[PATCH] uml: sigio code - reduce spinlock hold time
authorPaolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Thu, 19 Jan 2006 01:42:57 +0000 (17:42 -0800)
committerLinus Torvalds <torvalds@g5.osdl.org>
Thu, 19 Jan 2006 03:20:21 +0000 (19:20 -0800)
In a previous patch I shifted an allocation to being atomic.

In this patch, a better but more intrusive solution is implemented, i.e.  hold
the lock only when really needing it, especially not over pipe operations, nor
over the culprit allocation.

Additionally, while at it, add a missing kfree in the failure path, and make
sure that if we fail in forking, write_sigio_pid is -1 and not, say, -ENOMEM.

And fix whitespace, at least for things I was touching anyway.

Signed-off-by: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Cc: Jeff Dike <jdike@addtoit.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
arch/um/kernel/sigio_user.c

index 62e5cfdf21881ca1d8bad038e497a7026afa2f6e..f7b18e157d3509d9b4302b1dc8dfd39c02756630 100644 (file)
@@ -337,70 +337,103 @@ int ignore_sigio_fd(int fd)
        return(err);
 }
 
-static int setup_initial_poll(int fd)
+static struct pollfd* setup_initial_poll(int fd)
 {
        struct pollfd *p;
 
-       p = um_kmalloc_atomic(sizeof(struct pollfd));
-       if(p == NULL){
+       p = um_kmalloc(sizeof(struct pollfd));
+       if (p == NULL) {
                printk("setup_initial_poll : failed to allocate poll\n");
-               return(-1);
+               return NULL;
        }
        *p = ((struct pollfd) { .fd     = fd,
                                .events         = POLLIN,
                                .revents        = 0 });
-       current_poll = ((struct pollfds) { .poll        = p,
-                                          .used        = 1,
-                                          .size        = 1 });
-       return(0);
+       return p;
 }
 
 void write_sigio_workaround(void)
 {
        unsigned long stack;
+       struct pollfd *p;
        int err;
+       int l_write_sigio_fds[2];
+       int l_sigio_private[2];
+       int l_write_sigio_pid;
 
+       /* We call this *tons* of times - and most ones we must just fail. */
        sigio_lock();
-       if(write_sigio_pid != -1)
-               goto out;
+       l_write_sigio_pid = write_sigio_pid;
+       sigio_unlock();
 
-       err = os_pipe(write_sigio_fds, 1, 1);
+       if (l_write_sigio_pid != -1)
+               return;
+
+       err = os_pipe(l_write_sigio_fds, 1, 1);
        if(err < 0){
                printk("write_sigio_workaround - os_pipe 1 failed, "
                       "err = %d\n", -err);
-               goto out;
+               return;
        }
-       err = os_pipe(sigio_private, 1, 1);
+       err = os_pipe(l_sigio_private, 1, 1);
        if(err < 0){
-               printk("write_sigio_workaround - os_pipe 2 failed, "
+               printk("write_sigio_workaround - os_pipe 1 failed, "
                       "err = %d\n", -err);
                goto out_close1;
        }
-       if(setup_initial_poll(sigio_private[1]))
+
+       p = setup_initial_poll(l_sigio_private[1]);
+       if(!p)
                goto out_close2;
 
-       write_sigio_pid = run_helper_thread(write_sigio_thread, NULL, 
+       sigio_lock();
+
+       /* Did we race? Don't try to optimize this, please, it's not so likely
+        * to happen, and no more than once at the boot. */
+       if(write_sigio_pid != -1)
+               goto out_unlock;
+
+       write_sigio_pid = run_helper_thread(write_sigio_thread, NULL,
                                            CLONE_FILES | CLONE_VM, &stack, 0);
 
-       if(write_sigio_pid < 0) goto out_close2;
+       if (write_sigio_pid < 0)
+               goto out_clear;
 
-       if(write_sigio_irq(write_sigio_fds[0])) 
+       if (write_sigio_irq(l_write_sigio_fds[0]))
                goto out_kill;
 
- out:
+       /* Success, finally. */
+       memcpy(write_sigio_fds, l_write_sigio_fds, sizeof(l_write_sigio_fds));
+       memcpy(sigio_private, l_sigio_private, sizeof(l_sigio_private));
+
+       current_poll = ((struct pollfds) { .poll        = p,
+                                          .used        = 1,
+                                          .size        = 1 });
+
        sigio_unlock();
        return;
 
  out_kill:
-       os_kill_process(write_sigio_pid, 1);
+       l_write_sigio_pid = write_sigio_pid;
        write_sigio_pid = -1;
+       sigio_unlock();
+       /* Going to call waitpid, avoid holding the lock. */
+       os_kill_process(l_write_sigio_pid, 1);
+       goto out_free;
+
+ out_clear:
+       write_sigio_pid = -1;
+ out_unlock:
+       sigio_unlock();
+ out_free:
+       kfree(p);
  out_close2:
-       os_close_file(sigio_private[0]);
-       os_close_file(sigio_private[1]);
+       os_close_file(l_sigio_private[0]);
+       os_close_file(l_sigio_private[1]);
  out_close1:
-       os_close_file(write_sigio_fds[0]);
-       os_close_file(write_sigio_fds[1]);
-       sigio_unlock();
+       os_close_file(l_write_sigio_fds[0]);
+       os_close_file(l_write_sigio_fds[1]);
+       return;
 }
 
 int read_sigio_fd(int fd)