Revert "crypto: caam - get rid of tasklet"
authorHoria Geantă <horia.geanta@nxp.com>
Wed, 9 Nov 2016 08:46:21 +0000 (10:46 +0200)
committerHerbert Xu <herbert@gondor.apana.org.au>
Sun, 13 Nov 2016 09:45:13 +0000 (17:45 +0800)
commit2b163b5bce04546da72617bfb6c8bf07a45c4b17
tree478c7c297a4c74a3b2023ec84aed7a95e65a8258
parentc73e36e82b2394b1e199cb975485444cf05ed662
Revert "crypto: caam - get rid of tasklet"

This reverts commit 66d2e2028091a074aa1290d2eeda5ddb1a6c329c.

Quoting from Russell's findings:
https://www.mail-archive.com/linux-crypto@vger.kernel.org/msg21136.html

[quote]
Okay, I've re-tested, using a different way of measuring, because using
openssl speed is impractical for off-loaded engines.  I've decided to
use this way to measure the performance:

dd if=/dev/zero bs=1048576 count=128 | /usr/bin/time openssl dgst -md5

For the threaded IRQs case gives:

0.05user 2.74system 0:05.30elapsed 52%CPU (0avgtext+0avgdata 2400maxresident)k
0.06user 2.52system 0:05.18elapsed 49%CPU (0avgtext+0avgdata 2404maxresident)k
0.12user 2.60system 0:05.61elapsed 48%CPU (0avgtext+0avgdata 2460maxresident)k
=> 5.36s => 25.0MB/s

and the tasklet case:

0.08user 2.53system 0:04.83elapsed 54%CPU (0avgtext+0avgdata 2468maxresident)k
0.09user 2.47system 0:05.16elapsed 49%CPU (0avgtext+0avgdata 2368maxresident)k
0.10user 2.51system 0:04.87elapsed 53%CPU (0avgtext+0avgdata 2460maxresident)k
=> 4.95 => 27.1MB/s

which corresponds to an 8% slowdown for the threaded IRQ case.  So,
tasklets are indeed faster than threaded IRQs.

[...]

I think I've proven from the above that this patch needs to be reverted
due to the performance regression, and that there _is_ most definitely
a deterimental effect of switching from tasklets to threaded IRQs.
[/quote]

Signed-off-by: Horia Geantă <horia.geanta@nxp.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
drivers/crypto/caam/intern.h
drivers/crypto/caam/jr.c