|
IO::AIO - Asynchronous Input/Output |
IO::AIO - Asynchronous Input/Output
use IO::AIO;
aio_open "/etc/passwd", O_RDONLY, 0, sub {
my $fh = shift
or die "/etc/passwd: $!";
...
};
aio_unlink "/tmp/file", sub { };
aio_read $fh, 30000, 1024, $buffer, 0, sub {
$_[0] > 0 or die "read error: $!";
};
# version 2+ has request and group objects use IO::AIO 2;
aioreq_pri 4; # give next request a very high priority
my $req = aio_unlink "/tmp/file", sub { };
$req->cancel; # cancel request if still in queue
my $grp = aio_group sub { print "all stats done\n" };
add $grp aio_stat "..." for ...;
# AnyEvent integration
open my $fh, "<&=" . IO::AIO::poll_fileno or die "$!";
my $w = AnyEvent->io (fh => $fh, poll => 'r', cb => sub { IO::AIO::poll_cb });
# Event integration
Event->io (fd => IO::AIO::poll_fileno,
poll => 'r',
cb => \&IO::AIO::poll_cb);
# Glib/Gtk2 integration
add_watch Glib::IO IO::AIO::poll_fileno,
in => sub { IO::AIO::poll_cb; 1 };
# Tk integration
Tk::Event::IO->fileevent (IO::AIO::poll_fileno, "",
readable => \&IO::AIO::poll_cb);
# Danga::Socket integration
Danga::Socket->AddOtherFds (IO::AIO::poll_fileno =>
\&IO::AIO::poll_cb);
This module implements asynchronous I/O using whatever means your operating system supports.
Asynchronous means that operations that can normally block your program (e.g. reading from disk) will be done asynchronously: the operation will still block, but you can do something else in the meantime. This is extremely useful for programs that need to stay interactive even when doing heavy I/O (GUI programs, high performance network servers etc.), but can also be used to easily do operations in parallel that are normally done sequentially, e.g. stat'ing many files, which is much faster on a RAID volume or over NFS when you do a number of stat operations concurrently.
While most of this works on all types of file descriptors (for example sockets), using these functions on file descriptors that support nonblocking operation (again, sockets, pipes etc.) is very inefficient. Use an event loop for that (such as the Event module): IO::AIO will naturally fit into such an event loop itself.
In this version, a number of threads are started that execute your requests and signal their completion. You don't need thread support in perl, and the threads created by this module will not be visible to perl. In the future, this module might make use of the native aio functions available on many operating systems. However, they are often not well-supported or restricted (GNU/Linux doesn't allow them on normal files currently, for example), and they would only support aio_read and aio_write, so the remaining functionality would have to be implemented using threads anyway.
Although the module will work in the presence of other (Perl-) threads,
it is currently not reentrant in any way, so use appropriate locking
yourself, always call poll_cb from within the same thread, or never
call poll_cb (or other aio_ functions) recursively.
This is a simple example that uses the Event module and loads /etc/passwd asynchronously:
use Fcntl; use Event; use IO::AIO;
# register the IO::AIO callback with Event
Event->io (fd => IO::AIO::poll_fileno,
poll => 'r',
cb => \&IO::AIO::poll_cb);
# queue the request to open /etc/passwd
aio_open "/etc/passwd", O_RDONLY, 0, sub {
my $fh = shift
or die "error while opening: $!";
# stat'ing filehandles is generally non-blocking
my $size = -s $fh;
# queue a request to read the file
my $contents;
aio_read $fh, 0, $size, $contents, 0, sub {
$_[0] == $size
or die "short read: $!";
close $fh;
# file contents now in $contents
print $contents;
# exit event loop and program
Event::unloop;
};
};
# possibly queue up other requests, or open GUI windows, # check for sockets etc. etc.
# process events as long as there are some: Event::loop;
Every aio_* function creates a request. which is a C data structure not
directly visible to Perl.
If called in non-void context, every request function returns a Perl object representing the request. In void context, nothing is returned, which saves a bit of memory.
The perl object is a fairly standard ref-to-hash object. The hash contents are not used by IO::AIO so you are free to store anything you like in it.
During their existance, aio requests travel through the following states, in order:
While request submission and execution is fully asynchronous, result
processing is not and relies on the perl interpreter calling poll_cb
(or another function with the same effect).
poll_cb.
The poll_cb function will process all outstanding aio requests by
calling their callbacks, freeing memory associated with them and managing
any groups they are contained in.
All the aio_* calls are more or less thin wrappers around the syscall
with the same name (sans aio_). The arguments are similar or identical,
and they all accept an additional (and optional) $callback argument
which must be a code reference. This code reference will get called with
the syscall return code (e.g. most syscalls return -1 on error, unlike
perl, which usually delivers ``false'') as it's sole argument when the given
syscall has been executed asynchronously.
All functions expecting a filehandle keep a copy of the filehandle internally until the request has finished.
All functions return request objects of type the IO::AIO::REQ manpage that allow further manipulation of those requests while they are in-flight.
The pathnames you pass to these routines must be absolute and encoded as octets. The reason for the former is that at the time the request is being executed, the current working directory could have changed. Alternatively, you can make sure that you never change the current working directory anywhere in the program and then use relative paths.
To encode pathnames as octets, either make sure you either: a) always pass in filenames you got from outside (command line, readdir etc.) without tinkering, b) are ASCII or ISO 8859-1, c) use the Encode module and encode your pathnames to the locale (or other) encoding in effect in the user environment, d) use Glib::filename_from_unicode on unicode filenames or e) use something else to ensure your scalar has the correct contents.
This works, btw. independent of the internal UTF-8 bit, which IO::AIO handles correctly wether it is set or not.
$pri is given, sets the priority for the next aio request.
The default priority is 0, the minimum and maximum priorities are -4
and 4, respectively. Requests with higher priority will be serviced
first.
The priority will be reset to 0 after each call to one of the aio_*
functions.
Example: open a file with low priority, then read something from it with higher priority so the read request is serviced before other low priority open requests (potentially spamming the cache):
aioreq_pri -3;
aio_open ..., sub {
return unless $_[0];
aioreq_pri -2;
aio_read $_[0], ..., sub {
...
};
};
aioreq_pri, but subtracts the given value from the current
priority, so the effect is cumulative.
The pathname passed to aio_open must be absolute. See API NOTES, above,
for an explanation.
The $flags argument is a bitmask. See the Fcntl module for a
list. They are the same as used by sysopen.
Likewise, $mode specifies the mode of the newly created file, if it
didn't exist and O_CREAT has been given, just like perl's sysopen,
except that it is mandatory (i.e. use 0 if you don't create new files,
and 0666 or 0777 if you do). Note that the $mode will be modified
by the umask in effect then the request is being executed, so better never
change the umask.
Example:
aio_open "/etc/passwd", O_RDONLY, 0, sub {
if ($_[0]) {
print "open successful, fh is $_[0]\n";
...
} else {
die "open failed: $!\n";
}
};
Unfortunately, you can't do this to perl. Perl insists very strongly on closing the file descriptor associated with the filehandle itself. Here is what aio_close will try:
1. dup()licate the fd 2. asynchronously close() the duplicated fd 3. dup()licate the fd once more 4. let perl close() the filehandle 5. asynchronously close the duplicated fd
The idea is that the first close() flushes stuff to disk that closing an
fd will flush, so when perl closes the fd, nothing much will need to be
flushed. The second async. close() will then flush stuff to disk that
closing the last fd to the file will flush.
Just FYI, SuSv3 has this to say on close:
All outstanding record locks owned by the process on the file associated with the file descriptor shall be removed.
If fildes refers to a socket, close() shall cause the socket to be destroyed. ... close() shall block for up to the current linger interval until all data is transmitted. [this actually sounds like a specification bug, but who knows]
And at least Linux additionally actually flushes stuff on every close, even when the file itself is still open.
Sounds enourmously inefficient and complicated? Yes... please show me how to nuke perl's fd out of existence...
$length bytes from the specified $fh and $offset
into the scalar given by $data and offset $dataoffset and calls the
callback without the actual number of bytes read (or -1 on error, just
like the syscall).
If $offset is undefined, then the current file descriptor offset will
be used (and updated), otherwise the file descriptor offset will not be
changed by these calls.
If $length is undefined in aio_write, use the remaining length of $data.
If $dataoffset is less than zero, it will be counted from the end of
$data.
The $data scalar MUST NOT be modified in any way while the request
is outstanding. Modifying it can result in segfaults or World War III (if
the necessary/optional hardware is installed).
Example: Read 15 bytes at offset 7 into scalar $buffer, starting at
offset 0 within the scalar:
aio_read $fh, 7, 15, $buffer, 0, sub {
$_[0] > 0 or die "read error: $!";
print "read $_[0] bytes: <$buffer>\n";
};
$length bytes from $in_fh to $out_fh. It starts
reading at byte offset $in_offset, and starts writing at the current
file offset of $out_fh. Because of that, it is not safe to issue more
than one aio_sendfile per $out_fh, as they will interfere with each
other.
This call tries to make use of a native sendfile syscall to provide
zero-copy operation. For this to work, $out_fh should refer to a
socket, and $in_fh should refer to mmap'able file.
If the native sendfile call fails or is not implemented, it will be
emulated, so you can call aio_sendfile on any type of filehandle
regardless of the limitations of the operating system.
Please note, however, that aio_sendfile can read more bytes from
$in_fh than are written, and there is no way to find out how many
bytes have been read from aio_sendfile alone, as aio_sendfile only
provides the number of bytes written to $out_fh. Only if the result
value equals $length one can assume that $length bytes have been
read.
aio_readahead populates the page cache with data from a file so that
subsequent reads from that file will not block on disk I/O. The $offset
argument specifies the starting point from which data is to be read and
$length specifies the number of bytes to be read. I/O is performed in
whole pages, so that offset is effectively rounded down to a page boundary
and bytes are read up to the next page boundary greater than or equal to
(off-set+length). aio_readahead does not read beyond the end of the
file. The current file offset of the file is left unchanged.
If that syscall doesn't exist (likely if your OS isn't Linux) it will be emulated by simply reading the data, which would have a similar effect.
stat or lstat in void context. The callback will
be called after the stat and the results will be available using stat _
or -s _ etc...
The pathname passed to aio_stat must be absolute. See API NOTES, above,
for an explanation.
Currently, the stats are always 64-bit-stats, i.e. instead of returning an error when stat'ing a large file, the results will be silently truncated unless perl itself is compiled with large file support.
Example: Print the length of /etc/passwd:
aio_stat "/etc/passwd", sub {
$_[0] and die "stat failed: $!";
print "size is ", -s _, "\n";
};
utime function (including the special case of $atime
and $mtime being undef). Fractional times are supported if the underlying
syscalls support them.
When called with a pathname, uses utimes(2) if available, otherwise
utime(2). If called on a file descriptor, uses futimes(2) if available,
otherwise returns ENOSYS, so this is not portable.
Examples:
# set atime and mtime to current time (basically touch(1)): aio_utime "path", undef, undef; # set atime to current time and mtime to beginning of the epoch: aio_utime "path", time, undef; # undef==0
chown function, except that undef for either $uid
or $gid is being interpreted as ``do not change'' (but -1 can also be used).
Examples:
# same as "chown root path" in the shell: aio_chown "path", 0, -1; # same as above: aio_chown "path", 0, undef;
truncate(2) or ftruncate(2).
chmod function.
Asynchronously create a device node (or fifo). See mknod(2).
The only (POSIX-) portable way of calling this function is:
aio_mknod $path, IO::AIO::S_IFIFO | $mode, 0, sub { ...
$srcpath at
the path $dstpath and call the callback with the result code.
$srcpath at
the path $dstpath and call the callback with the result code.
$path and pass it to
the callback. If an error occurs, nothing or undef gets passed to the
callback.
$srcpath to $dstpath, just as
rename(2) and call the callback with the result code.
$mode will be modified by the umask at the time the
request is executed, so do not change your umask.
aio_readdir reads an entire
directory (i.e. opendir + readdir + closedir). The entries will not be
sorted, and will NOT include the . and .. entries.
The callback a single argument which is either undef or an array-ref
with the filenames.
$srcpath to $dstpath and call the callback with
the 0 (error) or -1 ok.
This is a composite request that it creates the destination file with
mode 0200 and copies the contents of the source file into it using
aio_sendfile, followed by restoring atime, mtime, access mode and
uid/gid, in that order.
If an error occurs, the partial destination file will be unlinked, if possible, except when setting atime, mtime, access mode and uid/gid, where errors are being ignored.
$srcpath to $dstpath and call the callback with
the 0 (error) or -1 ok.
This is a composite request that tries to rename(2) the file first. If
rename files with EXDEV, it copies the file with aio_copy and, if
that is successful, unlinking the $srcpath.
aio_readdir) but additionally tries to
efficiently separate the entries of directory $path into two sets of
names, directories you can recurse into (directories), and ones you cannot
recurse into (everything else, including symlinks to directories).
aio_scandir is a composite request that creates of many sub requests_
$maxreq specifies the maximum number of outstanding aio requests that
this function generates. If it is <= 0, then a suitable default
will be chosen (currently 4).
On error, the callback is called without arguments, otherwise it receives two array-refs with path-relative entry names.
Example:
aio_scandir $dir, 0, sub {
my ($dirs, $nondirs) = @_;
print "real directories: @$dirs\n";
print "everything else: @$nondirs\n";
};
Implementation notes.
The aio_readdir cannot be avoided, but stat()'ing every entry can.
After reading the directory, the modification time, size etc. of the directory before and after the readdir is checked, and if they match (and isn't the current time), the link count will be used to decide how many entries are directories (if >= 2). Otherwise, no knowledge of the number of subdirectories will be assumed.
Then entries will be sorted into likely directories (everything without
a non-initial dot currently) and likely non-directories (everything
else). Then every entry plus an appended /. will be stat'ed,
likely directories first. If that succeeds, it assumes that the entry
is a directory or a symlink to directory (which will be checked
seperately). This is often faster than stat'ing the entry itself because
filesystems might detect the type of the entry without reading the inode
data (e.g. ext2fs filetype feature).
If the known number of directories (link count - 2) has been reached, the rest of the entries is assumed to be non-directories.
This only works with certainty on POSIX (= UNIX) filesystems, which fortunately are the vast majority of filesystems around.
It will also likely work on non-POSIX filesystems with reduced efficiency as those tend to return 0 or 1 as link counts, which disables the directory counting heuristic.
$path, return the
status of the final rmdir only. This is a composite request that
uses aio_scandir to recurse into and rmdir directories, and unlink
everything else.
If this call isn't available because your OS lacks it or it couldn't be
detected, it will be emulated by calling fsync instead.
Returns an object of class the IO::AIO::GRP manpage. See its documentation below for more info.
Example:
my $grp = aio_group sub {
print "all stats done\n";
};
add $grp
(aio_stat ...),
(aio_stat ...),
...;
While this request does nothing, it still goes through the execution phase and still requires a worker thread. Thus, the callback will not be executed immediately but only after other requests in the queue have entered their execution phase. This can be used to measure request latency.
While it is theoretically handy to have simple I/O scheduling requests like sleep and file handle readable/writable, the overhead this creates is immense (it blocks a thread for a long time) so do not use this function except to put your application under artificial I/O pressure.
All non-aggregate aio_* functions return an object of this class when
called in non-void context.
This class is a subclass of the IO::AIO::REQ manpage, so all its methods apply to objects of this class, too.
A IO::AIO::GRP object is a special request that can contain multiple other aio requests.
You create one by calling the aio_group constructing function with a
callback that will be called when all contained requests have entered the
done state:
my $grp = aio_group sub {
print "all requests are done\n";
};
You add requests by calling the add method with one or more
IO::AIO::REQ objects:
$grp->add (aio_unlink "...");
add $grp aio_stat "...", sub {
$_[0] or return $grp->result ("error");
# add another request dynamically, if first succeeded
add $grp aio_open "...", sub {
$grp->result ("ok");
};
};
This makes it very easy to create composite requests (see the source of
aio_move for an application) that work and feel like simple requests.
IO::AIO::poll_cb, just like any other request.Their lifetime, simplified, looks like this: when they are empty, they
will finish very quickly. If they contain only requests that are in the
done state, they will also finish. Otherwise they will continue to
exist.
That means after creating a group you have some time to add requests. And in the callbacks of those requests, you can add further requests to the group. And only when all those requests have finished will the the group itself finish.
Returns all its arguments.
value(s) that will be passed to the group callback when all
subrequests have finished and set thre groups errno to the current value
of errno (just like calling errno without an error number). By default,
no argument will be passed and errno is zero.
$errno, or the current value of errno
when the argument is missing.
Every aio request has an associated errno value that is restored when the callback is invoked. This method lets you change this value from its default (0).
Calling result will also set errno, so make sure you either set $!
before the call to result, or call c<errno> after it.
aio_scandir might generate hundreds of thousands aio_stat
requests, delaying any later requests for a long time.
To avoid this, and allow incremental generation of requests, you can
instead a group and set a feeder on it that generates those requests. The
feed callback will be called whenever there are few enough (see limit,
below) requests active in the group itself and is expected to queue more
requests.
The feed callback can queue as many requests as it likes (i.e. add does
not impose any limits).
If the feed does not queue more requests when called, it will be automatically removed from the group.
If the feed limit is 0, it will be set to 2 automatically.
Example:
# stat all files in @files, but only ever use four aio requests concurrently:
my $grp = aio_group sub { print "finished\n" };
limit $grp 4;
feed $grp sub {
my $file = pop @files
or return;
add $grp aio_stat $file, sub { ... };
};
Setting the limit to 0 will pause the feeding process.
poll_cb to check the results.
See poll_cb for an example.
IO::AIO::max_poll_req and IO::AIO::max_poll_time.
If not all requests were processed for whatever reason, the filehandle
will still be ready when poll_cb returns.
Example: Install an Event watcher that automatically calls IO::AIO::poll_cb with high priority:
Event->io (fd => IO::AIO::poll_fileno,
poll => 'r', async => 1,
cb => \&IO::AIO::poll_cb);
0, meaning infinity)
that are being processed by IO::AIO::poll_cb in one call, respectively
the maximum amount of time (default 0, meaning infinity) spent in
IO::AIO::poll_cb to process requests (more correctly the mininum amount
of time poll_cb is allowed to use).
Setting max_poll_time to a non-zero value creates an overhead of one
syscall per request processed, which is not normally a problem unless your
callbacks are really really fast or your OS is really really slow (I am
not mentioning Solaris here). Using max_poll_reqs incurs no overhead.
Setting these is useful if you want to ensure some level of interactiveness when perl is not fast enough to process all requests in time.
For interactive programs, values such as 0.01 to 0.1 should be fine.
Example: Install an Event watcher that automatically calls IO::AIO::poll_cb with low priority, to ensure that other parts of the program get the CPU sometimes even under high AIO load.
# try not to spend much more than 0.1s in poll_cb IO::AIO::max_poll_time 0.1;
# use a low priority so other tasks have priority
Event->io (fd => IO::AIO::poll_fileno,
poll => 'r', nice => 1,
cb => &IO::AIO::poll_cb);
select on the filehandle. This is useful if you want to
synchronously wait for some requests to finish).
See nreqs for an example.
Returns the number of requests processed, but is otherwise strictly equivalent to:
IO::AIO::poll_wait, IO::AIO::poll_cb
Strictly equivalent to:
IO::AIO::poll_wait, IO::AIO::poll_cb
while IO::AIO::nreqs;
$nthreads. The current
default is 8, which means eight asynchronous operations can execute
concurrently at any one time (the number of outstanding requests,
however, is unlimited).
IO::AIO starts threads only on demand, when an AIO request is queued and no free thread exists. Please note that queueing up a hundred requests can create demand for a hundred threads, even if it turns out that everything is in the cache and could have been processed faster by a single thread.
It is recommended to keep the number of threads relatively low, as some Linux kernel versions will scale negatively with the number of threads (higher parallelity => MUCH higher latency). With current Linux 2.6 versions, 4-32 threads should be fine.
Under most circumstances you don't need to call this function, as the module selects a default that is suitable for low to moderate load.
$nthreads. If more than the
specified number of threads are currently running, this function kills
them. This function blocks until the limit is reached.
While $nthreads are zero, aio requests get queued but not executed
until the number of threads has been increased again.
This module automatically runs max_parallel 0 at program end, to ensure
that all threads are killed and that there are no outstanding requests.
Under normal circumstances you don't need to call this function.
$nthreads other threads are also
idle, it will free its resources and exit.
This is useful when you allow a large number of threads (e.g. 100 or 1000) to allow for extremely high load situations, but want to free resources under normal circumstances (1000 threads can easily consume 30MB of RAM).
The default is probably ok in most situations, especially if thread creation is fast. If thread creation is very slow on your system you might want to use larger values.
aio_group together with a feed callback.
Sets the maximum number of outstanding requests to $nreqs. If you
do queue up more than this number of requests, the next call to the
poll_cb (and poll_some and other functions calling poll_cb)
function will block until the limit is no longer exceeded.
The default value is very large, so there is no practical limit on the number of outstanding requests.
You can still queue as many requests as you want. Therefore,
max_oustsanding is mainly useful in simple scripts (with low values) or
as a stop gap to shield against fatal memory overflow (with large values).
Example: wait till there are no outstanding requests anymore:
IO::AIO::poll_wait, IO::AIO::poll_cb
while IO::AIO::nreqs;
This module should do ``the right thing'' when the process using it forks:
Before the fork, IO::AIO enters a quiescent state where no requests can be added in other threads and no results will be processed. After the fork the parent simply leaves the quiescent state and continues request/result processing, while the child frees the request/result queue (so that the requests started before the fork will only be handled in the parent). Threads will be started on demand until the limit set in the parent process has been reached again.
In short: the parent will, after a short pause, continue as if fork had not been called, while the child will act as if IO::AIO has not been used yet.
Per-request usage:
Each aio request uses - depending on your architecture - around 100-200 bytes of memory. In addition, stat requests need a stat buffer (possibly a few hundred bytes), readdir requires a result buffer and so on. Perl scalars and other data passed into aio requests will also be locked and will consume memory till the request has entered the done state.
This is not awfully much, so queuing lots of requests is not usually a problem.
Per-thread usage:
In the execution phase, some aio requests require more memory for temporary buffers, and each thread requires a stack and other data structures (usually around 16k-128k, depending on the OS).
Known bugs will be fixed in the next release.
Marc Lehmann <schmorp@schmorp.de> http://home.schmorp.de/
|
IO::AIO - Asynchronous Input/Output |