Re: smp race fix between invalidate_inode_pages* and do_no_page

From: Nick Piggin
Date: Wed Jan 11 2006 - 04:32:47 EST


Andrea Arcangeli wrote:
On Wed, Jan 11, 2006 at 03:08:31PM +1100, Nick Piggin wrote:

I'd be inclined to think a lock_page is not a big SMP scalability
problem because the struct page's cacheline(s) will be written to
several times in the process of refcounting anyway. Such a workload
would also be running into tree_lock as well.


I seem to recall you wanted to make the tree_lock a readonly lock for
readers for the exact same scalability reason? do_no_page is quite a

I think Bill Irwin or Peter Chubb made the tree_lock a reader-writer
lock back in the day.

I have some patches (ref:lockless pagecache) that completely removes
the tree_lock from read-side operations like find_get_page and
find_lock_page, and turns the write side back into a regular spinlock.
You must be thinking of that?

fast path for the tree lock too. But I totally agree the unavoidable is
the atomic_inc though, good point, so it worth more to remove the
tree_lock than to remove the page lock, the tree_lock can be avoided the
atomic_inc on page->_count not.


Yep, my thinking as well.

The other bonus that makes this attractive is that then we can drop the
*whole* vm_truncate_count mess... vm_truncate_count and
inode->trunate_count exists for the only single reason that do_no_page
must not map into the pte a page that is under truncation. We can
provide the same guarantee with the page lock doing like
invalidate_inode_pages2_range (that is to check page_mapping under the
page_lock and executing unmap_mapping_range with the page lock held if
needed). That will free 4 bytes per vma (without even counting the
truncate_count on every inode out there! that could be an even larger
gain), on my system I have 9191 vmas in use, that's 36K saved of ram in
my system, and that's 36K saved on x86, on x86-64 it's 72K saved of
physical ram since it's an unsigned long after a pointer, and vma must
not be hw aligned (and infact it isn't so the saving is real). On the
indoes side it saves 4 bytes
* 1384 on my current system, on a busy nfs server it can save a lot
more. The inode also most not be hw aligned and correctly it isn't. On a
server with lot more of vmas and lot more of inodes it'll save more ram.

So if I make this change this could give me a grant for lifetime
guarantee of seccomp in the kernel that takes less than 1kbyte on a x86,
right? (on a normal desktop I'll save at minimum 30 times more than what
I cost to the kernel users ;) Just kidding of course...


Sounds like a good idea (and your proposed implementation -
lock_page and recheck mapping in do_no_page sounds sane).

Thanks,
Nick

--
SUSE Labs, Novell Inc.

Send instant messages to your online friends http://au.messenger.yahoo.com -
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/