Re: [hybi] Background info: Properties of sendfile()

[Greg Wilkins <[e-mail]>]

On 6 August 2010 11:28, Jamie Lokier <[e-mail]> wrote:
> Roy T. Fielding wrote:
>> On Aug 5, 2010, at 5:49 PM, Jamie Lokier wrote:
>>
>> > It's time to explain sendfile() better.
>> >
>> > 0. Nearly all OSes used for web serving have something called
>> >   sendfile() or under a different name.  It has become a standard
>> >   method used in high performance static servers, and is beginning to
>> >   be used in dynamic servers that do smart caching.  It is given an
>> >   open file handle and range, and writes content of that file to a
>> >   TCP socket as efficiently as possible.  Some OSes allow contents of
>> >   process memory to be sent as well.
>>
>> Unfortunately, you should also be aware that nearly all OSes have
>> severe kernel bugs in their implementations of sendfile and seem
>> to be getting progressively worse over time, to the point where
>> Apache httpd needs to disable it by default.  I don't know why.
>
> Thanks.  Ugh.  (I would like to know more, if you can point me at the details.)

For jetty using java's file mapped buffers, we've encountered 2 key
problems that cause us to turn it off.
Firstly because the file is mapped to address space, even if it is not
actually loaded into user space, there was a problem with leaking
address space. Once mapped the address range was never recovered, even
if the buffer was discarded. It was possible (on 32 bit machines) to
consume all the available address space by repeatedly creating and
sending large files.  I'm not sure if this was an OS flaw or just the
way the JVMs expose the feature as a memory mapped buffer.
We also have issues with the windows implementation which causes the
file system to lock the file if it is mapped.  This can prevent hot
redeploys and efficient development cycles.

But even if the send file semantic is not always available, it is a
big gain when it is available.  So having support for large frames is
good.

> I.e. it just means "don't mandate that chunks must be small".
> It's not asking much.

So is 16 exebytes enough for you (see Ian's email).
Not sure how I'm going to write a unit test for that one:)



cheers