commands: print chunk type in debugrevlog Each data entry ("chunk") in a revlog has a type based on the first byte of the data. This type indicates how to interpret the data. This seems like a useful thing to be able to query through a debug command. So let's add that to `hg debugrevlog`. This does make `hg debugrevlog` slightly slower, as it has to read more than just the index. However, even on the mozilla-unified manifest (which is ~200MB spread over ~350K revisions), this takes <400ms.

perf: add command for measuring revlog chunk operations Upcoming commits will teach revlogs to leverage the new compression engine API so that new compression formats can more easily be leveraged in revlogs. We want to be sure this refactoring doesn't regress performance. So this commit introduces "perfrevchunks" to explicitly test performance of reading, decompressing, and recompressing revlog chunks. Here is output when run on the mozilla-unified repo: $ hg perfrevlogchunks -c ! read ! wall 0.346603 comb 0.350000 user 0.340000 sys 0.010000 (best of 28) ! read w/ reused fd ! wall 0.337707 comb 0.340000 user 0.320000 sys 0.020000 (best of 30) ! read batch ! wall 0.013206 comb 0.020000 user 0.000000 sys 0.020000 (best of 221) ! read batch w/ reused fd ! wall 0.013259 comb 0.030000 user 0.010000 sys 0.020000 (best of 222) ! chunk ! wall 1.909939 comb 1.910000 user 1.900000 sys 0.010000 (best of 6) ! chunk batch ! wall 1.750677 comb 1.760000 user 1.740000 sys 0.020000 (best of 6) ! compress ! wall 5.668004 comb 5.670000 user 5.670000 sys 0.000000 (best of 3) $ hg perfrevlogchunks -m ! read ! wall 0.365834 comb 0.370000 user 0.350000 sys 0.020000 (best of 26) ! read w/ reused fd ! wall 0.350160 comb 0.350000 user 0.320000 sys 0.030000 (best of 28) ! read batch ! wall 0.024777 comb 0.020000 user 0.000000 sys 0.020000 (best of 119) ! read batch w/ reused fd ! wall 0.024895 comb 0.030000 user 0.000000 sys 0.030000 (best of 118) ! chunk ! wall 2.514061 comb 2.520000 user 2.480000 sys 0.040000 (best of 4) ! chunk batch ! wall 2.380788 comb 2.380000 user 2.360000 sys 0.020000 (best of 5) ! compress ! wall 9.815297 comb 9.820000 user 9.820000 sys 0.000000 (best of 3) We already see some interesting data, such as how much slower non-batched chunk reading is and that zlib compression appears to be >2x slower than decompression. I didn't have the data when I wrote this commit message, but I ran this on Mozilla's NFS-based Mercurial server and the time for reading with a reused file descriptor was faster. So I think it is worth testing both with and without file descriptor reuse so we can make informed decisions about recycling file descriptors.

setup: add flag to build_ext to control building zstd Downstream packagers will inevitably want to disable building the vendored python-zstandard Python package. Rather than force them to patch setup.py, let's give them a knob to use. distutils Command classes support defining custom options. It requires setting certain class attributes (yes, class attributes: instance attributes don't work because the class type is consulted before it is instantiated). We already have a custom child class of build_ext, so we set these class attributes, implement some scaffolding, and override build_extensions to filter the Extension instance for the zstd extension if the `--no-zstd` argument is specified. Example usage: $ python setup.py build_ext --no-zstd

drawdag: update test repos by drawing the changelog DAG in ASCII Currently, we have "debugbuilddag" which is a powerful tool to build test cases but not intuitive. We may end up running "hg log" in the test to make the test more readable. This patch adds a "drawdag" extension with a "debugdrawdag" command for similar testing purpose. Unlike the cryptic "debugbuilddag" command, it reads an ASCII graph that is intuitive to human, so the test case can be more readable. Unlike "debugbuilddag", "drawdag" does not require an empty repo. So it can be used to add new changesets to an existing repo. Since the "drawdag" logic is not that trivial and only makes sense for testing purpose, the extension is added to the "tests" directory, to make the core logic clean. If we find it useful (for example, to demonstrate cases and help user understand some cases) and want to ship it by default in the future, we can move it to a ship-by-default "debugdrawdag" at that time.

posix: give checklink a fast path that cache the check file and is read only util.checklink would create a symlink and remove it again. That would sometimes happen multiple times. Write operations are relatively expensive and give disk tear and noise for applications monitoring file system activity. Instead of creating a symlink and deleting it again, just create it once and leave it in .hg/cache/check-link . If the file exists, just verify that os.islink reports true. We will assume that this check is as good as symlink creation not failing. Note: The symlink left in .hg/cache has to resolve to a file - otherwise 'make dist' will fail ... test-symlink-os-yes-fs-no.py does some monkey patching to simulate a platform without symlink support. The slightly different testing method requires additional monkeying.

posix: move checklink test file to .hg/cache This avoids unnecessary churn in the working directory. It is not necessarily a fully valid assumption that .hg/cache is on the same filesystem as the working directory, but I think it is an acceptable approximation. It could also be the case that different parts of the working directory is on different mount points so checking in the root folder could also be wrong.

posix: give checkexec a fast path; keep the check files and test read only Before, Mercurial would create a new temporary file every time, stat it, change its exec mode, stat it again, and delete it. Most of this dance was done to handle the rare and not-so-essential case of VFAT mounts on unix. The cost of that was paid by the much more common and important case of using normal file systems. Instead, try to create and preserve .hg/cache/checkisexec and .hg/cache/checknoexec with and without exec flag set. If the files exist and have correct exec flags set, we can conclude that that file system supports the exec flag. Best case, the whole exec check can thus be done with two stat calls. Worst case, we delete the wrong files and check as usual. That will be because temporary loss of exec bit or on file systems without support for the exec bit. In that case we check as we did before, with the additional overhead of one extra stat call. It is possible that this different test algorithm in some cases on odd file systems will give different behaviour. Again, I think it will be rare and special cases and I think it is worth the risk. test-clone.t happens to show the situation where checkisexec is left behind from the old style check, while checknoexec only will be created next time a exec check will be performed.

posix: simplify checkexec check Use a slightly simpler logic that in some cases can avoid an unnecessary chmod and stat. Instead of flipping the X bits, make it more clear that we rely on no X bits being set on initial file creation, and that at least some of them stick after they all have been set.

posix: move checkexec test file to .hg/cache This avoids unnecessary churn in the working directory. It is not necessarily a fully valid assumption that .hg/cache is on the same filesystem as the working directory, but I think it is an acceptable approximation. It could also be the case that different parts of the working directory is on different mount points so checking in the root folder could also be wrong.

manifest: move manifestctx creation into manifestlog.get() Most manifestctx creation already happened in manifestlog.get(), but there was one spot in the manifestctx class itself that created an instance manually. This patch makes that one instance go through the manifestlog. This means extensions can just wrap manifestlog.get() and it will cover all manifestctx creations. It also means this code path now hits the manifestlog cache.

util: implement zstd compression engine Now that zstd is vendored and being built (in some configurations), we can implement a compression engine for zstd! The zstd engine is a little different from existing engines. Because it may not always be present, we have to defer load the module in case importing it fails. We facilitate this via a cached property that holds a reference to the module or None. The "available" method is implemented to reflect reality. The zstd engine declares its ability to handle bundles using the "zstd" human name and the "ZS" internal name. The latter was chosen because internal names are 2 characters (by only convention I think) and "ZS" seems reasonable. The engine, like others, supports specifying the compression level. However, there are no consumers of this API that yet pass in that argument. I have plans to change that, so stay tuned. Since all we need to do to support bundle generation with a new compression engine is implement and register the compression engine, bundle generation with zstd "just works!" Tests demonstrating this have been added. How does performance of zstd for bundle generation compare? On the mozilla-unified repo, `hg bundle --all -t <engine>-v2` yields the following on my i7-6700K on Linux: engine CPU time bundle size vs orig size throughput none 97.0s 4,054,405,584 100.0% 41.8 MB/s bzip2 (l=9) 393.6s 975,343,098 24.0% 10.3 MB/s gzip (l=6) 184.0s 1,140,533,074 28.1% 22.0 MB/s zstd (l=1) 108.2s 1,119,434,718 27.6% 37.5 MB/s zstd (l=2) 111.3s 1,078,328,002 26.6% 36.4 MB/s zstd (l=3) 113.7s 1,011,823,727 25.0% 35.7 MB/s zstd (l=4) 116.0s 1,008,965,888 24.9% 35.0 MB/s zstd (l=5) 121.0s 977,203,148 24.1% 33.5 MB/s zstd (l=6) 131.7s 927,360,198 22.9% 30.8 MB/s zstd (l=7) 139.0s 912,808,505 22.5% 29.2 MB/s zstd (l=12) 198.1s 854,527,714 21.1% 20.5 MB/s zstd (l=18) 681.6s 789,750,690 19.5% 5.9 MB/s On compression, zstd for bundle generation delivers: * better compression than gzip with significantly less CPU utilization * better than bzip2 compression ratios while still being significantly faster than gzip * ability to aggressively tune compression level to achieve significantly smaller bundles That last point is important. With clone bundles, a server can pre-generate a bundle file, upload it to a static file server, and redirect clients to transparently download it during clone. The server could choose to produce a zstd bundle with the highest compression settings possible. This would take a very long time - a magnitude longer than a typical zstd bundle generation - but the result would be hundreds of megabytes smaller! For the clone volume we do at Mozilla, this could translate to petabytes of bandwidth savings per year and faster clones (due to smaller transfer size). I don't have detailed numbers to report on decompression. However, zstd decompression is fast: >1 GB/s output throughput on this machine, even through the Python bindings. And it can do that regardless of the compression level of the input. By the time you have enough data to worry about overhead of decompression, you have plenty of other things to worry about performance wise. zstd is wins all around. I can't wait to implement support for it on the wire protocol and in revlogs.

hghave: add check for zstd support Not all configurations will support zstd. Add a check so we can conditionalize tests.

exchange: obtain compression engines from the registrar util.compengines has knowledge of all registered compression engines and the metadata that associates them with various bundle types. This patch removes the now redundant declaration of this metadata from exchange.py and obtains it from the new source. The effect of this patch is that once a new compression engine is registered with util.compengines, `hg bundle -t <engine>` will just work.

bundle2: equate 'UN' with no compression An upcoming patch will change the "alg" argument passed to this function from None to "UN" when no compression is wanted. The existing implementation of bundle2 does not set a "Compression" parameter if no compression is used. In theory, setting "Compression=UN" should work. But I haven't audited the code to see if all client versions supporting bundle2 will accept this. Rather than take the risk, avoid the BC breakage and treat "UN" the same as None.

util: check for compression engine availability before returning If a requested compression engine is registered but not available, requesting it will now abort. To be honest, I'm not sure if this is the appropriate mechanism for handling optional compression engines. I won't know until all uses of compression (bundles, wire protocol, revlogs, etc) are using the new API and zstd (our planned optional engine) is implemented. So this API could change.

util: expose an "available" API on compression engines When the zstd compression engine is introduced, it won't work in all installations, namely pure Python installs. So, we need a mechanism to declare whether a compression engine is available. We don't want to conditionally register the compression engine because it is sometimes useful to know when a compression engine name or encountered data is valid but just not available versus unknown.

setup: compile zstd C extension Now that zstd and python-zstandard are vendored, we can start compiling them as part of the install. python-zstandard provides a self-contained Python function that returns a distutils.extension.Extension, so it is really easy to add zstd to our setup.py without having to worry about defining source files, include paths, etc. The function even allows specifying the module name the extension should be compiled as. This conveniently allows us to compile the module into the "mercurial" package so "our" version won't collide with a version installed under the canonical "zstd" module name.

zstd: vendor python-zstandard 0.5.0 As the commit message for the previous changeset says, we wish for zstd to be a 1st class citizen in Mercurial. To make that happen, we need to enable Python to talk to the zstd C API. And that requires bindings. This commit vendors a copy of existing Python bindings. Why do we need to vendor? As the commit message of the previous commit says, relying on systems in the wild to have the bindings or zstd present is a losing proposition. By distributing the zstd and bindings with Mercurial, we significantly increase our chances that zstd will work. Since zstd will deliver a better end-user experience by achieving better performance, this benefits our users. Another reason is that the Python bindings still aren't stable and the API is somewhat fluid. While Mercurial could be coded to target multiple versions of the Python bindings, it is safer to bundle an explicit, known working version. The added Python bindings are mostly a fully-featured interface to the zstd C API. They allow one-shot operations, streaming, reading and writing from objects implements the file object protocol, dictionary compression, control over low-level compression parameters, and more. The Python bindings work on Python 2.6, 2.7, and 3.3+ and have been tested on Linux and Windows. There are CFFI bindings, but they are lacking compared to the C extension. Upstream work will be needed before we can support zstd with PyPy. But it will be possible. The files added in this commit come from Git commit e637c1b214d5f869cf8116c550dcae23ec13b677 from https://github.com/indygreg/python-zstandard and are added without modifications. Some files from the upstream repository have been omitted, namely files related to continuous integration. In the spirit of full disclosure, I'm the maintainer of the "python-zstandard" project and have authored 100% of the code added in this commit. Unfortunately, the Python bindings have not been formally code reviewed by anyone. While I've tested much of the code thoroughly (I even have tests that fuzz APIs), there's a good chance there are bugs, memory leaks, not well thought out APIs, etc. If someone wants to review the code and send feedback to the GitHub project, it would be greatly appreciated. Despite my involvement with both projects, my opinions of code style differ from Mercurial's. The code in this commit introduces numerous code style violations in Mercurial's linters. So, the code is excluded from most lints. However, some violations I agree with. These have been added to the known violations ignore list for now.

zstd: vendor zstd 1.1.1 zstd is a new compression format and it is awesome, yielding higher compression ratios and significantly faster compression and decompression operations compared to zlib (our current compression engine of choice) across the board. We want zstd to be a 1st class citizen in Mercurial and to eventually be the preferred compression format for various operations. This patch starts the formal process of supporting zstd by vendoring a copy of zstd. Why do we need to vendor zstd? Good question. First, zstd is relatively new and not widely available yet. If we didn't vendor zstd or distribute it with Mercurial, most users likely wouldn't have zstd installed or even available to install. What good is a feature if you can't use it? Vendoring and distributing the zstd sources gives us the highest liklihood that zstd will be available to Mercurial installs. Second, the Python bindings to zstd (which will be vendored in a separate changeset) make use of zstd APIs that are only available via static linking. One reason they are only available via static linking is that they are unstable and could change at any time. While it might be possible for the Python bindings to attempt to talk to different versions of the zstd C library, the safest thing to do is link against a specific, known-working version of zstd. This is why the Python zstd bindings themselves vendor zstd and why we must as well. This also explains why the added files are in a "python-zstandard" directory. The added files are from the 1.1.1 release of zstd (Git commit 4c0b44f8ced84c4c8edfa07b564d31e4fa3e8885 from https://github.com/facebook/zstd) and are added without modifications. Not all files from the zstd "distribution" have been added. Notably missing are files to support interacting with "legacy," pre-1.0 versions of zstd. The decision of which files to include is made by the upstream python-zstandard project (which I'm the author of). The files in this commit are a snapshot of the files from the 0.5.0 release of that project, Git commit e637c1b214d5f869cf8116c550dcae23ec13b677 from https://github.com/indygreg/python-zstandard.

bdiff: give slight preference to removing trailing lines [This change could be folded into the previous changeset to minimize the repo churn ...] Similar to the previous change, introduce an exception to the general preference for matches in the middle of bdiff ranges: If the best match on the B side starts at the beginning of the bdiff range, don't aim for the middle-most A side match but for the earliest. New (later) matches on the A side will only be considered better if the corresponding match on the B side *not* is at the beginning of the range. Thus, if the best (middle-most) match on the B side turns out to be at the beginning of the range, the earliest match on the A side will be used. The bundle size for 4.0 (hg bundle --base null -r 4.0 x.hg) happens to go from 22807275 to 22808120 bytes - a 0.004% increase.

bdiff: give slight preference to appending lines [This change could be folded into the previous changeset to minimize the repo churn ...] The general preference to matches in the middle of bdiff ranges helps getting balanced recursion and efficient computation. But, as previous changes have shown, it might also give diffs that seems "obviously wrong". To mitigate that: If the best match on the A side starts at the beginning of the bdiff range, don't aim for the middle-most B side match but for the earliest. This will make the matches balanced (by both sides being "early") even though the bisection will be less balanced. Still, this case only apply if the *best* and middle-most match was fully unbalanced on the A side. Each recursion will thus even in this worst case reduce the problem significantly and we are not re-introducing the problem that was fixed in f1ca249696ed. The bundle size for 4.0 (hg bundle --base null -r 4.0 x.hg) happens to go from 22806817 to 22807275 bytes - a 0.002% increase. This make the recent test-bdiff.py changes give a more pretty output ... but they no longer show that the recursion is around middle matches (because it in these cases isn't).

bdiff: give slight preference to longest matches in the middle of the B side We already have a slight preference for matches close to the middle on the A side. Now, do the same on the B side. j is iterating the b range backwards and we thus accept a new j if the previous match was in the upper half. This makes the test-bhalf diff "correct". It obviously also gives more preference to balanced recursion than to appending to sequences. That is kind of correct, but will also unfortunately make some bundles bigger. No doubt, we can also create examples where it will make them smaller ... The bundle size for 4.0 (hg bundle --base null -r 4.0 x.hg) happens to go from 22803824 to 22806817 bytes - an 0.01% increase.

bdiff: rearrange the "better longest match" code This is primarily to make the code more managable and prepare for later changes. More specific assignments might also be slightly faster, even thought it also might generate a bit more code.

bdiff: adjust criteria for getting optimal longest match in the A side middle We prefer matches closer to the middle to balance recursion, as introduced in f1ca249696ed. For ranges with uneven length, matches starting exactly in the middle should have preference. That will be optimal for matches of length 1. We will thus accept equality in the half check. For ranges with even length, half was ceil'ed when calculated but we got the preference for low matches from the 'less than half' check. To get the same result as before when we also accept equality, floor it. Without that, test-annotate.t would show some different (still correct but less optimal) results. This will change the heuristics. Tests shows a slightly different output - and sometimes slightly smaller bundles. The bundle size for 4.0 (hg bundle --base null -r 4.0 x.hg) happens to go from 22804885 to 22803824 bytes - an 0.005% reduction.

tests: explore some bdiff cases

tests: make test-bdiff.py easier to maintain Add more stdout logging to help navigate the .out file.

perf: unbust perfbdiff --alldata This broke in f84fc6a92817 due to a refactored manifest API. The fix is a bit hacky - perfbdiff doesn't yet support tree manifests for example. But it gets the job done. A test has been added for --alldata so this doesn't happen again.

worker: discard waited pid by anyone who noticed it first This makes sure all waited pids are removed before calling killworkers() even if waitpid()-pids.discard() sequence is interrupted by another SIGCHLD.

worker: kill workers after all zombie processes are reaped Since we now wait child processes in non-blocking way (changed by 7bc25549e084 and e8fb03cfbbde), we don't have to kill them in the middle of the waitpid() loop. This change will help solving a possible race of waitpid()-pids.discard() sequence and another SIGCHLD. waitforworkers() is called by cleanup(), in which case we do killworkers() beforehand so we can remove killworkers() from waitforworkers().

worker: make sure killworkers() never be interrupted by another SIGCHLD killworkers() iterates over pids, which can be updated by SIGCHLD handler. So we should either copy pids or prevent killworkers() from being interrupted by SIGCHLD. I chose the latter as it is simpler and can make pids handling more consistent. This fixes a possible "set changed size during iteration" error at killworkers() before cleanup().