[lkml]   [2005]   [Sep]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: "Read my lips: no more merges" - aka Linux 2.6.14-rc1
Benjamin LaHaise a écrit :
> On Tue, Sep 13, 2005 at 09:04:32AM +0200, Eric Dumazet wrote:
>>I wish a process param could allow open() to take any free fd available,
>>not the lowest one. One can always use fcntl(fd, F_DUPFD, slot) to move a
>>fd on a specific high slot and always keep the 64 first fd slots free to
>>speedup the kernel part at open()/dup()/socket() time.
> The overhead is easy to avoid by making use of dup2() and close() to keep
> the lowest file descriptors in the table free, allowing open() and socket()
> to always return 3 or 4.

Yes, this is what I described :) Maybe this was not clear.

> Alternatively, the kernel could track available file descriptors using a
> tree to efficiently insert freed slots into an ordered list of free
> regions (something similar to the avl tree used in vmas). Is it worth
> doing?

Well no, since a user app can manage itself this part if it happens to be
performance critical.

Sample of a user land lib : Each time a new fd is returned by
open()/socket()/pipe()/accept()... the thread should call

fd = fdcache_dupfd(fd);

And close the file using fdcache_closefd(fd) instead of close(fd);

* Unix kernel has an expensive get_unused_fd() function :
* This is because semantics of Unix mandates that a open()/pipe()/socket()/ call always returns the lowest fd, not a random one.
* Linux use a linear scan of a table of bits.
* A program handling 1.000.000 files scans about 128 KB of ram, with a spinlock held : No other thread can get a fd.
* The trick is to use this library to make sure 64 low fds are available, so that the standard unix functions
* dont have to scan a lot of fd before finding a free one.
* And remap them using fcntl(F_DUPFD) at precise slots we manage ourselfs.
#include <pthread.h>
#include <fcntl.h>
#include <stdlib.h>
#include <unistd.h>

# define MAXFDS 1500000

struct {
pthread_mutex_t lock;
unsigned int cache_fd;
unsigned int next_alloc;
unsigned int *cache_tab;
} fdd;

void fdcache_init()
pthread_mutex_init(&fdd.lock, NULL);
fdd.cache_tab = calloc(MAXFDS, sizeof(unsigned int));
fdd.next_alloc = 64;

int fdcache_dupfd(int fd)
int ret;
if (fdd.cache_fd == 0)
fdd.cache_fd = fdd.next_alloc++;
ret = fcntl(fd, F_DUPFD, fdd.cache_fd);
if (ret != -1) {
fdd.cache_fd = fdd.cache_tab[ret];
return ret;
else {
return fd;

void fdcache_closefd(int fd)
if (fd == -1)


fdd.cache_tab[fd] = fdd.cache_fd;
fdd.cache_fd = fd;
 \ /
  Last update: 2005-09-15 22:44    [W:0.063 / U:1.944 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site