Re: [bug] SLOB, tipc_init(), WARNING: at arch/x86/mm/highmem_32.c:52 kmap_atomic_prot()

[Matt Mackall <mpm@selenic.com>]

On Thu, Nov 29, 2007 at 03:39:20PM +0100, Ingo Molnar wrote:
> 
> * Ingo Molnar <mingo@elte.hu> wrote:
> 
> > i'm getting this on 32-bit (with the kmap-atomic debugging patch 
> > applied):
> > 
> >  ---------------->
> >  Calling initcall 0x78b67c00: tipc_init+0x0/0xc0()
> >  TIPC: Activated (version 1.6.2 compiled Nov 29 2007 15:04:36)
> >  WARNING: at arch/x86/mm/highmem_32.c:52 kmap_atomic_prot()
> >  Pid: 1, comm: swapper Not tainted 2.6.24-rc3-cfs-v24 #45
> >   [<78107272>] show_trace_log_lvl+0x12/0x40
> >   [<781072ad>] show_trace+0xd/0x20
> >   [<781086f8>] dump_stack+0x58/0x60
> >   [<7811541f>] kmap_atomic_prot+0x1bf/0x240
> >   [<781154ae>] kmap_atomic+0xe/0x20
> >   [<78157be5>] get_page_from_freelist+0x225/0x420
> >   [<78157e4d>] __alloc_pages+0x6d/0x3a0
> >   [<78169a7b>] slob_new_page+0x1b/0x60
> >   [<78169be4>] slob_alloc+0x124/0x1e0
> >   [<78169e0f>] __kmalloc_node+0x6f/0xa0
> >   [<7884f7c2>] reg_init+0x42/0x80
> >   [<7884f80a>] tipc_reg_start+0xa/0x40
> >   [<7883e6c6>] tipc_core_start+0x66/0xc0
> >   [<78b67c81>] tipc_init+0x81/0xc0
> 
> this is due to the kzalloc() here:
> 
> 0x7884f1d0 is in reg_init (net/tipc/user_reg.c:88).
> 83              spin_lock_bh(&reg_lock);
> 84              if (!users) {
> 85                      users = kzalloc(USER_LIST_SIZE, GFP_ATOMIC);
> 86                      if (users) {
> 87                              for (i = 1; i <= MAX_USERID; i++) {
> 88                                      users[i].next = i - 1;
> 
> which does a:
> 
> 120     static inline void clear_highpage(struct page *page)
> 121     {
> 122             void *kaddr = kmap_atomic(page, KM_USER0);
> 123             clear_page(kaddr);
> 124             kunmap_atomic(kaddr, KM_USER0);
> 125     }
> 
> but ... why does the debug code think it's in softirq context?
> 
> plus, and this is a slob question i guess, how come we drop into 
> clear_highpage() for a kzalloc()??

Good question. Looks like kzalloc switched from doing a memset to
passing a GFP_ZERO flag down to kmalloc. Slob didn't get completely
updated to reflect this, so it blindly propagates the flag onto
__alloc_pages and does a harmless double-clear.

Someone should remind us what the point of moving the kzalloc memset
down into the allocators was. We now have all three allocators doing:

        if (unlikely((flags & __GFP_ZERO) && ptr))
                memset(ptr, 0, obj_size(cachep));

and needing to mask flags before passing them to page allocators,
which hardly seems better than kzalloc unconditionally doing the
memset. Wouldn't it be better/faster/smaller to make kzalloc a
non-inline?

Slob also has a nice second path for large kmallocs where it just
calls the page allocator directly which also needs this treatment.
Which does the right thing with non-highmem systems, but can hit this
bug otherwise.

Below is a totally untested patch. Alternately, we could simply tweak
clear_highpage to remove the limitation, but that would leave slob
doing an extra clear.

diff -r c60016ba6237 mm/slob.c
--- a/mm/slob.c	Tue Nov 13 09:09:36 2007 -0800
+++ b/mm/slob.c	Thu Nov 29 15:15:54 2007 -0600
@@ -223,6 +231,7 @@ static void *slob_new_page(gfp_t gfp, in
 {
 	void *page;
 
+	gfp &= ~__GFP_ZERO;
 #ifdef CONFIG_NUMA
 	if (node != -1)
 		page = alloc_pages_node(node, gfp, order);
@@ -457,6 +470,8 @@ void *__kmalloc_node(size_t size, gfp_t 
 			page = virt_to_page(ret);
 			page->private = size;
 		}
+		if (unlikely((gfp & __GFP_ZERO) && ret))
+			memset(ret, 0, size);
 		return ret;
 	}
 }

-- 
Mathematics is the supreme nostalgia of our time.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/