On 5/28/20 5:06 PM, Song Liu wrote:

On Thu, May 28, 2020 at 3:18 AM Michal Soltys [off-list ref] wrote:

This is slow because we are processing 4kB at a time. But it is slower than
my expectation. What are the speeds measured at boot? Something like the
following in dmesg:

[    3.851064] raid6: avx2x4   gen() 17328 MB/s
[    3.877404] raid6: avx2x4   xor() 11033 MB/s
[    3.903410] raid6: avx2x2   gen() 14710 MB/s
[    3.929410] raid6: avx2x2   xor()  8921 MB/s
[    3.955408] raid6: avx2x1   gen() 12687 MB/s
[    3.981405] raid6: avx2x1   xor()  8414 MB/s
[    4.007409] raid6: sse2x4   gen()  9105 MB/s
[    4.033403] raid6: sse2x4   xor()  5724 MB/s
[    4.059408] raid6: sse2x2   gen()  7570 MB/s
[    4.085410] raid6: sse2x2   xor()  4880 MB/s
[    4.111409] raid6: sse2x1   gen()  6484 MB/s
[    4.137402] raid6: sse2x1   xor()  4412 MB/s

These are results for this machine:

[   17.707948] raid6: sse2x4   gen() 12196 MB/s
[   17.871946] raid6: sse2x4   xor()  8169 MB/s
[   18.039947] raid6: sse2x2   gen() 10035 MB/s
[   18.195950] raid6: sse2x2   xor()  6954 MB/s
[   18.343951] raid6: sse2x1   gen()  7983 MB/s
[   18.495948] raid6: sse2x1   xor()  6089 MB/s
[   18.577606] raid6: using algorithm sse2x4 gen() 12196 MB/s
[   18.675268] raid6: .... xor() 8169 MB/s, rmw enabled
[   18.774428] raid6: using ssse3x2 recovery algorithm

For the next step, could you please try the following the beginning of
--assemble
run?

    trace.py -M 10 'r::r5l_recovery_verify_data_checksum()(retval != 0)'

Thanks,
Song

Will do and report.

PS: Some key functions here are inlined and thus not easily traceable. We will
probably need to patch the kernel for further debugging.

Won't be a problem.