[PATCH 3 of 7] vfs: use checkambigatclosing in checkambig=True but atomictemp=False case

[Pierre-Yves David]

On 09/21/2016 10:04 AM, FUJIWARA Katsunori wrote:
> At Wed, 21 Sep 2016 15:20:39 +0900,
> FUJIWARA Katsunori wrote:
>>
>> At Tue, 20 Sep 2016 18:12:58 +0200,
>> Pierre-Yves David wrote:
>>>
>>> On 09/16/2016 10:51 PM, FUJIWARA Katsunori wrote:
>>>> # HG changeset patch
>>>> # User FUJIWARA Katsunori <foozy at lares.dti.ne.jp>
>>>> # Date 1474057495 -32400
>>>> #      Sat Sep 17 05:24:55 2016 +0900
>>>> # Node ID 71b6b49f8a7ab6c894028b9153290f4bbf0f54f6
>>>> # Parent  ad999fb789fcb86b11c98334ab98b31a17ee2d25
>>>> vfs: use checkambigatclosing in checkambig=True but atomictemp=False case
>>>>
>>>> In Mercurial source tree, opening a file in "a"/"a+" mode like below
>>>> doesn't specify atomictemp=True for vfs, and this avoids file stat
>>>> ambiguity check by atomictempfile.
>>>>
>>>>   - writing changes out in revlog layer uses "a+" mode
>>>>   - truncation in repair.strip() uses "a" mode
>>>>   - truncation in transaction._playback() uses "a" mode
>>>>
>>>> If steps below occurs at "the same time in sec", all of mtime, ctime
>>>> and size are same between (1) and (3).
>>>>
>>>>   1. append data to revlog-style file (and close transaction)
>>>>   2. discard appended data by truncation (strip or rollback)
>>>>   3. append same size but different data to revlog-style file again
>>>>
>>>> Therefore, cache validation doesn't work after (3) as expected.
>>>>
>>>> This patch uses checkambigatclosing in checkambig=True but
>>>> atomictemp=False case, to check (and get rid of) file stat ambiguity
>>>> at closing.
>>>>
>>>> This is a part of ExactCacheValidationPlan.
>>>>
>>>>     https://www.mercurial-scm.org/wiki/ExactCacheValidationPlan
>>>>
>>>> diff --git a/mercurial/scmutil.py b/mercurial/scmutil.py
>>>> --- a/mercurial/scmutil.py
>>>> +++ b/mercurial/scmutil.py
>>>> @@ -587,6 +587,10 @@ class vfs(abstractvfs):
>>>>          if nlink == 0:
>>>>              self._fixfilemode(f)
>>>>
>>>> +        if checkambig:
>>>> +            assert mode not in ('r', 'rb'), "valid only at writing"
>>>> +            fp = checkambigatclosing(fp)
>>>
>>> It sound a bit too much like a real logic check with assert. Instead we
>>> should probably either:
>>> - have a hard check with an abort.
>>> - ignore the bad state with a devel warning (probably the best).
>>
>> OK, I'll send revised one with the latter.
>
> Oops, I forgot that vfs doesn't have any reference path to ui object :-<
>
> Should we raise Abort('implementation error: mode %s is not valid for
> checkambig=True') or so ? (or ignore bad state silently ?)

We could either carry some function to issue this warning, but this 
seems a bit cumbersome :-/

I'm a bit afraid of raising an Abort as is might eventually backfire on 
some innocent user. However, silently ignoring it means developer could 
misuse the API without ever realizing. So I guess we should go with the 
raise Abort.

Cheers,

-- 
Pierre-Yves David