[PATCH 0/2] packfile: small syscall reductions

[PATCH 0/2] packfile: small syscall reductions

[Eric Wong]

Being an avid user of strace, redundant syscalls clutter up the
output and bother me, so I removed some.  Folks with CPU
mitigations enabled might also benefit from fewer syscalls.

Eric Wong (2):
  packfile: remove redundant fcntl F_GETFD/F_SETFD
  packfile: replace lseek+read with pread

 packfile.c | 16 ++--------------
 1 file changed, 2 insertions(+), 14 deletions(-)

[PATCH 1/2] packfile: remove redundant fcntl F_GETFD/F_SETFD

[Eric Wong]

git_open sets close-on-exec since cd66ada06588f797
("sha1_file: open window into packfiles with O_CLOEXEC").
There's no reason to keep using fcntl to set the close-on-exec
flag, anymore.

Signed-off-by: Eric Wong <e@80x24.org>
---
 packfile.c | 11 -----------
 1 file changed, 11 deletions(-)

diff --git a/packfile.c b/packfile.c
index f0dc63e92f..1821cb7a3d 100644
--- a/packfile.c
+++ b/packfile.c
@@ -510,7 +510,6 @@ static int open_packed_git_1(struct packed_git *p)
 	struct pack_header hdr;
 	unsigned char hash[GIT_MAX_RAWSZ];
 	unsigned char *idx_hash;
-	long fd_flag;
 	ssize_t read_result;
 	const unsigned hashsz = the_hash_algo->rawsz;
 
@@ -554,16 +553,6 @@ static int open_packed_git_1(struct packed_git *p)
 	} else if (p->pack_size != st.st_size)
 		return error("packfile %s size changed", p->pack_name);
 
-	/* We leave these file descriptors open with sliding mmap;
-	 * there is no point keeping them open across exec(), though.
-	 */
-	fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
-	if (fd_flag < 0)
-		return error("cannot determine file descriptor flags");
-	fd_flag |= FD_CLOEXEC;
-	if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
-		return error("cannot set FD_CLOEXEC");
-
 	/* Verify we recognize this pack file format. */
 	read_result = read_in_full(p->pack_fd, &hdr, sizeof(hdr));
 	if (read_result < 0)

[PATCH 2/2] packfile: replace lseek+read with pread

[Eric Wong]

We already have pread emulation for portability, so there's
there's no reason to make two syscalls where one suffices.

Furthermore, readers of the packfile will be using mmap
(or pread to emulate mmap), anyways, so the file description
offset does not matter in this case.

Signed-off-by: Eric Wong <e@80x24.org>
---
 packfile.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/packfile.c b/packfile.c
index 1821cb7a3d..7e7c04e4d8 100644
--- a/packfile.c
+++ b/packfile.c
@@ -576,9 +576,8 @@ static int open_packed_git_1(struct packed_git *p)
 			     " while index indicates %"PRIu32" objects",
 			     p->pack_name, ntohl(hdr.hdr_entries),
 			     p->num_objects);
-	if (lseek(p->pack_fd, p->pack_size - hashsz, SEEK_SET) == -1)
-		return error("end of packfile %s is unavailable", p->pack_name);
-	read_result = read_in_full(p->pack_fd, hash, hashsz);
+	read_result = pread_in_full(p->pack_fd, hash, hashsz,
+					p->pack_size - hashsz);
 	if (read_result < 0)
 		return error_errno("error reading from %s", p->pack_name);
 	if (read_result != hashsz)

Re: [PATCH 2/2] packfile: replace lseek+read with pread

[Junio C Hamano]

Eric Wong <e@80x24.org> writes:

> We already have pread emulation for portability, so there's
> there's no reason to make two syscalls where one suffices.
>
> Furthermore, readers of the packfile will be using mmap
> (or pread to emulate mmap), anyways, so the file description
> offset does not matter in this case.

s/description/descriptor/ probably.

After seeking to the packfile trailer and reading the pack id hash
using lseek+read, this helper function does not read from the file
descriptor, and the sole caller of it closes the file descriptor
immediately after it returns, which means the read file offset after
reading the packfile trailer does not matter.

So this conversion is correct.  Thanks for a careful analysis.

Will queue both patches.

> Signed-off-by: Eric Wong <e@80x24.org>
> ---
>  packfile.c | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/packfile.c b/packfile.c
> index 1821cb7a3d..7e7c04e4d8 100644
> --- a/packfile.c
> +++ b/packfile.c
> @@ -576,9 +576,8 @@ static int open_packed_git_1(struct packed_git *p)
>  			     " while index indicates %"PRIu32" objects",
>  			     p->pack_name, ntohl(hdr.hdr_entries),
>  			     p->num_objects);
> -	if (lseek(p->pack_fd, p->pack_size - hashsz, SEEK_SET) == -1)
> -		return error("end of packfile %s is unavailable", p->pack_name);
> -	read_result = read_in_full(p->pack_fd, hash, hashsz);
> +	read_result = pread_in_full(p->pack_fd, hash, hashsz,
> +					p->pack_size - hashsz);
>  	if (read_result < 0)
>  		return error_errno("error reading from %s", p->pack_name);
>  	if (read_result != hashsz)

Re: [PATCH 2/2] packfile: replace lseek+read with pread

[Junio C Hamano]

Junio C Hamano <gitster@pobox.com> writes:

> Eric Wong <e@80x24.org> writes:
>
>> We already have pread emulation for portability, so there's
>> there's no reason to make two syscalls where one suffices.
>>
>> Furthermore, readers of the packfile will be using mmap
>> (or pread to emulate mmap), anyways, so the file description
>> offset does not matter in this case.
>
> s/description/descriptor/ probably.
>
> After seeking to the packfile trailer and reading the pack id hash
> using lseek+read, this helper function does not read from the file
> descriptor, and the sole caller of it closes the file descriptor
> immediately after it returns, which means the read file offset after
> reading the packfile trailer does not matter.

Oops, that was not right.  When we successfully open the packfile,
we leave the file descriptor open, so we do need the "we never read
using read(2) from the file descriptor" guarantee for this change to
be correct.

But we do have the guarantee, and existing code does depend on the
guarantee, so the patch is good.

Thanks.

Re: [PATCH 2/2] packfile: replace lseek+read with pread

[Eric Wong]

Junio C Hamano <gitster@pobox.com> wrote:
> Eric Wong <e@80x24.org> writes:
> >
> > Furthermore, readers of the packfile will be using mmap
> > (or pread to emulate mmap), anyways, so the file description
> > offset does not matter in this case.
> 
> s/description/descriptor/ probably.

No, I meant "description" :)  The offset is shared in case of dup{,2,3}
syscalls, which only creates a new descriptor, not a new description.
Both the Linux and POSIX lseek(2) manpages say "description".