x86_64 page fault NMI-safe
in [Kernel]
|
Prev: [PATCH] enable readback to get HPET working on ATI SB4x00, kernel 2.6.35_rc5
Next: input: Fix wrong dimensions check for synaptics
From: Mathieu Desnoyers on 14 Jul 2010 19:20 * Frederic Weisbecker (fweisbec(a)gmail.com) wrote: > On Wed, Jul 14, 2010 at 06:31:07PM -0400, Mathieu Desnoyers wrote: > > * Frederic Weisbecker (fweisbec(a)gmail.com) wrote: > > > On Wed, Jul 14, 2010 at 12:54:19PM -0700, Linus Torvalds wrote: > > > > On Wed, Jul 14, 2010 at 12:36 PM, Frederic Weisbecker > > > > <fweisbec(a)gmail.com> wrote: > > > > > > > > > > There is also the fact we need to handle the lost NMI, by defering its > > > > > treatment or so. That adds even more complexity. > > > > > > > > I don't think your read my proposal very deeply. It already handles > > > > them by taking a fault on the iret of the first one (that's why we > > > > point to the stack frame - so that we can corrupt it and force a > > > > fault). > > > > > > > > > Ah right, I missed this part. > > > > Hrm, Frederic, I hate to ask that but.. what are you doing with those percpu 8k > > data structures exactly ? :) > > > > Mathieu > > > > So, when an event triggers in perf, we sometimes want to capture the stacktrace > that led to the event. > > We want this stacktrace (here we call that a callchain) to be recorded > locklessly. So we want this callchain buffer per cpu, with the following > type: Ah OK, so you mean that perf now has 2 different ring buffer implementations ? How about using a single one that is generic enough to handle perf and ftrace needs instead ? (/me runs away quickly before the lightning strikes) ;) Mathieu > > #define PERF_MAX_STACK_DEPTH 255 > > struct perf_callchain_entry { > __u64 nr; > __u64 ip[PERF_MAX_STACK_DEPTH]; > }; > > > That makes 2048 bytes. But per cpu is not enough for the callchain to be recorded > locklessly, we also need one buffer per context: task, softirq, hardirq, nmi, as > an event can trigger in any of these. > Since we disable preemption, none of these contexts can nest locally. In > fact hardirqs can nest but we just don't care about this corner case. > > So, it makes 2048 * 4 = 8192 bytes. And that per cpu. > -- Mathieu Desnoyers Operating System Efficiency R&D Consultant EfficiOS Inc. http://www.efficios.com -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Linus Torvalds on 14 Jul 2010 19:30 On Wed, Jul 14, 2010 at 4:09 PM, Andi Kleen <andi(a)firstfloor.org> wrote: > > It can happen in theory, but for such a rare case take a lock > and walking everything should be fine. Actually, that's _exactly_ the wrong kind of thinking. Bad latency is bad latency, even when it happens rarely. So latency problems kill - even when they are rare. So you want to avoid them. And walking every possible page table is a _huge_ latency problem when it happens. In contrast, what's the advantage of doing thigns synchronously while holding a lock? It's that you can avoid a few page faults, and get better CPU use. But that's _stupid_ if it's something that is very rare to begin with. So the very rarity argues for the lazy approach. If it wasn't rare, there would be a much stronger argument for trying to do things up-front. Linus -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Tejun Heo on 14 Jul 2010 19:40 Hello, On 07/14/2010 10:08 PM, H. Peter Anvin wrote: >> I suspect the low level per cpu allocation functions should >> just call it. >> > > Yes, specifically the point at which we allocate new per cpu memory > blocks. We can definitely do that but walking whole page table list is too heavy to do automatically at that level especially when all users other than NMI would be fine w/ the default lazy approach. If Linus' approach doesn't pan out, I think the right thing to do would be adding a wrapper to vmalloc_sync_all() and let perf code call it after percpu allocation. Thanks. -- tejun -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Frederic Weisbecker on 14 Jul 2010 19:40 On Wed, Jul 14, 2010 at 07:11:17PM -0400, Mathieu Desnoyers wrote: > * Frederic Weisbecker (fweisbec(a)gmail.com) wrote: > > On Wed, Jul 14, 2010 at 06:31:07PM -0400, Mathieu Desnoyers wrote: > > > * Frederic Weisbecker (fweisbec(a)gmail.com) wrote: > > > > On Wed, Jul 14, 2010 at 12:54:19PM -0700, Linus Torvalds wrote: > > > > > On Wed, Jul 14, 2010 at 12:36 PM, Frederic Weisbecker > > > > > <fweisbec(a)gmail.com> wrote: > > > > > > > > > > > > There is also the fact we need to handle the lost NMI, by defering its > > > > > > treatment or so. That adds even more complexity. > > > > > > > > > > I don't think your read my proposal very deeply. It already handles > > > > > them by taking a fault on the iret of the first one (that's why we > > > > > point to the stack frame - so that we can corrupt it and force a > > > > > fault). > > > > > > > > > > > > Ah right, I missed this part. > > > > > > Hrm, Frederic, I hate to ask that but.. what are you doing with those percpu 8k > > > data structures exactly ? :) > > > > > > Mathieu > > > > > > > > So, when an event triggers in perf, we sometimes want to capture the stacktrace > > that led to the event. > > > > We want this stacktrace (here we call that a callchain) to be recorded > > locklessly. So we want this callchain buffer per cpu, with the following > > type: > > Ah OK, so you mean that perf now has 2 different ring buffer implementations ? > How about using a single one that is generic enough to handle perf and ftrace > needs instead ? > > (/me runs away quickly before the lightning strikes) ;) > > Mathieu :-) That's no ring buffer. It's a temporary linear buffer to fill a stacktrace, and get its effective size before committing it to the real ring buffer. Sure that involves two copies. But I don't have a better solution in mind than using a pre-buffer for that, since we can't know the size of the stacktrace in advance. We could always reserve the max stacktrace size, but that would be wasteful. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Steven Rostedt on 14 Jul 2010 19:50
[ /me removes the duplicate email of himself! ] On Wed, 2010-07-14 at 19:11 -0400, Mathieu Desnoyers wrote: > > > > So, when an event triggers in perf, we sometimes want to capture the stacktrace > > that led to the event. > > > > We want this stacktrace (here we call that a callchain) to be recorded > > locklessly. So we want this callchain buffer per cpu, with the following > > type: > > Ah OK, so you mean that perf now has 2 different ring buffer implementations ? > How about using a single one that is generic enough to handle perf and ftrace > needs instead ? > > (/me runs away quickly before the lightning strikes) ;) > To be fair, that's just a temp buffer. -- Steve (/me sends this to try to remove the dup email he's getting ) -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ |