quoted
quoted
quoted
quoted
"Junio" == Junio C Hamano [off-list ref] writes:
Junio> If input_buffer[ret] _were_ the last octet read, [the] loop would
Junio> be discarding that octet when [it] call[s] more than one xread()
Junio> to fill the buffer.
That did sem more likely, but I thought I'd throw out the possibility.
I just tried instrumenting fill().
In the first call to fill() during a fetch, xread() returns 4096. In
the second call to fill(), xread() returns 4095. In all subsequent
calls to fill(), xread() returns 4096 again.
So, after the second call to xread(), consumed_bytes & 0xFFF == 0xFFF.
It always follows the pattern:
xread() read 0x1000 from fd 0 at 0x80a0900
ret = 0x1000
consumed_bytes = 0
input_len = 0x1000
xread() read 0xFFF from fd 0 at 0x80a0900
ret = 0xFFF
consumed_bytes = 0x1000
input_len = 0xFFF
xread() read 0x1000 from fd 0 at 0x80a0900
ret = 0x1000
consumed_bytes = 0x1FFF
input_len = 0x1000
with all subsequent calls reading 0x1000 and adding 0x1000 to
consumed_bytes, until the final chuck of tha pack is read.
Also, all calls to xread() where the status lines are being sent from
the remote server also return 0x1000 octets. Only the second chunck of
a pack ever returns 0xFFF.
I've tested against a number of remote servers and the pattern holds for
a wide range of remote server versions.
The pattern also holds for clones over ssh.
Does anyone have an idea of why the second call to read(2), when
receiving a pack from a remote, always leaves the last octet of the
buffer free, whereas all other read(2)s fill it?
-JimC
--
James Cloos [off-list ref] OpenPGP: 1024D/ED7DAEA6
James Cloos schrieb:
quoted
quoted
quoted
quoted
quoted
"Junio" == Junio C Hamano [off-list ref] writes:
Junio> If input_buffer[ret] _were_ the last octet read, [the] loop would
Junio> be discarding that octet when [it] call[s] more than one xread()
Junio> to fill the buffer.
That did sem more likely, but I thought I'd throw out the possibility.
I just tried instrumenting fill().
In the first call to fill() during a fetch, xread() returns 4096. In
the second call to fill(), xread() returns 4095. In all subsequent
calls to fill(), xread() returns 4096 again.
So, after the second call to xread(), consumed_bytes & 0xFFF == 0xFFF.
It always follows the pattern:
xread() read 0x1000 from fd 0 at 0x80a0900
ret = 0x1000
consumed_bytes = 0
input_len = 0x1000
xread() read 0xFFF from fd 0 at 0x80a0900
ret = 0xFFF
consumed_bytes = 0x1000
input_len = 0xFFF
xread() read 0x1000 from fd 0 at 0x80a0900
ret = 0x1000
consumed_bytes = 0x1FFF
input_len = 0x1000
with all subsequent calls reading 0x1000 and adding 0x1000 to
consumed_bytes, until the final chuck of tha pack is read.
Also, all calls to xread() where the status lines are being sent from
the remote server also return 0x1000 octets. Only the second chunck of
a pack ever returns 0xFFF.
I've tested against a number of remote servers and the pattern holds for
a wide range of remote server versions.
The pattern also holds for clones over ssh.
Does anyone have an idea of why the second call to read(2), when
receiving a pack from a remote, always leaves the last octet of the
buffer free, whereas all other read(2)s fill it?
You could instrument upload-pack.c. There is an xread() around l.237; does
it aways return 4096? And send_client_data() at around l.254; when does it
send 4096 and when 4095 bytes? Read the comment in this if-branch.
upload-pack is run on the server-side; therefore, you can test this only
on local repositories (unless you can replace upload-pack on the server, too).
-- Hannes