SEC1 doesn't support S/G in descriptors so for hash operations,
the CPU has to build a buffer containing the buffered block and
the incoming data. This generates a lot of memory copies which
represents more than 50% of CPU time of a md5sum operation as
shown below with a 'perf record'.
|--86.24%-- kcapi_md_digest
| |
| |--86.18%-- _kcapi_common_vmsplice_chunk_fd
| | |
| | |--83.68%-- splice
| | | |
| | | |--83.59%-- ret_from_syscall
| | | | |
| | | | |--83.52%-- sys_splice
| | | | | |
| | | | | |--83.49%-- splice_from_pipe
| | | | | | |
| | | | | | |--83.04%-- __splice_from_pipe
| | | | | | | |
| | | | | | | |--80.67%-- pipe_to_sendpage
| | | | | | | | |
| | | | | | | | |--78.25%-- hash_sendpage
| | | | | | | | | |
| | | | | | | | | |--60.08%-- ahash_process_req
| | | | | | | | | | |
| | | | | | | | | | |--56.36%-- sg_copy_buffer
| | | | | | | | | | | |
| | | | | | | | | | | |--55.29%-- memcpy
| | | | | | | | | | | |
However, unlike SEC2+, SEC1 offers the possibility to chain
descriptors. It is therefore possible to build a first descriptor
pointing to the buffered data and a second descriptor pointing to
the incoming data, hence avoiding the memory copy to a single
buffer.
With this patch, the time necessary for a md5sum on a 90Mbytes file
is approximately 3 seconds. Without the patch it takes 6 seconds.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
37b5e8897e