert it in the common
      subrepo location.  I don't see any sanitizing on the path config options, so I
      fixed those too.  Note that while the behavior is fixed here, there are still
      places where 'comparing with C:' gets printed out, and that's not great for
      debugging purposes.  (Specifically I saw it in `hg incoming -B C:`, without
      subrepos.)  While clone will write out an absolute default path, I wonder what
      would happen if a user edited that path to be 'C:'.  (I don't think supporting
      relative paths in .hgrc is a sane thing to do, but while we're poking holes in
      things...)
      
      Since this is such an oddball case, it still leaks through in places, and there
      seems to be a lot of duplicate url parsing, maybe the url parsing should be
      moved to dispatch, and provide the command with a url object?  Then we could
      convert this to an absolute path once, and not have to worry about it in the
      rest of the code.
      
      I also checked '--cwd C:' on the command line, and it was previously working
      because os.chdir() will DTRT.
      
      Finally, one other note from the url.localpath() experimenting.  I don't see any
      cases where 'self._hostport' can hold a drive letter.  So I'm wondering if that
      is wrong/old code.
      0c0689a7565e
  9. Jan 22, 2018
    • Matt Harbison's avatar
      dummysmtpd: don't die on client connection errors · 75bae69747f0
      Matt Harbison authored
      The connection refused error in test-patchbomb-tls.t[1] is sporadic, but one of
      the more often seen errors on Windows.  I added enough logging to a file and
      dumped it out at the end to make the following observations:
      
        - The listening socket is successfully created and bound to the port, and the
          "listening at..." message is always logged.
      
        - Generally, the following is the entire log output, with the "accepted ..."
          message having been added after `sslutil.wrapserversocket`:
      
                listening at localhost:$HGPORT
                $LOCALIP ssl error
                accepted connect
                accepted connect
                $LOCALIP from=quux to=foo, bar
                $LOCALIP ssl error
      
        - In the cases that fail, asyncore.loop() in the run() method is exiting, but
          not with an exception.
      
        - In the cases that fail, the following is logged right after "listening ...":
      
              Traceback (most recent call last):
                File "c:\\Python27\\lib\\asyncore.py", line 83, in read
                  obj.handle_read_event()
                File "c:\\Python27\\lib\\asyncore.py", line 443, in handle_read_event
                  self.handle_accept()
                File "../tests/dummysmtpd.py", line 80, in handle_accept
                  conn = sslutil.wrapserversocket(conn, ui, certfile=self._certfile)
                File "..\\mercurial\\sslutil.py", line 570, in wrapserversocket
                  return sslcontext.wrap_socket(sock, server_side=True)
                File "c:\\Python27\\lib\\ssl.py", line 363, in wrap_socket
                  _context=self)
                File "c:\\Python27\\lib\\ssl.py", line 611, in __init__
                  self.do_handshake()
                File "c:\\Python27\\lib\\ssl.py", line 840, in do_handshake
                  self._sslobj.do_handshake()
              error: [Errno 10054] $ECONNRESET$
      
        - If the base class handler is overridden completely, the the first "ssl
          error" line is replaced by the stacktrace, but the other lines are
          unchanged.  The client behaves no differently, whether or not the server
          stacktraced.
      
      In general, `./run-tests.py --local -j9 -t9000 test-patchbomb-tls.t
      --runs-per-test 20` would show the issue after a run or two.  With this change,
      `./run-tests.py --local -j9 -t9000 test-patchbomb-tls.t --loop` ran 800 times
      without a hiccup.  This makes me wonder if the other connection refused messages
      that bubble up on occasion are caused by a similar issue.  It seems a bit
      drastic to kill the whole server on account of a single communication failure
      with a client.
      
      # no-check-commit because of handle_error()
      
      [1] https://buildbot.mercurial-scm.org/builders/Win7%20x86_64%20hg%20tests/builds/421/steps/run-tests.py%20%28python%202.7.13%29/logs/stdio
      75bae69747f0
  10. Jan 21, 2018
  11. Jan 20, 2018
  12. Jan 17, 2018
    • Boris Feld's avatar
      streamclone: also stream caches to the client · 5f5fb279fd39
      Boris Feld authored
      When stream clone is used over bundle2, relevant cache files are also streamed.
      This is expected to be a massive performance win for clone since no important
      cache will have to be recomputed.
      
      Some performance numbers:
      
      (All times are wall-clock times in seconds, 2 attempts per case.)
      
      # Mozilla-Central
      
      ## Clone over ssh over lan
      V1 streaming: 234.3 239.6
      V2 streaming: 248.4 243.7
      
      ## Clone over ssh over Internet
      V1 streaming: 175.5 110.9
      V2 streaming: 109.1 111.0
      
      ## Clone over HTTP over lan
      V1 streaming: 105.3 105.6
      V2 streaming: 112.7 111.4
      
      ## Clone over HTTP over internet
      V1 streaming: 105.6 114.6
      V2 streaming: 226.7 225.9
      
      ## Hg tags
      V1 streaming (no cache): 1.084 1.071
      V2 streaming (cache):    0.312 0.325
      
      ## Hg branches
      V1 streaming (no cache):   14.047 14.148
      V2 streaming (with cache):  0.312  0.333
      
      # Pypy
      
      ## Clone over ssh over internet
      V1 streaming: 29.4 30.1
      V2 streaming: 31.2 30.1
      
      ## Clone over http over internet
      V1 streaming: 29.7 29.7
      V2 streaming: 75.2 72.9
      
      (since ssh and lan are not affected, there seems to be an issue with how we
      read/write the http stream on connection with latency, unrelated to the format)
      
      ## Hg tags
      V1 streaming (no cache):   1.752 1.664
      V2 streaming (with cache): 0.274 0.260
      
      ## Hg branches
      V1 streaming (no cache):   4.469 4.728
      V2 streaming (with cache): 0.318 0.321
      
      # Private repository:
          * 500K revision revisions
          * 11K topological heads
          * 28K branch heads
      
      ## hg tags
      no cache:   1543.332
      with cache:    4.900
      
      ## hg branches
      no cache:   91.828
      with cache:  2.955
      5f5fb279fd39
    • Boris Feld's avatar
      caches: make 'cachetocopy' available in scmutil · 72fdd99eb526
      Boris Feld authored
      For more code to use this information, we need it to be more publicly available.
      72fdd99eb526
    • Boris Feld's avatar
      streamclone: add support for cloning non append-only file · 56c30b31afbe
      Boris Feld authored
      The phaseroots are stored in a non append-only file in the repository. We
      include them in the stream too. Since they are not append-only, we have to
      keep a copy around while we hold the lock to be able to stream them later.
      
      Since phase get exchanged within the stream we can skip requesting them
      independently.
      
      As a side effect, this will fixes issue5648 once the feature is enabled by
      default.
      56c30b31afbe
  13. Jan 18, 2018
  14. Jan 17, 2018
Loading