A low-level (packet-oriented) API would handle these issues and allow a higher-level (message-oriented) API to emit messages as a sequence of packets.
IIUC, writing STM device directly can fulfill this request, like below:
/* write the data via STM device file */ write(fd, (char*)trace_data, sizeof(unsigned int) * TEST_DATA_SIZE);But I think it may be not fast enough for your requirement.
Yes you're write that if you have a large amount of data you could write it out using write() to the STM device driver, but the overhead would depend on the size of the message.
The cost of going via a syscall is 1 syscall overhead (~1000ns?) plus ~30ns per STM packet.
The cost of going via a low-level userspace API is ~30ns per packet.
So you would have to be writing hundreds of packets in one call, for the syscall overhead to become insignificant. Even some of the general-purpose higher level message protocols are designed to only use about 10 packets per message - that would be ~300ns userspace, ~1300 with write(). For me the whole attraction of STM is that you can write messages 'little and often' and they get individual timestamps and channel identifiers. You could do quite a lot with just one packet per message.
Apart from this way, what else I can think out is to implement a user space wrapper program based on the mmap() system call I have finished in kernel space.
Ok I'll look forward to it.
We'll be putting the CoreSight Access Library up on ARM's public github area soon, and that might be a way to share ideas for an STM userspace API.
Al
Thanks, Chunyan
I'm wondering what's the best way for me to put my proposal up for review. Does Linaro have a Confluence system?
Al
Thanks, Chunyan
[1] https://git.linaro.org/people/zhang.chunyan/linux.git/shortlog/refs/h eads/stm- mmap-test-v4.6 [2] https://pastebin.linaro.org/view/21995dfa
On Fri, Apr 1, 2016 at 5:41 PM, Ola Liljedahl Ola.Liljedahl@arm.com
wrote:
On 01/04/2016, 11:21, "CoreSight on behalf of Al Grant" <coresight-bounces@lists.linaro.org on behalf of Al.Grant@arm.com>
wrote:
Ah, right, using this address (map) directly has some thing wrong for now, I haven't found the root cause. Now I have to focus the work on getting STM driver upstreamed first, and then I will take a look at what caused the problem. I will get back to you once STM mmap() interface works well.
Ok I look forward to it. I think a userspace write API could be quite useful for low-overhead instrumentation. On my Cortex-A57 server I can write a 32-bit STM timestamped data item, guaranteed delivery, from userspace every 30ns. It's not as fast as writing an untimestamped item to a ring buffer in local cache (~1ns) but it's less disruptive on cache, and a lot faster than calling clock_gettime() and writing a timestamped item to a ring buffer (~250ns). And I'm fairly sure that I could make the same API work on Intel STH too although I haven't got one to try out.
I approve of this.
‹ Ola
Al _______________________________________________ CoreSight mailing list CoreSight@lists.linaro.org https://lists.linaro.org/mailman/listinfo/coresight