New sendfile(2) syscall. A joint effort of NGINX and Netflix from 2013 and

up to now.

The new sendfile is the code that Netflix uses to send their multiple tens
of gigabits of data per second. The new implementation features asynchronous
I/O, when I/O operations are launched, but not awaited to be complete. An
explanation of why such behavior is beneficial compared to old one is
going to be too long for a commit message, so we will skip it here.

Additional features of new syscall are extra flags, which provide an
application more control over data sent. The SF_NOCACHE flag tells
kernel that data shouldn't be cached after it was sent. The SF_READAHEAD()
macro allows to specify readahead size in pages.

The new syscalls is a drop in replacement. No modifications are required
to applications. One can take nginx binary for stable/10 and run it
successfully on head. Although SF_NODISKIO lost its original sense, as now
sendfile doesn't block, and now means something completely different (tm),
using the new sendfile the old way is absolutely safe.

Celebrates:	Netflix global launch!
Sponsored by:	Nginx, Inc.
Sponsored by:	Netflix
Relnotes:	yes

1 files changed, 91 insertions, 32 deletions

diff --git a/lib/libc/sys/sendfile.2 b/lib/libc/sys/sendfile.2
index b363382..2b52dd9 100644
--- a/lib/libc/sys/sendfile.2
+++ b/lib/libc/sys/sendfile.2
@@ -25,7 +25,7 @@
 .\"
 .\" $FreeBSD$
 .\"
-.Dd January 7, 2010
+.Dd January 7, 2016
 .Dt SENDFILE 2
 .Os
 .Sh NAME
@@ -46,7 +46,7 @@
 The
 .Fn sendfile
 system call
-sends a regular file specified by descriptor
+sends a regular file or shared memory object specified by descriptor
 .Fa fd
 out a stream socket specified by descriptor
 .Fa s .
@@ -101,32 +101,55 @@ the system will write the total number of bytes sent on the socket to the
 variable pointed to by
 .Fa sbytes .
 .Pp
-The
+The least significant 16 bits of
 .Fa flags
 argument is a bitmap of these values:
-.Bl -item -offset indent
-.It
-.Dv SF_NODISKIO .
-This flag causes any
-.Fn sendfile
-call which would block on disk I/O to instead
-return
-.Er EBUSY .
-Busy servers may benefit by transferring requests that would
-block to a separate I/O worker thread.
-.It
-.Dv SF_MNOWAIT .
-Do not wait for some kernel resource to become available,
-in particular,
-.Vt mbuf
-and
-.Vt sf_buf .
-The flag does not make the
-.Fn sendfile
-syscall truly non-blocking, since other resources are still allocated
-in a blocking fashion.
-.It
-.Dv SF_SYNC .
+.Bl -tag -offset indent
+.It Dv SF_NODISKIO
+This flag causes
+.Nm
+to return
+.Er EBUSY
+instead of blocking when a busy page is encountered.
+This rare situation can happen if some other process is now working
+with the same region of the file.
+It is advised to retry the operation after a short period.
+.Pp
+Note that in older
+.Fx
+versions the
+.Dv SF_NODISKIO
+had slightly different notion.
+The flag prevented
+.Nm
+to run I/O operations in case if an invalid (not cached) page is encountered,
+thus avoiding blocking on I/O.
+Starting with
+.Fx 11
+.Nm
+sending files off the
+.Xr ffs 7
+filesystem doesn't block on I/O
+(see 
+.Sx IMPLEMENTATION NOTES
+), so the condition no longer applies.
+However, it is safe if an application utilizes
+.Dv SF_NODISKIO
+and on
+.Er EBUSY
+performs the same action as it did in
+older
+.Fx
+versions, e.g.
+.Xr aio_read 2,
+.Xr read 2
+or
+.Nm
+in a different context.
+.It Dv SF_NOCACHE
+The data sent to socket will not be cached by the virtual memory system,
+and will be freed directly to the pool of free pages.
+.It Dv SF_SYNC
 .Nm
 sleeps until the network stack no longer references the VM pages
 of the file, making subsequent modifications to it safe.
@@ -134,6 +157,22 @@ Please note that this is not a guarantee that the data has actually
 been sent.
 .El
 .Pp
+The most significant 16 bits of
+.Fa flags
+specify amount of pages that
+.Nm
+may read ahead when reading the file.
+A macro
+.Fn SF_FLAGS
+is provided to combine readahead amount and flags.
+Example shows specifing readahead of 16 pages and
+.Dv SF_NOCACHE
+flag:
+.Pp
+.Bd -literal -offset indent -compact
+	SF_FLAGS(16, SF_NOCACHE)
+.Ed
+.Pp
 When using a socket marked for non-blocking I/O,
 .Fn sendfile
 may send fewer bytes than requested.
@@ -149,6 +188,18 @@ The
 .Fx
 implementation of
 .Fn sendfile
+doesn't block on disk I/O when it sends a file off the
+.Xr ffs 7
+filesystem.
+The syscall returns success before the actual I/O completes, and data
+is put into the socket later unattended.
+However, the order of data in the socket is preserved, so it is safe
+to do further writes to the socket.
+.Pp
+The
+.Fx
+implementation of
+.Fn sendfile
 is "zero-copy", meaning that it has been optimized so that copying of the file data is avoided.
 .Sh TUNING
 On some architectures, this system call internally uses a special
@@ -232,12 +283,10 @@ The
 argument
 is not a valid socket descriptor.
 .It Bq Er EBUSY
-Completing the entire transfer would have required disk I/O, so
-it was aborted.
-Partial data may have been sent.
-(This error can only occur when
+A busy page was encountered and
 .Dv SF_NODISKIO
-is specified.)
+had been specified.
+Partial data may have been sent.
 .It Bq Er EFAULT
 An invalid address was specified for an argument.
 .It Bq Er EINTR
@@ -310,9 +359,19 @@ first appeared in
 .Fx 3.0 .
 This manual page first appeared in
 .Fx 3.1 .
+In
+.Fx 10
+support for sending shared memory descriptors had been introduced.
+In
+.Fx 11
+a non-blocking implementation had been introduced.
 .Sh AUTHORS
-The
+The initial implementation of
 .Fn sendfile
 system call
 and this manual page were written by
 .An David G. Lawrence Aq Mt dg@dglawrence.com .
+The
+.Fx 11
+implementation was written by
+.An Gleb Smirnoff Aq Mt glebius@FreeBSD.org .