Re: [PATCH v2 1/2] sha1_file: introduce close_one_pack() to close packs on fd pressure

From: Brandon Casey <drafnel@gmail.com>
Date: Aug 1, 2013, 19:16 UTC
Message-ID: <CA+sFfMer+5bhqxF=_zAQhZ+8sQD6EAYb8HBtYpuhQY_0uj-m9A@mail.gmail.com>
In-Reply-To: <7v4nb9s1az.fsf@alter.siamese.dyndns.org>

On Thu, Aug 1, 2013 at 11:39 AM, Junio C Hamano <gitster@pobox.com> wrote:

Show 16 quoted lines

> Brandon Casey <drafnel@gmail.com> writes:
>
>> I've been looking closer at uses of p->windows everywhere, and it
>> seems that we always open_packed_git() before we try to create new
>> windows.  There doesn't seem to be any reason that we can't continue
>> to use the existing open windows even after closing the pack file.
>> ...
>> If we don't need to close_pack_windows(), find_lru_pack() doesn't
>> strictly need to reject packs that have windows in use.
>
> That makes me feel somewhat uneasy.  Yes, you can open/mmap/close
> and hold onto the contents of a file still mapped in-core, and it
> may not count as "open filedescriptor", but do OSes allow infinite
> such mmapped regions to us?  We do keep track of number of open
> windows, but is there a way for us to learn how close we are to the
> limit?

Not that I know of, but xmmap() does already try to unmap existing windows when mmap() fails, and then retries the mmap. It calls release_pack_memory() which calls unuse_one_window(). mmap returns ENOMEM when either there is no available memory or if the limit of mmap mappings has been exceeded.

So, I think we'll be ok. It's the same situation we'd be in if there were many large packs (but fewer than pack_max_fds) and a small packedGitWindowSize, requiring many mmap windows. We'd try to map an additional segment, fail, release some unused segments, and retry.

The memory usage of all mmap segments would still be bounded by packedGitLimit. It's just that now, when we're only under file descriptor pressure, we won't close the mmap windows unnecessarily when they may be needed again.

-Brandon