: Get perf counters working on D-cache aliasing cpus.
cover : Get perf counters working on D-cache aliasing cpus.
With Jen's Axboe's basic sparc counter patch in my tree I started playing with it on sparc64. Turns out it won't work on anything pre-Niagara.
The issue is D-cache aliasing between the kernel side mapping and the user side mapping of the perf event ring buffer.
The following two patches attempt to address this issue and are working properly on my test machine.
Quoted 9 lines · David Miller
With Jen's Axboe's basic sparc counter patch in my tree I started playing with it on sparc64. Turns out it won't work on anything pre-Niagara.
The issue is D-cache aliasing between the kernel side mapping and the user side mapping of the perf event ring buffer.
The following two patches attempt to address this issue and are working properly on my test machine.
Oh, great!
The only (small) worry would be that this uses vmalloc space on 32-bit platforms which is generally a scarce resource.
IMHO it's not a big issue as the typical workflow would be to have just a handful of perf ring-buffers around and most of the performance work happens in the 64-bit space anyway.
One practical complication is lots of tracepoints and perf trace:
but we can solve that via output buffer multiplexing:PERF_COUNTER_IOC_SET_OUTPUT - which is implemented in tip:perfcounters/core but not yet propagated to the tools. That would be desirable to do anyway, to enable time-coherent trace entries, etc.
Ingo
From: Ingo Molnar <[email protected]>
Date: Tue, 8 Sep 2009 14:36:10 +0200
Quoted 2 lines
The only (small) worry would be that this uses vmalloc space on 32-bit platforms which is generally a scarce resource.
Yes, I considered this too.
But as you said there are things that mitigate how much of a real issue this actually is.
Quoted 9 lines · David Miller
With Jen's Axboe's basic sparc counter patch in my tree I started playing with it on sparc64. Turns out it won't work on anything pre-Niagara.
The issue is D-cache aliasing between the kernel side mapping and the user side mapping of the perf event ring buffer.
The following two patches attempt to address this issue and are working properly on my test machine.
Dave,
Ingo just reminded me that we might want to do splice support for perf stuff. My plan was to have the splice thing allocate a new page, flip with a filled one from the buffer and send the filled one down to the splice consumer.
Now having vmap'ed all that complicates stuff enourmously.
Would it also work if we use order-1 pages on your platform instead of order-0, so that they are properly aligned for the d-cache?
From: Peter Zijlstra <[email protected]>
Date: Fri, 18 Sep 2009 14:31:36 +0200
Quoted 9 lines
Ingo just reminded me that we might want to do splice support for perf stuff. My plan was to have the splice thing allocate a new page, flip with a filled one from the buffer and send the filled one down to the splice consumer.
Now having vmap'ed all that complicates stuff enourmously.
Would it also work if we use order-1 pages on your platform instead of order-0, so that they are properly aligned for the d-cache?
You still have to make sure the order-1 page is order-1 virtually aligned in userspace. But that would help only sparc64 because SHMLBA happens to be 16K (2 * smallest supported PAGE_SIZE).
So, this won't handle sparc32 where the SHMLBA is 4MB.
MIPS has the same exact problems and has a large SHMLBA too.
Quoted 13 lines · David Miller
From: Peter Zijlstra <[email protected]>
Date: Fri, 18 Sep 2009 14:31:36 +0200Quoted 9 lines
Ingo just reminded me that we might want to do splice support for perf stuff. My plan was to have the splice thing allocate a new page, flip with a filled one from the buffer and send the filled one down to the splice consumer.
Now having vmap'ed all that complicates stuff enourmously.
Would it also work if we use order-1 pages on your platform instead of order-0, so that they are properly aligned for the d-cache?
Quoted 7 lines
You still have to make sure the order-1 page is order-1 virtually aligned in userspace. But that would help only sparc64 because SHMLBA happens to be 16K (2 * smallest supported
PAGE_SIZE).So, this won't handle sparc32 where the SHMLBA is 4MB.
MIPS has the same exact problems and has a large SHMLBA too.
Ah, ok, vmalloc it is then. I'll have to come up with another scheme for splice() then.
Thanks!