Messages in this thread |  | | From | Alexander Shishkin <> | Subject | Re: [PATCH v0 04/71] itrace: Infrastructure for instruction flow tracing units | Date | Wed, 18 Dec 2013 16:01:04 +0200 |
| |
Peter Zijlstra <peterz@infradead.org> writes:
> On Wed, Dec 18, 2013 at 03:23:41PM +0200, Alexander Shishkin wrote: >> Peter Zijlstra <peterz@infradead.org> writes: >> >> > On Wed, Dec 11, 2013 at 02:36:16PM +0200, Alexander Shishkin wrote: >> >> Instruction tracing PMUs are capable of recording a log of instruction >> >> execution flow on a cpu core, which can be useful for profiling and crash >> >> analysis. This patch adds itrace infrastructure for perf events and the >> >> rest of the kernel to use. >> >> >> >> Since such PMUs can produce copious amounts of trace data, it may be >> >> impractical to process it inside the kernel in real time, but instead export >> >> raw trace streams to userspace for subsequent analysis. Thus, itrace PMUs >> >> may export their trace buffers, which can be mmap()ed to userspace from a >> >> perf event fd with a PERF_EVENT_ITRACE_OFFSET offset. To that end, perf >> >> is extended to work with multiple ring buffers per event, reusing the >> >> ring_buffer code in an attempt to reduce complexity. >> > >> > Please read the thread here: https://lkml.org/lkml/2008/12/4/64 >> > >> > On my thoughts of this creative mmap() usage. >> >> That's unfortunate, it made sense to me. But let's then have a look at >> the alternative approaches. Bearing in mind that it is crucial for us to >> export trace buffers to userspace as opposed to processing the trace >> data in the kernel, the fact that we still need the normal perf data >> stream and your dislike for mmap trickery, we need two separate file >> descriptors: one for the perf data and one for the trace data. > > Why don't you start by explaining _why_ you need a second stream to > begin with?
Oh, I'm sure I've explained it earlier ([1], [2]), but why not. The data in the second stream is generated at a rate which is hundreds of megabytes per second per core. Decoding this data is ~1000 times slower than generating it. Ergo, can't be done in kernel, needs to be exported as-is to userspace for later retreival and decoding. Doing it via perf stream means an extra copy, which at these rates is a waste. Ergo, a second buffer.
[1] https://lkml.org/lkml/2013/12/11/213 [2] https://lkml.org/lkml/2013/12/11/358
Regards, -- Alex
|  |