Hi Seth, Dan, Vitaly, Herbert, Using crypto_comp APIs, zswap is not able to use the hardware accelators which are only ported to cryto_acomp nowadays. So Mahipal Challa tried to solve this problem by the below patch a long time ago: mm: zswap - Add crypto acomp/scomp framework support [1] At that time, the test was based on acomp with scomp backend. It was not a real async platform. On a platform with real acomp support like hisilicon-zip, the patch will lead to serious "sleep on atomic" issues. To leverage the power of hardware accelerator, right now, I am sending a new patch which will remove the atomic context and permit crypto to sleep in zswap. Literally, using an async compressor, people can dynamically allocate acomp_req and queue those requests to acrytp drivers, and finally use the callback to notify the completion of compression/decompression. but this will require dynamic memory allocation and various synchronizations in zswap, and it is too complex. Alternatively, this patch pre-allocates the acomp_req with the same number of CPUs. For each acomp_req, one mutex and one wait are bound with it. The mutex is used for the race protection of the acomp_req and other percpu resources. Even though the preempt-disabled atomic context is replaced by sleepable context, threads might migrate, but the mutex can still protect the race between CPUs for same resources. Tested on hisilicon zip driver on a SMP enviorment and on lz4 scomp-based acomp as well. To use scomp-based acomp, another patch I sent before is needed: crypto: acomp - search acomp with scomp backend in crypto_has_acomp [2] [1] https://www.spinics.net/lists/linux-mm/msg122455.html [2] https://marc.info/?l=linux-crypto-vger&m=158822346227760&w=2 Barry Song (1): mm/zswap: move to use crypto_acomp API for hardware acceleration mm/zswap.c | 150 ++++++++++++++++++++++++++++++++++++++--------------- 1 file changed, 108 insertions(+), 42 deletions(-) -- 2.23.0
