What protects cpu_tlbstate?
Hi,
What protects the cpu_tlbstate? I see in i386/kernel/smp.c that its always used in a non-preemptable area, but what prevents races with interrupts? For example, what prevents leave_mm() called via the flush_tlb_all IPI from racing with, say, enter_lazy_tlb? Couldn't a race leave cpu_tlbstate in an inconsistent state?
Or does it simply not matter? But if that were true, it seems to me that there should be at least some barriers or something to make sure the final state is consistent.
Thanks,
J
Quoted 11 lines · On Thursday 05 April 2007 21:44:10 Jeremy Fitzhardinge wrote:
Hi,
What protects the
cpu_tlbstate? I see ini386/kernel/smp.cthat its always used in a non-preemptable area, but what prevents races with interrupts? For example, what preventsleave_mm()called via theflush_tlb_allIPI from racing with, say,enter_lazy_tlb? Couldn't a race leavecpu_tlbstatein an inconsistent state?Or does it simply not matter? But if that were true, it seems to me that there should be at least some barriers or something to make sure the final state is consistent.
The interrupts can only happen when the other CPU is already lazy and enter_lazy_tlb would be a nop then. The flushers itself are synchronized by the page_table_lock or the mm semaphore.
Against switch_mm it tries to protect with ordering.
wmb()s are not needed on x86 (ok minus errata on ppro and VIA magic mode but which is UP only). That would leave some rmb()s, but I don't see any place they would be needed.
-Andi
Quoted 10 lines · Andi Kleen
The interrupts can only happen when the other CPU is already lazy and
enter_lazy_tlbwould be a nop then. The flushers itself are synchronized by thepage_table_lockor the mm semaphore.Against
switch_mmit tries to protect with ordering.wmb()s are not needed on x86 (ok minus errata on ppro and VIA magic mode but which is UP only). That would leave some rmb()s, but I don't see any place they would be needed.
Hm, I was more wondering about simple compiler reordering. Does the relative order of setting and reading cpu_tlbstate.state, active_mm and the mm->cpu_vm_mask matter?
J
Quoted 7 lines · On Thursday 05 April 2007 23:00:22 Jeremy Fitzhardinge wrote:
Quoted 6 lines · Andi Kleen
The interrupts can only happen when the other CPU is already lazy and
enter_lazy_tlbwould be a nop then. The flushers itself are synchronized by thepage_table_lockor the mm semaphore.Against
switch_mmit tries to protect with ordering.
Quoted 8 lines
Quoted 4 lines
wmb()s are not needed on x86 (ok minus errata on ppro and VIA magic mode but which is UP only). That would leave some rmb()s, but I don't see any place they would be needed.
Hm, I was more wondering about simple compiler reordering. Does the relative order of setting and reading cpu_tlbstate.state,
active_mmand themm->cpu_vm_maskmatter?
Hmm, perhaps a barrier between state and active_mm might be a good idea. Setting active_mm after state might be problematic. cpu_vm_mask should be already a memory barrier.
-Andi
Quoted 9 lines · Andi Kleen
Quoted 4 lines
Hm, I was more wondering about simple compiler reordering. Does the relative order of setting and reading cpu_tlbstate.state,
active_mmand themm->cpu_vm_maskmatter?Hmm, perhaps a barrier between state and
active_mmmight be a good idea. Settingactive_mmafter state might be problematic.cpu_vm_maskshould be already a memory barrier.
Should leave_mm update active_mm and set the state to TLBSTATE_OK? Just reloading cr3 seems a bit blunt.
J