patchosxkeychain: lock for exclusive execution

From: Koji Nakamaru <koji.nakamaru@gree.net>

Resolves "failed to store: -25299" when "fetch.parallel 0" is configured and there are many submodules.

The error code -25299 (errSecDuplicateItem) may be returned by SecItemUpdate() in add_internet_password() if multiple instances of git-credential-osxkeychain run in parallel. This patch introduces an exclusive lock to serialize execution for avoiding this and other potential issues.

Signed-off-by: Koji Nakamaru <koji.nakamaru@gree.net>
---
    osxkeychain: lock for exclusive execution

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1729%2FKojiNakamaru%2Ffeature%2Fosxkeychian_exlock-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1729/KojiNakamaru/feature/osxkeychian_exlock-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1729

 contrib/credential/osxkeychain/git-credential-osxkeychain.c | 3 +++
 1 file changed, 3 insertions(+)

Interesting.

SecItemUpdate returning errSecDuplicateItem didn’t make sense to make sense to me so I had a check to see what scenario this happens and it appears to be a scenario where updating in-place fails but replacing it entirely succeeds. However it seems the item might have ultimately still been updated: https://github.com/apple-oss-distributions/Security/blob/0600e7bab30fbac3adcafcb6c57d3981dc682304/OSX/libsecurity_keychain/lib/SecItem.cpp#L2398

The behaviour is a bit odd and the associated code comment referencing an Apple bug number is perhaps is indicative of that. I guess it perhaps makes sense if you are holding references, but that doesn’t apply to us.

I wonder if a fix here could be to treat errSecDuplicateItem as a successful operation for SecItemUpdate. Can you confirm the keychain item is successfully updated in that scenario?

A broader Git-wide question that you perhaps don’t know the answer to but someone else here might do is: why are we spamming updates to the credential helper? Every parallel fetch instance performing a store operation on the same host seems unexpected to me, particularly if there’s no actual changes.

Bo

On Fri, May 10, 2024 at 04:02:03PM +0100, Bo Anderson wrote:

The short answer is that Git always passes a credential which has been used successfully to the helpers to record (if they want to). That's how stuff gets stored in the first place. And those parallel fetches have no knowledge of what the other ones are doing, so they all try to store.

But the more interesting question is: why do we tell helpers to store a credential that we got from helpers in the first place? The behavior is mostly an artifact of how the original implementation behaved, as it did not record the source of the credential.

And I think there are several problems with that, besides inefficiency and locking. See this old patch, which fixes it by remembering when a credential came from a helper:

  https://lore.kernel.org/git/20120407033417.GA13914@sigill.intra.peff.net/

But we didn't merge it because some people rely on the behavior of helpers feeding back to themselves. I outlined some solutions there, but it would definitely be a change in behavior that people would have to adapt to.

Some possible alternatives:

  - we could remember _which_ helper we got the credential from, and
    avoid invoking it again.

  - we could record a bit saying that the credential came from a helper,
    and then feed that back to helpers when storing. So osxkeychain
    could then decide not to store it.

Both of those solve the repeated stores, but still let credentials populate across helpers (which I still think is a questionable thing to do by default, per the discussion in that thread, but is the very thing that some people rely on).

-Peff

On 2024-05-10 at 20:01:14, Jeff King wrote:

This will break the new `state[]` feature, which relies on being able to see the state after the fact to know whether the operation was successful. As an example of the functionality the current approach allows, authentication could use an HOTP (like TOTP, but using a counter instead of time) value, and storing the correct used counter on success would be important.

I agree it's not super important if we're just using a username and password, but considering I just added support for arbitrary authentication schemes, which can include things such as limited-use OAuth tokens, one-time use passcodes, and certain types of HMAC-based signing, we probably don't want to choose this approach.

This is actually possible with the new `state[]` feature. `osxkeychain` can simply set that field to something like `osxkeychain:seen=1` and simply do nothing if it sees that field.

All the credential helper needs to do is declare support for that functionality with the appropriate capability and emit the field if it gets that capability on standard input.

On Fri, May 10, 2024 at 08:33:08PM +0000, brian m. carlson wrote:

Yeah, I think it makes sense to keep the Git side as general as possible. So invoking the helper but giving it extra information (so it can decide whether to be a noop or not) seems like the better approach.

Makes sense. Back in that old thread I showed a patch which would let helpers pass arbitrary fields to each other (or back to themselves), and this works in roughly the same way.

If I understand the protocol, it is just:

  printf("capability[]=state\n");
  printf("state[]=osxkeychain:seen=1\n");

in the helper when it returns a username/password? And I guess the matching parse/check on "store".

Sounds like that would be easy for folks on macOS to play with.

-Peff

On 2024-05-10 at 22:07:15, Jeff King wrote:

Yup. It may receive `state[]` fields from other helpers, so it needs to check that the entries are its own (presumably starting with `osxkeychain:`) when it reads them, but otherwise, that's it.

Jeff King <peff@peff.net> writes:

Would "refreshing the last-time-used record" a valid use case for the behaviour that feeds the successful one back to where the credential came from? Such a helper could instead log the last-time the credential was asked for, and assume that the lack of an explicit "reject" call signals that the use of the value it returned earlier was auccessfully used, but it is a less obvious way to implement such a "this hasn't been successfully used for a long time, perhaps we should expire/ask again/do something else?" logic.

On Fri, May 10, 2024 at 01:40:29PM -0700, Junio C Hamano wrote:

There was some discussion in that old thread about whether that was important or not. I don't have a strong opinion there. Not refreshing is a more secure default, but possibly more annoying (and a change from the status quo).

I do think brian's suggestion to use state[] to pass it back means that the decision is then in the hands of the helper. So "credential-cache", for example, could decide whether to refresh its ttl or not, or we could even make it configurable with a command-line option for the helper.

-Peff

Jeff King <peff@peff.net> writes:

Yeah, I read your discussion with brian, and the state[] thing all made sense to me.

Thanks.

Koji Nakamaru (2):
  osxkeychain: lock for exclusive execution
  osxkeychain: state[] seen=1 to skip unnecessary store operations

 .../osxkeychain/git-credential-osxkeychain.c       | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

base-commit: 0f3415f1f8478b05e64db11eb8aaa2915e48fef6
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1729%2FKojiNakamaru%2Ffeature%2Fosxkeychian_exlock-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1729/KojiNakamaru/feature/osxkeychian_exlock-v2
Pull-Request: https://github.com/gitgitgadget/git/pull/1729

Range-diff vs v1:

 1:  309c17c78f3 = 1:  309c17c78f3 osxkeychain: lock for exclusive execution
 -:  ----------- > 2:  1f57718abff osxkeychain: state[] seen=1 to skip unnecessary store operations

From: Koji Nakamaru <koji.nakamaru@gree.net>

Resolves "failed to store: -25299" when "fetch.parallel 0" is configured and there are many submodules.

The error code -25299 (errSecDuplicateItem) may be returned by SecItemUpdate() in add_internet_password() if multiple instances of git-credential-osxkeychain run in parallel. This patch introduces an exclusive lock to serialize execution for avoiding this and other potential issues.

Signed-off-by: Koji Nakamaru <koji.nakamaru@gree.net>
---
 contrib/credential/osxkeychain/git-credential-osxkeychain.c | 3 +++
 1 file changed, 3 insertions(+)

"Koji Nakamaru via GitGitGadget" <gitgitgadget@gmail.com> writes:

Use of third-person singular without subject for the "observation" part is highly unusual the log messages in our codebase.

The usual way to compose a log message of this project is to

 - Give an observation on how the current system work in the present
   tense (so no need to say "Currently X is Y", just "X is Y"), and
   discuss what you perceive as a problem in it.

 - Propose a solution (optional---often, problem description
   trivially leads to an obvious solution in reader's minds).

 - Give commands to the codebase to "become like so".

in this order.

"This patch introduces" -> "Introduce"

Is this step still needed, though?

Thank you for the instruction about a log message. I'll follow it.

For solving the issue I originally had, just utilizing state[] to skip "store" operations is enough.

Since the osxkeychain implementation doesn't seem to be aware that it can run in parallel, I thought it would be better to leave this step in case a similar problem occurs. If this reason is weak, I'll remove this step.

Koji Nakamaru

From: Koji Nakamaru <koji.nakamaru@gree.net>

Records whether credentials come from get operations and skips unnecessary store operations by utilizing the state[] feature, as suggested by brian m. carlson.

Signed-off-by: Koji Nakamaru <koji.nakamaru@gree.net>
---
 .../osxkeychain/git-credential-osxkeychain.c          | 11 +++++++++++
 1 file changed, 11 insertions(+)

"Koji Nakamaru via GitGitGadget" <gitgitgadget@gmail.com> writes:

This step has a problem description that is even sketchier than the previous one. Anticipate questions by the other developers who read this commit 6 months after it is accepted in the mainline (e.g., What problem is there in the current system, why it is bad and worth solving, and how is the patch trying to solve it?) and let your proposed log message answer the questions, as you won't be always sitting next to these developers.

Thanks.

Thank you. I'll clean up the whole patch and its description after getting a final reply about exlock.

Koji Nakamaru

Koji Nakamaru (2):
  osxkeychain: exclusive lock to serialize execution of operations
  osxkeychain: state to skip unnecessary store operations

 .../osxkeychain/git-credential-osxkeychain.c       | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

base-commit: 83f1add914c6b4682de1e944ec0d1ac043d53d78
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1729%2FKojiNakamaru%2Ffeature%2Fosxkeychian_exlock-v3
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1729/KojiNakamaru/feature/osxkeychian_exlock-v3
Pull-Request: https://github.com/gitgitgadget/git/pull/1729

Range-diff vs v2:

 1:  309c17c78f3 ! 1:  3341346c5e6 osxkeychain: lock for exclusive execution
     @@ Metadata
      Author: Koji Nakamaru <koji.nakamaru@gree.net>
      
       ## Commit message ##
     -    osxkeychain: lock for exclusive execution
     +    osxkeychain: exclusive lock to serialize execution of operations
      
     -    Resolves "failed to store: -25299" when "fetch.parallel 0" is configured
     -    and there are many submodules.
     +    git passes a credential that has been used successfully to the helpers
     +    to record. If "git-credential-osxkeychain store" commands run in
     +    parallel (with fetch.parallel configuration and/or by running multiple
     +    git commands simultaneously), some of them may exit with the error
     +    "failed to store: -25299". This is because SecItemUpdate() in
     +    add_internet_password() may return errSecDuplicateItem (-25299) in this
     +    situation. Apple's documentation [1] also states as below:
      
     -    The error code -25299 (errSecDuplicateItem) may be returned by
     -    SecItemUpdate() in add_internet_password() if multiple instances of
     -    git-credential-osxkeychain run in parallel. This patch introduces an
     -    exclusive lock to serialize execution for avoiding this and other
     -    potential issues.
     +      In macOS, some of the functions of this API block while waiting for
     +      input from the user (for example, when the user is asked to unlock a
     +      keychain or give permission to change trust settings). In general, it
     +      is safe to use this API in threads other than your main thread, but
     +      avoid calling the functions from multiple operations, work queues, or
     +      threads concurrently. Instead, serialize function calls or confine
     +      them to a single thread.
     +
     +    The error has not been noticed before, because the former implementation
     +    ignored the error.
     +
     +    Introduce an exclusive lock to serialize execution of operations.
     +
     +    [1] https://developer.apple.com/documentation/security/certificate_key_and_trust_services/working_with_concurrency
      
          Signed-off-by: Koji Nakamaru <koji.nakamaru@gree.net>
      
 2:  1f57718abff ! 2:  146b0ae9146 osxkeychain: state[] seen=1 to skip unnecessary store operations
     @@ Metadata
      Author: Koji Nakamaru <koji.nakamaru@gree.net>
      
       ## Commit message ##
     -    osxkeychain: state[] seen=1 to skip unnecessary store operations
     +    osxkeychain: state to skip unnecessary store operations
      
     -    Records whether credentials come from get operations and skips
     -    unnecessary store operations by utilizing the state[] feature, as
     -    suggested by brian m. carlson.
     +    git passes a credential that has been used successfully to the helpers
     +    to record. If a credential is already stored,
     +    "git-credential-osxkeychain store" just records the credential returned
     +    by "git-credential-osxkeychain get", and unnecessary (sometimes
     +    problematic) SecItemAdd() and/or SecItemUpdate() are performed.
      
     +    We can skip such unnecessary operations by marking a credential returned
     +    by "git-credential-osxkeychain get". This marking can be done by
     +    utilizing the "state[]" feature:
     +
     +    - The "get" command sets the field "state[]=osxkeychain:seen=1".
     +
     +    - The "store" command skips its actual operation if the field
     +      "state[]=osxkeychain:seen=1" exists.
     +
     +    Introduce a new state "state[]=osxkeychain:seen=1".
     +
     +    Suggested-by: brian m. carlson <sandals@crustytoothpaste.net>
          Signed-off-by: Koji Nakamaru <koji.nakamaru@gree.net>
      
       ## contrib/credential/osxkeychain/git-credential-osxkeychain.c ##

From: Koji Nakamaru <koji.nakamaru@gree.net>

git passes a credential that has been used successfully to the helpers to record. If "git-credential-osxkeychain store" commands run in parallel (with fetch.parallel configuration and/or by running multiple git commands simultaneously), some of them may exit with the error "failed to store: -25299". This is because SecItemUpdate() in add_internet_password() may return errSecDuplicateItem (-25299) in this situation. Apple's documentation [1] also states as below:

  In macOS, some of the functions of this API block while waiting for
  input from the user (for example, when the user is asked to unlock a
  keychain or give permission to change trust settings). In general, it
  is safe to use this API in threads other than your main thread, but
  avoid calling the functions from multiple operations, work queues, or
  threads concurrently. Instead, serialize function calls or confine
  them to a single thread.

The error has not been noticed before, because the former implementation ignored the error.

Introduce an exclusive lock to serialize execution of operations.

[1] https://developer.apple.com/documentation/security/certificate_key_and_trust_services/working_with_concurrency

Signed-off-by: Koji Nakamaru <koji.nakamaru@gree.net>
---
 contrib/credential/osxkeychain/git-credential-osxkeychain.c | 3 +++
 1 file changed, 3 insertions(+)

From: Koji Nakamaru <koji.nakamaru@gree.net>

git passes a credential that has been used successfully to the helpers to record. If a credential is already stored, "git-credential-osxkeychain store" just records the credential returned by "git-credential-osxkeychain get", and unnecessary (sometimes problematic) SecItemAdd() and/or SecItemUpdate() are performed.

We can skip such unnecessary operations by marking a credential returned by "git-credential-osxkeychain get". This marking can be done by utilizing the "state[]" feature:

- The "get" command sets the field "state[]=osxkeychain:seen=1".

- The "store" command skips its actual operation if the field
  "state[]=osxkeychain:seen=1" exists.

Introduce a new state "state[]=osxkeychain:seen=1".

Suggested-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Koji Nakamaru <koji.nakamaru@gree.net>
---
 .../osxkeychain/git-credential-osxkeychain.c          | 11 +++++++++++
 1 file changed, 11 insertions(+)

I thought about the issue further. The approach with the "state[]" feature reduces problematic "store" operations a lot in most cases, it is however not perfect to avoid the original issue about parallel execution.

For example (though quite artificial), let's suppose if two git commands start simultaneously where the credential is input automatically by "expect" command. Two "osxkeychain store" commands will then run simultaneously.

I hope new descriptions are convincing.

Koji Nakamaru

Thank you for detailed insights.

I tested osxkeychain with the modification at the end of this note and got the following log for "git fetch --all --prune --recurse-submodules". SecItemUpdate() sometimes returns errSecDuplicateItem but the keychain item seems okay -- its value is correct after the command finished -- perhaps because one of successful operations stores the correct value. Even if every store operation fails, perhaps the originally stored value is kept and no damage occurs.

  XXX: get
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: wwwauth[]=Basic realm="GitHub"
  XXX: store
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: username=jenkins
  XXXX: password=ghq_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
  XXXX: -25299
  XXXXX: 0
  XXX: get
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: wwwauth[]=Basic realm="GitHub"
  XXX: store
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: username=jenkins
  XXXX: password=ghq_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
  XXXX: -25299
  XXXXX: 0
  XXX: get
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: wwwauth[]=Basic realm="GitHub"
  XXX: store
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: username=jenkins
  XXXX: password=ghq_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
  XXXX: -25299
  XXXXX: 0
  XXX: get
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: wwwauth[]=Basic realm="GitHub"
  XXX: store
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: username=jenkins
  XXXX: password=ghq_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
  XXXX: -25299
  XXXXX: -25299
  failed to store: -25299
  XXX: get
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: wwwauth[]=Basic realm="GitHub"
  XXX: store
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: username=jenkins
  XXXX: password=ghq_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
  XXXX: -25299
  XXXXX: 0
  XXX: get
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: wwwauth[]=Basic realm="GitHub"
  XXX: store
  XXXX: protocol=https
  XXXX: host=github.com
  XXXX: username=jenkins
  XXXX: password=ghq_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
  XXXX: -25299
  XXXXX: -25299
  failed to store: -25299
  ...

This issue however occurs only when fetch.parallel is configured. If fetch.parallel is not configured, we should not ignore any error (including errSecDuplicateItem). Also, the above unstable behaviour is essentially caused by running osxkeychain instances in parallel where some of them treat "get" and others treat "store". I've also considered treating errSecDuplicateItem of SecItemUpdate() as errSecSuccess, but ended up with the current patch for these reasons.

I agree on this point and would like to know the reason.

Koji Nakamaru