crypto: Remove CRYPTO_TFM_REQ_MAY_SLEEP flag in AES-NI accelerated ecb/cbc mode
Subsystems (2)
PATCH 2/3 crypto: Remove CRYPTO_TFM_REQ_MAY_SLEEP flag in AES-NI accelerated ecb/cbc mode
Because AES-NI instructions will touch XMM state, corresponding code must be enclosed within kernel_fpu_begin/end, which used preempt_disable/enable. So sleep should be prevented between kernel_fpu_begin/end.
--- arch/x86/crypto/aesni-intel_glue.c | 4 ++++ 1 file changed, 4 insertions(+)
arch/x86/crypto/aesni-intel_glue.c +4 -0
@@ -198,6 +198,7 @@ static int ecb_encrypt(struct blkcipher_blkcipher_walk_init (&walk, dst, src, nbytes); err =blkcipher_walk_virt (desc, &walk); + desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP ;kernel_fpu_begin ();while ((nbytes = walk.nbytes)) {
@@ -221,6 +222,7 @@ static int ecb_decrypt(struct blkcipher_blkcipher_walk_init (&walk, dst, src, nbytes); err =blkcipher_walk_virt (desc, &walk); + desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP ;kernel_fpu_begin ();while ((nbytes = walk.nbytes)) {
@@ -266,6 +268,7 @@ static int cbc_encrypt(struct blkcipher_blkcipher_walk_init (&walk, dst, src, nbytes); err =blkcipher_walk_virt (desc, &walk); + desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP ;kernel_fpu_begin ();while ((nbytes = walk.nbytes)) {
@@ -289,6 +292,7 @@ static int cbc_decrypt(struct blkcipher_blkcipher_walk_init (&walk, dst, src, nbytes); err =blkcipher_walk_virt (desc, &walk); + desc->flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP ;kernel_fpu_begin ();while ((nbytes = walk.nbytes)) {
Herbert Xu
maintainerQuoted 6 lines · Huang Ying
Because AES-NI instructions will touch XMM state, corresponding code must be enclosed within kernel_fpu_begin/end, which used preempt_disable/enable. So sleep should be prevented between kernel_fpu_begin/end.
Signed-off-by Huang Ying <[email protected]>
I'll apply this for now. But it would be much better to allow sleeping in this case.
Couldn't we simply move the kernel_fpu_begin/end inside the loop? It shouldn't be too expensive when it runs the 2nd time since it doesn't have to save anything, right?
Cheer,
Signature · 4 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txtQuoted 6 lines · Herbert Xu
Quoted 5 lines · Huang Ying
Because AES-NI instructions will touch XMM state, corresponding code must be enclosed within kernel_fpu_begin/end, which used preempt_disable/enable. So sleep should be prevented between kernel_fpu_begin/end.
Quoted 8 lines
Quoted 1 line
Signed-off-by Huang Ying <[email protected]>I'll apply this for now. But it would be much better to allow sleeping in this case.
Couldn't we simply move the kernel_fpu_begin/end inside the loop? It shouldn't be too expensive when it runs the 2nd time since it doesn't have to save anything, right?
I will test the performance difference.
Best Regards,
Huang Ying