> Yeah, semaphores can't be used from hardirq contexts for much the same
> reasons. But its all ia64 code, right? So I'm not directly seeing how
> other archs are affected here.
The root of the problem is a call chain like this:
[<a00000010005c200>] ia64_global_tlb_purge+0x540/0xa40
sp=e0000001c28afca0 bsp=e0000001c28a0ea0
[<a00000010005c8a0>] flush_tlb_range+0x1a0/0x240
sp=e0000001c28afca0 bsp=e0000001c28a0e70
[<a000000100105a90>] page_referenced_one+0x170/0x260
sp=e0000001c28afca0 bsp=e0000001c28a0e20
[<a000000100105f20>] page_referenced+0x180/0x320
sp=e0000001c28afcb0 bsp=e0000001c28a0dd0
[<a0000001000ec640>] shrink_active_list+0x520/0xbc0
sp=e0000001c28afcc0 bsp=e0000001c28a0d48
[<a0000001000ece30>] shrink_zone+0x150/0x200
sp=e0000001c28afd80 bsp=e0000001c28a0d00
[<a0000001000edaf0>] kswapd+0x690/0x940
sp=e0000001c28afd80 bsp=e0000001c28a0c40
[<a0000001000a6860>] kthread+0xa0/0x120
sp=e0000001c28afe30 bsp=e0000001c28a0c10
[<a000000100013b70>] kernel_thread_helper+0x30/0x60
sp=e0000001c28afe30 bsp=e0000001c28a0be0
[<a0000001000089c0>] start_kernel_thread+0x20/0x40
sp=e0000001c28afe30 bsp=e0000001c28a0be0
Some place in the upper (arch independent) parts of this trace the
code acquires an anon_vma_lock and holds it while calling into arch
specific code (there is presumably some inlining going on here because
the source doesn't have an obvious call from page_referenced_one() to
flush_tlb_range()).
My concern for other architectures (especially ones that require IPI to
complete the TLB shootdown) is that holding anon_vma_lock while doing
the shootdown leaves them open for some deadlock scenarios in this
code path. But perhaps I completley misunderstand how IPI shootdown
happens (very possible).
-Tony
--
To unsubscribe from this list: send the line "unsubscribe linux-ia64" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
|