On Sat, 2003-01-25 at 00:17, Rasmus Lerdorf wrote:
> Take this script:
>
> <? include '/home/rasmus/foo/u2.inc' ?>
>
> So, a simple include with an absolute pathname to avoid any include_path
> searching and all safe_mode and open_basedir checking turned off. The
> latter of course being an absolute killer. We still need a lot of system
> calls to handle this include:
>
> getcwd("/home/rasmus", 4096) = 13
> lstat64("/home", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
> lstat64("/home/rasmus", {st_mode=S_IFDIR|0777, st_size=45056, ...}) = 0
> lstat64("/home/rasmus/foo", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
> lstat64("/home/rasmus/foo/u2.inc", {st_mode=S_IFREG|0644, st_size=12, ...}) = 0
> open("/home/rasmus/foo/u2.inc", O_RDONLY) = 3
> fstat64(3, {st_mode=S_IFREG|0644, st_size=12, ...}) = 0
> fstat64(3, {st_mode=S_IFREG|0644, st_size=12, ...}) = 0
> fstat64(3, {st_mode=S_IFREG|0644, st_size=12, ...}) = 0
> mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) =
>0x4002c000
> _llseek(3, 0, [0], SEEK_CUR) = 0
> lseek(3, 0, SEEK_SET) = 0
>
> The 4 lstat64() calls here are due to calling realpath() which causes
> every dir leading up to the filename to be stat'ed to check if it is a
> link so we can normalize the filename before adding it to the included
> files list. It's handy and nice to normalize the path here, but it is
> really really expensive! And this overhead doesn't go away with opcode
> caches either as they still need to go through this step to determine if
> the included file has been cached or not. I also don't see the point of
> those three fstat64() calls that I guess are a result of the sanity
> checking we do after the fopen.
>
> Due to our current implementation, if you have a lot of includes, you
> really should put all your include files in / and you will see some
> impressive improvements. This is annoying, but I am not sure how to fix
> it. A couple of things crossed my mind:
>
> 1. Don't normalize the path names, but instead do a single stat on the
> file and record the device and inode of it and use that to match it
> up later to check for multiple inclusion and the various opcode
> caches would use that as a cache key.
>
> 2. Add a fast_include directive, or something equally lame-sounding,
> which skips the getcwd(), realpath() and sanity checks and just
> includes the file.
>
> 3. For absolute paths, don't call realpath() to normalize, just strip
> out multiple /'s and any .'s or whatever we think is necessary to
> try to normalize it short of having to stat every single dir leading
> up to the file.
>
> 4. Do some sort of caching on this info so it doesn't happen on every
> include on every request.
Zend Accellerator has a config setting for disabling getcwd(), I think
that would be a viable option for PHP too, rather than a separate
include statement.
Don't you have any FreeBSD kernel hackers in Y!? You could probably
convince/bully one of them to fix it, that's what we always do with ours
here ;-)
- Stig
--
PHP Development Mailing List <http://www.php.net/>
To unsubscribe, visit: http://www.php.net/unsub.php
