homepage

In attachment.
This is actually just an excuse to store the patch somewhere and possibly collect opinions as I don't really think this should go in because:
- it's UNIX only
- as such, deciding whether using sendfile() should probably be done silently (no explicit argument)
- on the other hand, I don't think it's safe to decide this automatically because:
 - the input fd should be a regular file and it's not clear how to determine this beforehand 
 - in case of disconnection OSError is raised instead of socket.error (minor backward compatibility issue)

> deciding whether using sendfile() should probably be done silently (no explicit argument)
As an optimization taking advantage from OS support, I think this should be automatic too. But if there are too many issues, then explicit argument sounds better.
> the input fd should be a regular file and it's not clear how to determine this beforehand
Calling some function like os.stat that only works with real files? (not sure os.stat is the right function, just giving an idea)
> in case of disconnection OSError is raised instead of socket.error
PEP 3151!
>>> socket.error, OSError, IOError
(<class 'OSError'>, <class 'OSError'>, <class 'OSError'>)
:)

os.fstat wouldn't work since it succeeds with non-"regular" files, e.g. standard I/O:
>>> os.fstat(0)
posix.stat_result(st_mode=8592, st_ino=5, st_dev=11, st_nlink=1, st_uid=500, st_gid=5, st_size=0, st_atime=1323629303, st_mtime=1323629303, st_ctime=1323628616)
I think the best solution is to call sendfile() and catch OSError, then fallback on the generic loop. However, you must also guard against fileno() failing:
>>> io.BytesIO().fileno()
Traceback (most recent call last):
 File "<stdin>", line 1, in <module>
io.UnsupportedOperation: fileno

Here's the result of a benchmark sending a 1GB file over a Gb/s ethernet network:
vanilla:
real 0m9.446s
user 0m0.493s
sys 0m1.425s
sendfile:
real 0m9.143s
user 0m0.055s
sys 0m0.986s
The total time doesn't vary much (the reduction above is just jitter).
But it reduces user+sys time by almost a factor of 2.
Note that is changed Giampaolo's patch to call sendfile on the whole file, not by block.

> Note that is changed Giampaolo's patch to call sendfile 
> on the whole file, not by block.
That's not compatible across POSIX platforms.

> That's not compatible across POSIX platforms.
What do you mean ?
I juste call fstat() before calling senfile() to find out the file
size, and pass it as `count`.

Ah ok, I misinterpreted what you wrote then.
Generally speaking though, you don't need to know the file size: you just decide a blocksize (>= 65536 is usually ok) and use sendfile() as you use send().

> Ah ok, I misinterpreted what you wrote then.
> Generally speaking though, you don't need to know the file size: you just decide a blocksize (>= 65536 is usually ok) and use sendfile() as you use send().
But then you make much more syscalls than what's necessary, and your
heuristic for the block size is never going to match the choice the
kernel makes.
Anyway, here are the numbers, do you think it's interesting to merge
(I mean, you're the Python ftp expert ;-) ?

Specifying a big blocksize doesn't mean the transfer will be faster.
send/sendfile won't send more than a certain amount of bytes anyways.
If I'm not mistaken I recall from previous benchmarks that after a certain point (131072 or something) increasing the blocksize results in equal or even worse performances.
Another thing I don't like is that by doing so you implicitly assume that the file is "fstat-eable". I don't know if there are cases where it's not, but the less assumptions we do the better.
Note: I'm sure that for both send() and sendfile() blocksize>=65536 is faster than blocksize=8192 (the current default) so it probably makes sense to change that (I'll file a separate issue).

> Specifying a big blocksize doesn't mean the transfer will be faster.
> send/sendfile won't send more than a certain amount of bytes anyways.
The transfer won't be faster mainly because it's really I/O bound.
But it will use less CPU, only because you're making less syscalls.
> If I'm not mistaken I recall from previous benchmarks that after a certain point (131072 or something) increasing the blocksize results in equal or even worse performances.
I can perfectly believe this for a send loop, maybe because you're
exceeding the socket buffer size, or because your working set doesn't
fit into caches anymore, etc.
But for sendfile(), I don't see how calling it repeatedly could not be
slower than calling it once with the overall size: that's how netperf
and vsftpd use it, and probably others.
> Another thing I don't like is that by doing so you implicitly assume that the file is "fstat-eable". I don't know if there are cases where it's not, but the less assumptions we do the better.
Well, the file must be mmap-able, so I doubt fstat() is the biggest concern...

> The transfer won't be faster mainly because it's really I/O bound.
> But it will use less CPU, only because you're making less syscalls.
Have you actually measured this?

>> The transfer won't be faster mainly because it's really I/O bound.
>> But it will use less CPU, only because you're making less syscalls.
>
> Have you actually measured this?
"""
vanilla over Gb/s:
real 0m9.035s
user 0m0.523s
sys 0m1.412s
block-sendfile over Gb/s:
real 0m9.683s
user 0m0.253s
sys 0m1.212s
full-sendfile over Gb/s:
real 0m9.014s
user 0m0.059s
sys 0m1.000s
"""
As you can see, the throughput doesn't vary (the difference in "real
time" is just part of the variance).
However, the CPU usage (user+sys) is less for block-sendfile than send
loop, and less for full-sendfile than block-sendfile.
"""
vanilla over loopback:
real 0m3.200s
user 0m0.541s
sys 0m0.702s
block-sendfile over loopback:
real 0m2.713s
user 0m0.248s
sys 0m0.197s
full-sendfile over loopback:
real 0m1.718s
user 0m0.055s
sys 0m0.082s
"""
Same thing for loopback, except that here, zero-copy makes a
difference on the throughput because we're not I/O bound, but really
CPU/memory bound (and here sendfile of the complete file really
outperforms block-sendfile).
I don't have access to a 10Gb/s network, but basic math hints that
sendfile could make a difference on the overall throughput.

It seems you're right, sorry. We need to take that into account then.
In the meantime I rewrote the original patch and got rid of the "use_sendfile" explicit argument in order to attempt to use sendfile() by default and fall back on using send() if bytes sent were 0.
TSL_FTP related changes are still left out for now as I'm planning a little refactoring first.

> In the meantime I rewrote the original patch and got rid of the "use_sendfile" explicit argument in order to attempt to use sendfile() by default and fall back on using send() if bytes sent were 0.
"""
# block until socket is writable
select.select([], [sockno], [])
"""
I don't get it, why do you use select?

It's necessary because sendfile() can fail with EAGAIN.
As for your "blocksize = filesize" argument I changed my opinion: despite being less CPU consuming we might incur into problems if that number is too big. 'count' parameter on Linux, for example, is expected to be an unsigned int.
Other plarforms will also use different data types so we better stick with a fixed blocksize value (currently 8192).

> It's necessary because sendfile() can fail with EAGAIN.
It can fail with EAGAIN if the input FD is non-blocking, exactly like
the current implementation which calls fp.read(). Furthermore, since
sendfile actually supports only regular file and regular files don't
support non-blocking I/O, it's unlikely to ever happen.
> As for your "blocksize = filesize" argument I changed my opinion: despite being less CPU consuming we might incur into problems if that number is too big. 'count' parameter on Linux, for example, is expected to be an unsigned int.
'count' is size_t, like for mmap() and any other function accepting a
length, so nothing wrong can happen.
A platform which would have a sendfile prototype which doesn't support
sending a complete file at once would be completely broken...

> As for your "blocksize = filesize" argument I changed my opinion:
> despite being less CPU consuming we might incur into problems if
> that number is too big. 'count' parameter on Linux, for example, is
> expected to be an unsigned int.
> Other plarforms will also use different data types so we better stick
> with a fixed blocksize value (currently 8192).
If you really think a blocksize is necessary, you could choose a much
larger one for sendfile() (such as 16 MB). Then the overhead of system
calls would be much smaller.

> 'count' is size_t, like for mmap() and any other function accepting a length, so nothing wrong can happen.
Then why 'offset' and 'count' parameters have a different data type?
Linux: 
sendfile(..., off_t *offset, size_t count);
Solaris:
sendfile(..., off_t *off, size_t len);
HP-UX:
sendfile(..., off_t offset, bsize_t nbytes);
> A platform which would have a sendfile prototype which doesn't support
> sending a complete file at once would be completely broken...
You can't send a complete file at once in the first place unless it's very small. 
The usual way to send a file is chunk by chunk, so it wouldn't surprise me if sendfile() prototype does not support the use case you're describing.
Anyway, Antoine's suggestion makes sense to me: it's probably ok to just use a big value and be done with it.
16MB looks a little bit too much to me as the maximum amount of bytes sent per call is a lot less than 1MB, but even then it would probably be ok.
>> It's necessary because sendfile() can fail with EAGAIN.
> It can fail with EAGAIN if the input FD is non-blocking
It will. Try it yourself.
> Furthermore, since sendfile actually supports only regular file and regular 
> files don't support non-blocking I/O, it's unlikely to ever happen.
EAGAIN is caused by the socket fd not being ready yet, not the file fd.
Please try the patch before making such assumptions. We're going OT here.

> Then why 'offset' and 'count' parameters have a different data type?
Because offsets can be negative (e.g. for lseek), while a size can't.
That's why 'count' is size_t, not ssize_t.
>> Furthermore, since sendfile actually supports only regular file and regular
>> files don't support non-blocking I/O, it's unlikely to ever happen.
>
> EAGAIN is caused by the socket fd not being ready yet, not the file fd.
> Please try the patch before making such assumptions.
I didn't see the socket could be set to non-blocking.
In that case, there's a problem with the patch, since select can block
arbitrarily long because it doesn't take the socket timeout into
account.
Also, apparently socket.sendall() doesn't retry on EAGAIN, it doesn't
use BEGIN_SELECT_LOOP.
The risk of false positive (EAGAIN after select reported ready)
shouldn't be as bad as for sendto(), since usually you'll just get a
partial write for a stream oriented socket, but this could be bad for
e.g. a SCTP socket (since it's message-oriented).
> We're going OT here.
I'm leaving this topic, you can do as you like...

> Because offsets can be negative 
On Linux (and presumably on all POSIX platforms) passing a negative offset results in EINVAL.
> In that case, there's a problem with the patch, since select can block
> arbitrarily long because it doesn't take the socket timeout into
> account.
Right. I will fix that.
> Also, apparently socket.sendall() doesn't retry on EAGAIN, 
> it doesn't use BEGIN_SELECT_LOOP.
socket.sendall() is not supposed to return EAGAIN in the first place. And again, I don't see how this is related with the issue at hand.
> I'm leaving this topic, you can do as you like...
Bye.

A much larger patch which should address all issues is in attachment.
Updates:
- use poll() instead of select() whenever possible
- take socket timeout into account
- take SSL/FTPS into account
- when using select() look for EMFILE in case num fds > FD_SETSIZE
- look for (AttributeError, io.UnsupportedOperation) when invoking file.fileno() instead of Exception, which seemed too general

I don't understand why you must put the socket in non-blocking mode for sendfile().
Also I think the support code (_use_send() / _use_sendfile()) should be factored out somewhere. There could even be a socket.sendfile() method with the appropriate fallbacks?

> I don't understand why you must put the socket in 
> non-blocking mode for sendfile().
I did that mainly because I'm using select() / poll() and it seems kind of "natural" to set the socket in non-blocking mode (proftpd does the same).
I'm not sure whether it actually makes any difference though (on Linux it works either way, blocking or not).
I'm removing the non-blocking mode in the new attached patch. We can take take a look at the buildbots later and see how they behave.
> Also I think the support code (_use_send() / _use_sendfile()) should be factored out somewhere.
Agreed and turned them into methods.
> There could even be a socket.sendfile() method with the appropriate fallbacks?
Perhaps. The only thing which is not clear is how to deal with blocking vs. non-blocking sockets.
Also, Windows should also be covered and expose TransmitFile. 
It's probably better to discuss this elsewhere.

> > I don't understand why you must put the socket in 
> > non-blocking mode for sendfile().
> 
> I did that mainly because I'm using select() / poll() and it seems
> kind of "natural" to set the socket in non-blocking mode (proftpd does
> the same).
But why do you need to use select() / poll() ? Can't you just call
sendfile() right from the start?

Because otherwise sendfile() fails with EAGAIN many times before sending any actual data. 
select() / poll() make sure the while loop awakens only when the socket is ready to be written (as opposed to continuously catching EAGAIN and wait for sendfile() to succeed).

> Because otherwise sendfile() fails with EAGAIN many times before
> sending any actual data. 
EAGAIN on a blocking fd? Is it documented somewhere?
The Linux man page for sendfile() says:
 EAGAIN Nonblocking I/O has been selected using O_NONBLOCK and the
write would block.
FreeBSD apparently says something similar:
 [EAGAIN]		The socket is marked for non-blocking I/O and not all
			data was sent due to the socket buffer being filled.
			If specified, the number of bytes successfully sent
			will be returned in *sbytes.

I see. Well, what I'm experiencing right now if I remove the select() / poll() call is a sequence of EAGAIN errors alternated by successful sendfile() calls. 
Either the man page is wrong or I'm missing something.

> I see. Well, what I'm experiencing right now if I remove the
> select() / poll() call is a sequence of EAGAIN errors alternated by
> successful sendfile() calls. 
> Either the man page is wrong or I'm missing something.
Weird. I guess it would be nice to dig a bit more :-)

After digging a bit further it seems EAGAIN occurs in case a timeout was previously set against the socket as in ftplib.FTP(..., timeout=2) (at least on Linux, FWICT).
As such, we can debate whether avoid using select/poll if timeout was not set.
I'll that a look at the man pages of the other POSIX platforms and figure whether EAGAIN is interpreted as on Linux.
Other than that, the patch is reasonably ok to me and can be committed as-is and blocksize argument tuning can be discussed in a separate ticket.

> After digging a bit further it seems EAGAIN occurs in case a timeout was previously 
> set against the socket as in ftplib.FTP(..., timeout=2) (at least on Linux, FWICT).
Ah, indeed. That's because socket timeout makes the underlying fd non-blocking.
Which means there probably should be a higher-level sendfile() facility for sockets, taking into account the socket timeout...

I created issue 17552 in order to discuss about socket.sendfile() addition.

Updated patch which uses the newly added socket.sendfile() method (issue 17552).