patchgit: add --no-optional-locks option

Johannes, this is an adaptation of your 67e5ce7f63 (status: offer *not* to lock the index and update it, 2016-08-12). Folks working on GitHub Desktop complained to me that it's only available on Windows. :)

I expanded the scope a bit to let us give the same treatment to more commands in the long run. I'd also be OK with just cherry-picking your patch to non-Windows Git if you don't find my reasoning below compelling. But I think we need _something_ like this, as the other solutions I could come up with don't seem very promising.

-Peff

-- >8 -- Some tools like IDEs or fancy editors may periodically run commands like "git status" in the background to keep track of the state of the repository. Some of these commands may refresh the index and write out the result in an opportunistic way: if they can get the index lock, then they update the on-disk index with any updates they find. And if not, then their in-core refresh is lost and just has to be recomputed by the next caller.

But taking the index lock may conflict with other operations in the repository. Especially ones that the user is doing themselves, which _aren't_ opportunistic. In other words, "git status" knows how to back off when somebody else is holding the lock, but other commands don't know that status would be happy to drop the lock if somebody else wanted it.

There are a couple possible solutions:

  1. Have some kind of "pseudo-lock" that allows other
     commands to tell status that they want the lock.

     This is likely to be complicated and error-prone to
     implement (and maybe even impossible with just
     dotlocks to work from, as it requires some
     inter-process communication).

  2. Avoid background runs of commands like "git status"
     that want to do opportunistic updates, preferring
     instead plumbing like diff-files, etc.

     This is awkward for a couple of reasons. One is that
     "status --porcelain" reports a lot more about the
     repository state than is available from individual
     plumbing commands. And two is that we actually _do_
     want to see the refreshed index. We just don't want to
     take a lock or write out the result. Whereas commands
     like diff-files expect us to refresh the index
     separately and write it to disk so that they can depend
     on the result. But that write is exactly what we're
     trying to avoid.

  3. Ask "status" not to lock or write the index.

     This is easy to implement. The big downside is that any
     work done in refreshing the index for such a call is
     lost when the process exits. So a background process
     may end up re-hashing a changed file multiple times
     until the user runs a command that does an index
     refresh themselves.

This patch implements the option 3. The idea (and the test) is largely stolen from a Git for Windows patch by Johannes Schindelin, 67e5ce7f63 (status: offer *not* to lock the index and update it, 2016-08-12). The twist here is that instead of making this an option to "git status", it becomes a "git" option and matching environment variable.

The reason there is two-fold:

  1. An environment variable is carried through to
     sub-processes. And whether an invocation is a
     background process or not should apply to the whole
     process tree. So you could do "git --no-optional-locks
     foo", and if "foo" is a script or alias that calls
     "status", you'll still get the effect.

  2. There may be other programs that want the same
     treatment.

     I've punted here on finding more callers to convert,
     since "status" is the obvious one to call as a repeated
     background job. But "git diff"'s opportunistic refresh
     of the index may be a good candidate.

The test is taken from 67e5ce7f63, and it's worth repeating Johannes's explanation:

  Note that the regression test added in this commit does
  not *really* verify that no index.lock file was written;
  that test is not possible in a portable way. Instead, we
  verify that .git/index is rewritten *only* when `git
  status` is run without `--no-optional-locks`.

Signed-off-by: Jeff King <peff@peff.net>
---
 Documentation/git.txt | 13 +++++++++++++
 builtin/commit.c      |  5 ++++-
 cache.h               |  6 ++++++
 environment.c         |  5 +++++
 git.c                 |  4 ++++
 t/t7508-status.sh     | 10 ++++++++++
 6 files changed, 42 insertions(+), 1 deletion(-)

Jeff King <peff@peff.net> writes:

The phrase 'optional lock' does not answer this question clearly, though: does it make sense to extend the coverage of this option in the future to things more than the "opportunistic update to the index file"?

If the answer is no, then having 'index' instead of 'lock' in the name of the option would make more sense (and 'opportunistic' over 'optional', too), because what the change is about is to allow other processes that are directly interacting with the user to update the index, and 'lock' being hindrance is merely an implementation detail. The comment on the "test" in the log message mentions as if it were a short-coming that it does not check the lock but checks if the index is written, but I think that is testing what matters and preferable than testing "did we lock and then unlock it?"

On the other hand, if the answer is yes, then I am curious what other things this may extend to, and if these other things are also opportunistic optimizations.

Other than that, I think this change (including the part that this is done globally and down to subprocesses as needed) makes sense.

Thanks (and sorry for not being Johannes ;-).

On Thu, Sep 21, 2017 at 01:55:58PM +0900, Junio C Hamano wrote:

I left it intentionally vague exactly because I thought we might want to leave room for the answer to change to "yes" eventually. For instance, imagine that we had a ref storage format that required periodic compaction, and readers might sometimes choose to compact in order to save future readers from repeating some work they've done. If that compaction means holding a lock even for a brief period, I think it would fall under this option.

I admit that's just adding more hand-waving to the pile. But I don't think it really _hurts_ to leave that door open (aside from making the documentation a bit wishy-washy). And it helps because callers would pick up the new features automatically, without having to learn about a new option.

And I think that's really what this option is. It is less about the caller asking for some specific behavior, and more about them telling Git about the context in which it's executing so it can make intelligent decisions.

And in that sense, something descriptive like --background-process perhaps would be a better name. Except that I couldn't come up with a name that isn't confusing (certainly --background-process implies to me that Git would itself run in the background, which makes no sense here).

I also considered something like "--read-only" to tell Git that we should avoid writing to the repository. But that's really not what this does. It just avoids writes that may cause contention, not all writes.

I also considered using the word "opportunistic" in the option name, but decided it was too long and hard to spell.

So there. I am open to a better name, but I could not come up with one.

You lack his rugged good looks, but your review was still welcome.

-Peff

Jeff King <peff@peff.net> writes:

Oh, I do agree it is a good idea to leave that door open, so that scripts that rely on today's --no-optional-locks option will not have to be updated when a similar feature (like the "ref compaction" example you mentioned) against which "let's disable when we are not the primary process, in order to keep the interference to the minimum" would be something we would want to say. The option being added here should cover these future needs.

Yes, indeed.

Am 21.09.2017 um 06:32 schrieb Jeff King:

I don't think we should pass this environment variable to remote repositories. It should be listed in local_repo_env[] in environment.c.

-- Hannes

On Thu, Sep 21, 2017 at 08:25:50PM +0200, Johannes Sixt wrote:

I'm not sure I agree. This is really about the context in which the command is executing, not anything about the particular repository you're operating on.

For fetch/push operations that touch a remote, I doubt it would matter either way (and anyway, those often cross network boundaries that don't propagate environment variables anyway).

But imagine that "git status" learns to recurse into submodules and run "git status" inside them. Surely we would want the submodule repos to also avoid taking any unnecessary locks?

-Peff

On 9/22/2017 12:25 AM, Jeff King wrote:

https://github.com/git-for-windows/git/commit/ff63b51c22389139a864eb2e565c6cdc5a30f061

https://github.com/git-for-windows/git/pull/1004/commits/45bad66192352481acbc826f11d90c8928b39a7a

We should compare this with what we did in Git for Windows last fall. I guess those commits didn't get pushed upstream.

We added '--no-lock-index' to keep status from locking the index during status and effectively being read-only. This helped with problems with Visual Studio similar to the ones being described for KDevelop.

Jeff

On Fri, Sep 22, 2017 at 07:22:28AM -0400, Jeff Hostetler wrote:

Right. I think you missed the initial message in the thread that explains how this is an expanded version of ff63b51c22. :)

I didn't know about the environment thing in 45bad66192, though[1]. That makes me even more confident that this is the right approach.

-Peff

[1] Sorry for not doing my homework more carefully on the existing
    solution.  GitHub Desktop ran into the same situation and pointed me
    at ff63b51c22. I extrapolated the rest of it on my own. ;)

On Thu, Sep 21, 2017 at 9:25 PM, Jeff King <peff@peff.net> wrote:

You can teach Git to recurse into submodules already at home, just 'git config status.submoduleSummary none'. ;)

It occurs to me that the config name is badly choosen, as it stores an argument for git status --ignore-submodules[=mode]

On Fri, Sep 22, 2017 at 01:09:32PM -0700, Stefan Beller wrote:

Ah, thanks. I _thought_ we could already do that but when I went looking for the standard --recursive option I couldn't find it.

So yes, I would think we would want this option to apply recursively in that case, even when we cross repository boundaries.

-Peff

On Fri, Sep 22, 2017 at 2:25 PM, Jeff King <peff@peff.net> wrote:

Thanks for checking for submodules there.

I personally prefer --recurse-submodules despite the longer name, because a plain recursive option can mean anything in a sufficiently complex program such as Git (recurse into the tree (c.f. ls-tree), or for the algorithm used (c.f. merge, diff) or yet another dimension I did not think of).

Regarding the actual patch which is heavily aimed at coping with IDEs despite the command line being used, I wonder how many IDEs pass --ignore-submodules and recurse themselves (if needed). Reason for my suspicion is [1] which does pay attention to submodules:

[1] https://public-inbox.org/git/2bbb1d0f-ae06-1878-d185-112bd51f75c9@gmail.com/

There might be another option to cope with the situation:

 4. Teach all commands to spinlock / busywait shortly for important
     locks instead of giving up. In that case git-status rewriting
     the index ought to be fine?

Thanks, Stefan

On Fri, Sep 22, 2017 at 02:41:06PM -0700, Stefan Beller wrote:

Yeah, I agree that's a better name (mentioning --recursive is probably just showing my general ignorance of submodules).

Yeah, I think that's probably Visual Studio (which is what Johannes presumably wrote the original patch for). GitHub Desktop does not currently pass it, though perhaps should consider doing so (though of course the user could tweak their config, too).

We do have all the infrastructure in place to do a reasonable busywait with backoff. I think the patch is roughly just:

uh, too bad. The patch above looks really promising, though. :)

There is only one way to find out. Though we don't want to volunteer all users into this experiment, I'd presume.

Regarding larger indexes, I wonder if we can adapt the 500ms to the repo size. At first I thought the abbreviation length could be a good proxy to determine the maximum waiting time, but now I am not so sure any more.

Stefan

On Mon, Sep 25, 2017 at 11:51:31AM -0700, Stefan Beller wrote:

There are probably only a handful of other callers, and they'd just need to swap out s/update/&_timeout/. So it really is pretty trivial.

Yes. One of the nice things about the optional-locks approach is that it only affects callers who specify the option. And the general idea has gotten a year of testing in Visual Studio, which makes me feel good about it.

I think madness that way lies.

-Peff

On 09/20/2017 11:32 PM, Jeff King wrote:

Interestingly, this usually slaps me when performing an _interactive_ rebase.  It occurred to me that if I'm performing an interaction operation, it doesn't seem unreasonable for git wait up to 125ms or so for the lock and then prompting the user to ask if they want to continue waiting for the lock.

That is not necessarily the case.  I don't actually know git on the inside, but I would ask you to consider a read-write lock and a hybrid of one and three.

I don't know what dotlocks are, but I'm certain that you can implement a rw lock using lock files and no other IPC, although it does increase the complexity.  The way this works is that `git status' acquires a read lock and does its thing.  If it has real changes, instead of discarding them it attempts to upgrade to a write lock.  If that fails, you throw it away, otherwise you write them and release.

In order to implement rw locks with only lock files, "off the cuff" I say you have a single "lock list" file that should never be deleted and a "lock lock" file that is held in order to read or modify the list.  The format of the lock list would have a pair of 32-bit wrapping modification counts (or versions) at the top -- one for modifications to the lock list its self and another for modifications to the underlying data (i.e., the number of times a write lock has been acquired).  This header is followed by entries something like this:

<operation> <pid> <version> <timestamp>

<operation>  'r' if waiting for a read lock
             'R' if actively reading
             'w' if waiting for write lock
             'W' if actively writing
<pid>        The pid
<version>    If active, the version of the data at the time lock acquired or zero.
<timestamp>  Time began waiting or time lock acquired

An operation of 'r' or 'w' means that you are waiting and upper case means that you are active.  <version> is the version of the data at the time the lock became active and writers increment it when they acquire a lock.  You wait with file alteration notification on the lock list (if there is any doubt based upon timestamp precision then you can examine the lock list version).  When you want to read or write, you lock the lock-lock (this sounds like a joke... "Lock, lock.", "Who's there?", "Reader.", "Reader who?"....) and examine the lock list.  If it's empty, you add yourself as an active reader or writer with an upper case 'R' or 'W' and release the lock-lock.  If there are only readers, and you want to read, you add yourself as an active reader.  The version of the lock list is incremented every time it is modified.

Read-write locks need to be given a priority policy of either readers, writers, fifo or don't-care.  In this case that should probably be writers.  So if you want to read and there is an active or waiting writer, the prospective reader would either add themselves with an 'r' or fail if they don't want to wait.  If a process wants to write and there are active readers or writers, it adds its self to the list or fails as well.  When all active readers have exited, then which ever prospective writers gets the lock-lock first can make themselves the active writer.  When a process acquires a write lock, it increases the data version number.  If a reader lock tries to upgrade to a writer lock but the data version changed than it fails.

Is there not already a library somewhere that does this?  Either way, your current effort seems like a step in the right direction -- and thanks for that!

Daniel

On Fri, Sep 22, 2017 at 01:42:10AM -0500, Daniel Santos wrote:

Yes, lock timeouts would help with this situation, though not eliminate it entirely (and we've been adding some lock timeouts in a few places for related situations).

We generally avoid prompting, especially when a command can just be reissued. It sounds like "git rebase" gets into a funny state if it cannot grab the index lock. That's something that should be fixed. Even if it bails, you should be able to move forward with "rebase --continue" or similar.

By dotlocks, I just mean our strategy of creating O_EXCL "index.lock" files.

You're right that it can be done using two such locks, and communicating via the lockfile contents. So my "impossible" was overstating it. I do stand by my contention that it's much more complex than the existing scheme. :)

More importantly, though, it changes the locking contract completely between old and new versions (and between other implementations). There's probably only a small minority of users who might use two implementations to simultaneously access the same repository, but it is a use case we try to support. I think we'd have to require a repository-version bump to tell programs to use the new scheme.

-Peff

On Thursday 21 September 2017 10:02 AM, Jeff King wrote:

I might be missing something, shouldn't the IDEs be encouraged to use libgit2 instead? I thought it was meant for these use cases.

Note: I assume getting the status through libgit2 doesn't create an 
index.lock file.

So, if I get that correctly "git status --no-optional-locks" is a way to get the "current" status without updating the on disk index file?

The above sentence seems to be a little hard to interpret. How about,

    If set to `0`, Git will complete the requested operation without
    performing the optional sub-operations that require taking a lock.

--- Kaartic

Hi Kaartic,

On Sun, 24 Sep 2017, Kaartic Sivaraam wrote:

There are pros and cons. Visual Studio moved away from libgit2 e.g. to support SSH (back then, libgit2 did not support spawning processes, I have no idea whether that changed in the meantime).

Ciao, Johannes

On 9/25/2017 12:17 PM, Johannes Schindelin wrote:

There were other issues besides feature parity. The big one for VS was that it moved the git computations into a separate process and address space. You can't easily read/modify/write a 500MB .git/index file into memory (with however many copies of the data that that creates) in a 32-bit process.

Hi Peff,

On Thu, 21 Sep 2017, Jeff King wrote:

Okay, so this is trying to help me by upstreaming a patch from Git for Windows?

If so: thanks!

The changes, in particular the different semantics, will guarantee that I have to work on the consumer's side here, though, leaving me even less time for the Git mailing list, so I will need a lot more help with upstreaming patches.

Ciao, Dscho

On Mon, Sep 25, 2017 at 06:10:12PM +0200, Johannes Schindelin wrote:

Yes, in a round-about way. :)

I think you can get away with doing nothing for the time being. The two patches can co-exist in the GfW codebase[1], and people using the existing option can switch over at their leisure. Eventually you may decide to revert 67e5ce7f63.

-Peff

[1] Modulo some trivial conflict resolution when the two are merged.

This is an update of my --no-optional-locks patch. The only difference is that I updated the documentation to be a bit less clunky (Thanks, Kaartic).

For the other comments on v1, I decided not to make any changes:

  - JSixt suggested marking this as a local-repo variable. I think we
    really do want it to cross repo boundaries to handle submodules (and
    the GfW version does the same).

  - there was some discussion over the name. I didn't see other
    suggestions, and I didn't come up with anything better.

  - there was some discussion of alternatives, including reader/writer
    locking schemes and lock-retry timeouts for other callers. This
    approach has the benefit of simplicity and having been tested in the
    real world by GfW/VS.  I wouldn't mind if somebody wanted to explore
    lock retries, but I think we'd want to have this feature regardless.

So I'm happy with this version, though of course I'm open to more comments.

-- >8 --
Subject: git: add --no-optional-locks option

Some tools like IDEs or fancy editors may periodically run commands like "git status" in the background to keep track of the state of the repository. Some of these commands may refresh the index and write out the result in an opportunistic way: if they can get the index lock, then they update the on-disk index with any updates they find. And if not, then their in-core refresh is lost and just has to be recomputed by the next caller.

But taking the index lock may conflict with other operations in the repository. Especially ones that the user is doing themselves, which _aren't_ opportunistic. In other words, "git status" knows how to back off when somebody else is holding the lock, but other commands don't know that status would be happy to drop the lock if somebody else wanted it.

There are a couple possible solutions:

  1. Have some kind of "pseudo-lock" that allows other
     commands to tell status that they want the lock.

     This is likely to be complicated and error-prone to
     implement (and maybe even impossible with just
     dotlocks to work from, as it requires some
     inter-process communication).

  2. Avoid background runs of commands like "git status"
     that want to do opportunistic updates, preferring
     instead plumbing like diff-files, etc.

     This is awkward for a couple of reasons. One is that
     "status --porcelain" reports a lot more about the
     repository state than is available from individual
     plumbing commands. And two is that we actually _do_
     want to see the refreshed index. We just don't want to
     take a lock or write out the result. Whereas commands
     like diff-files expect us to refresh the index
     separately and write it to disk so that they can depend
     on the result. But that write is exactly what we're
     trying to avoid.

  3. Ask "status" not to lock or write the index.

     This is easy to implement. The big downside is that any
     work done in refreshing the index for such a call is
     lost when the process exits. So a background process
     may end up re-hashing a changed file multiple times
     until the user runs a command that does an index
     refresh themselves.

This patch implements the option 3. The idea (and the test) is largely stolen from a Git for Windows patch by Johannes Schindelin, 67e5ce7f63 (status: offer *not* to lock the index and update it, 2016-08-12). The twist here is that instead of making this an option to "git status", it becomes a "git" option and matching environment variable.

The reason there is two-fold:

  1. An environment variable is carried through to
     sub-processes. And whether an invocation is a
     background process or not should apply to the whole
     process tree. So you could do "git --no-optional-locks
     foo", and if "foo" is a script or alias that calls
     "status", you'll still get the effect.

  2. There may be other programs that want the same
     treatment.

     I've punted here on finding more callers to convert,
     since "status" is the obvious one to call as a repeated
     background job. But "git diff"'s opportunistic refresh
     of the index may be a good candidate.

The test is taken from 67e5ce7f63, and it's worth repeating Johannes's explanation:

  Note that the regression test added in this commit does
  not *really* verify that no index.lock file was written;
  that test is not possible in a portable way. Instead, we
  verify that .git/index is rewritten *only* when `git
  status` is run without `--no-optional-locks`.

Signed-off-by: Jeff King <peff@peff.net>
---
 Documentation/git.txt | 12 ++++++++++++
 builtin/commit.c      |  5 ++++-
 cache.h               |  6 ++++++
 environment.c         |  5 +++++
 git.c                 |  4 ++++
 t/t7508-status.sh     | 10 ++++++++++
 6 files changed, 41 insertions(+), 1 deletion(-)

Hi Peff,

On Wed, 27 Sep 2017, Jeff King wrote:

Indeed. We had it originally as a local-repo only variable, and you probably guessed that submodules made us rethink that approach.

We also weighed other options, and took this route for simplicity. It's possible to explain during an uninterrupted elevator ride what it does. Makes for less bugs.

And yes, we wanted to test it first before showing it to the Git mailing list ;-)

Ciao, Dscho

On Sun, Sep 24, 2017 at 01:08:41PM +0530, Kaartic Sivaraam wrote:

GitHub Desktop, at least, has actually moved _away_ from libgit2. I think there were a number of reasons. Some match what Dscho said regarding Visual Studio. But I think a main one is just that libgit2 doesn't quite match regular git for feature parity or pace of development. So you're stuck shelling out to regular Git half the time anyway, and now you get to handle dependencies on _two_ systems. :)

I don't know if that's the case or not.

It's actually "git --no-optional-locks status", since it's a git-wide option (it's just that "status" is the only one who respects it for now).

Yeah, my original is pretty clunky. I ended up with:

  If set to `0`, Git will complete any requested operation without
  performing any optional sub-operations that require taking a lock.

which swaps out "the" for "any".

Thanks.

-Peff

On Wed, 2017-09-27 at 02:40 -0400, Jeff King wrote:

Thanks for correcting the command. Now let me ask my (corrected) question again! So, if I get that correctly "git --no-optional- locks status" is a way to get the "current" status *without updating* the on disk index file?

--- Kaartic

On Wed, Sep 27, 2017 at 07:20:05PM +0530, Kaartic Sivaraam wrote:

Yes.

-Peff