On Sun, 28 Jan 2001, Marcelo Tosatti wrote:
> >
> > This is the smoking gun here, I bet, but I'd like to make sure I see the
> > whole thing. I don't see _why_ we'd have deadlocked on __wait_on_page(),
> > but I think this is the thread that hangs on to the mm semaphore.
>
> I was able to reproduce it here with dbench.
>
> Nothing is locked except this dbench thread (the only dbench thread):
>
> dbench D C1C9FE64 5200 1013 1 (L-TLB) 1370 785
> Call Trace: [___wait_on_page+130/160] [truncate_list_pages+100/404]
>[truncate_inode_pages+93/128] [iput+162/360] [dput+262/356] [fput+121/232]
>[exit_mmap+218/292]
> [mmput+56/80] [do_exit+208/680] [do_signal+566/656] [dput+25/356]
>[path_release+13/60] [sys_newstat+100/112] [sys_read+188/196] [signal_return+20/24]
Ok, this definitely seems to be the pattern.
I don't see _what_ is going on, though.
I know of one "known bug" in pre10: if you run out of swap-space with
shared memory segments, it will do the wrong thing (return 1 without
unlocking the page). xmms might trigger this, but I didn't think that
dbench used shared memory?
There's also an ugliness in the truncate ordering. I don't think it should
matter, but I do believe it's conceptually wrong as-is.
Does this patch make any difference at all?
Linus
-----
diff -u --recursive --new-file pre10/linux/mm/memory.c linux/mm/memory.c
--- pre10/linux/mm/memory.c Sat Jan 27 10:53:39 2001
+++ linux/mm/memory.c Sat Jan 27 19:12:35 2001
@@ -945,7 +945,6 @@
if (inode->i_size < offset)
goto do_expand;
inode->i_size = offset;
- truncate_inode_pages(mapping, offset);
spin_lock(&mapping->i_shared_lock);
if (!mapping->i_mmap && !mapping->i_mmap_shared)
goto out_unlock;
@@ -960,8 +959,7 @@
out_unlock:
spin_unlock(&mapping->i_shared_lock);
- /* this should go into ->truncate */
- inode->i_size = offset;
+ truncate_inode_pages(mapping, offset);
if (inode->i_op && inode->i_op->truncate)
inode->i_op->truncate(inode);
return;
diff -u --recursive --new-file pre10/linux/mm/shmem.c linux/mm/shmem.c
--- pre10/linux/mm/shmem.c Sat Jan 27 10:53:39 2001
+++ linux/mm/shmem.c Sat Jan 27 19:50:08 2001
@@ -217,8 +217,11 @@
info = &page->mapping->host->u.shmem_i;
swap = __get_swap_page(2);
- if (!swap.val)
- return 1;
+ if (!swap.val) {
+ set_page_dirty(page);
+ UnlockPage(page);
+ return -ENOMEM;
+ }
spin_lock(&info->lock);
shmem_recalc_inode(page->mapping->host);
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
Please read the FAQ at http://www.tux.org/lkml/
