RE: [PATCH v1 1/4] builtin/index-pack.c: change xwrite to write_in_full to allow large sizes.

From: rsbecker@nexbridge.com <rsbecker@nexbridge.com>
Date: Feb 27, 2024, 00:12 UTC
Message-ID: <027001da6911$a727b9d0$f5772d70$@nexbridge.com>
In-Reply-To: <xmqqa5nmkcuz.fsf@gitster.g>

On Monday, February 26, 2024 6:47 PM, Junio C Hamano wrote:

Show 14 quoted lines

><rsbecker@nexbridge.com> writes:
>
>>>The code above loops while input_len is non-zero, and correctly
>>>decrements it by the number of bytes written by xwrite() after each
>>>iteration.
>>>
>>>Assuming that xwrite()/write(2) works how I think it does on NonStop,
>>>I'm not sure I understand why this change is necessary.
>>
>> NonStop has a limited SSIZE_MAX. xwrite only handles that much so
>> anything beyond that gets dropped (not in the above code but in other
>> builtins)
>
>xwrite() caps a single write attempt to MAX_IO_SIZE and can return a

short-write, so anything beyound MAX_IO_SIZE will not even be

>sent to the underlying write(2).  There is a heuristic based on the value

of SSIZE_MAX to define MAX_IO_SIZE in <git-compat-util.h>,

>and if the value given by that heuristics is too large for your platform,

you can tweak your own MAX_IO_SIZE (see the comments in

>that header file).
>
>The caller of xwrite() must be prepared to see a write return with value

less than the length it used to call the function, either because

>of this MAX_IO_SIZE cut-off, or because of the underlying
>write(2) returning after a short write.  As long as the caller is prepared,

like Taylor pointed out, I am not sure why you'd need to change

>it.

I understand. I was involved in xwrite() a few years ago. The problem is that users of xwrite() did not account for that and t7704.9 failed as a result. These changes did fix the issue. I am not sure how to proceed based on the above, however. Continue or recode the callers (which is part of what this does)?