Skip to contentSkip to conversation

siw: Add support for CRC32C offload instruction using libcrypto crc32c-intel

PATCH In review Nicholas A. Bellinger 4 participants 10 messages Sep 22, 2010 – Sep 23, 2010

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 2/2 siw: Add support for CRC32C offload instruction using libcrypto crc32c-intel

From: Nicholas Bellinger <[email protected]>

This patch updates siw_create_qp() to check for the CONFIG_X86 + cpu_has_xmm4_2 dependent use of the CRC32C instruction offload using libcrypto crc32c-intel.ko. This patch will by default use crc32c-intel when available, and fall back to the legacy slicing by 1x libcrypto crc32c.ko code when the instruction offload is not availabe.

Because of the ability to selectively use c_tx->crc_enabled and c_rx->crc_enabled, this patch has to use a handful of less than appealing goto's in order to handle the default case where the offload is enabled by default, the crc32c-intel module is not available and we need to fall back to slicing by 1x crc32c.ko, or we need to fail all together.

Note that this code has only been compile tested so far, but a similar patch to LIO-Target (w/o the selectively use of TX/RX checksums) to use crc32c-intel.ko instruction offload is available here:

http://git.kernel.org/?p=linux/kernel/git/nab/lio-core-2.6.git;a=commitdiff;h=c92ef80235302a1f891f636691c928c3bb4e9ac1

---
 softiwarp/siw_verbs.c |   37 ++++++++++++++++++++++++++++++++++++-
 1 files changed, 36 insertions(+), 1 deletions(-)
softiwarp/siw_verbs.c +36 -1
index 27e2e5e..00ac03c 100644
@@ -348,7 +348,7 @@ struct ib_qp *siw_create_qp(struct ib_pd *ofa_pd, struct ib_qp_init_attr *attrs,
 	struct siw_iwarp_rx		*c_rx;
 	struct siw_uresp_create_qp	uresp;
 
-	int rv = 0;
+	int rv = 0, crc32c_offload = 1;
 
 	dprint(DBG_OBJ|DBG_CM, ": new QP on device %s\n",
 		ofa_dev->name);
@@ -460,6 +460,23 @@ struct ib_qp *siw_create_qp(struct ib_pd *ofa_pd, struct ib_qp_init_attr *attrs,
 	c_tx->crc_enabled = c_rx->crc_enabled = CONFIG_RDMA_SIW_CRC_ENFORCED;
 
 	if (c_tx->crc_enabled) {
+#ifdef CONFIG_X86
+		/*
+		 * Check for the Nehalem optimized crc32c-intel instructions
+		 * This is only currently available while running on bare-metal,
+		 * and is not yet available with QEMU-KVM guests.
+		 */
+		if (cpu_has_xmm4_2 && crc32c_offload) {
+			c_tx->mpa_crc_hd.tfm = crypto_alloc_hash("crc32c-intel",
+						0, CRYPTO_ALG_ASYNC);
+			if (IS_ERR(c_tx->mpa_crc_hd.tfm)) {
+				crc32c_offload = 0;
+				goto check_legacy_tx;
+			}
+			goto check_rx;
+		}
+check_legacy_tx:
+#endif /* CONFIG_X86 */
 		c_tx->mpa_crc_hd.tfm =
 			crypto_alloc_hash("crc32c", 0, CRYPTO_ALG_ASYNC);
 		if (IS_ERR(c_tx->mpa_crc_hd.tfm)) {
@@ -469,7 +486,24 @@ struct ib_qp *siw_create_qp(struct ib_pd *ofa_pd, struct ib_qp_init_attr *attrs,
 			goto remove_qp;
 		}
 	}
+check_rx:
 	if (c_rx->crc_enabled) {
+#ifdef CONFIG_X86
+		/*
+		 * Check for the Nehalem optimized crc32c-intel instructions
+		 * This is only currently available while running on bare-metal,
+		 * and is not yet available with QEMU-KVM guests.
+		 */
+		if (cpu_has_xmm4_2 && crc32c_offload) {
+			c_rx->mpa_crc_hd.tfm = crypto_alloc_hash("crc32c-intel",
+						0, CRYPTO_ALG_ASYNC);
+			if (IS_ERR(c_rx->mpa_crc_hd.tfm))
+				goto check_legacy_rx;
+
+			goto after_crc32c;
+		}
+check_legacy_rx:
+#endif /* CONFIG_X86 */
 		c_rx->mpa_crc_hd.tfm =
 			crypto_alloc_hash("crc32c", 0, CRYPTO_ALG_ASYNC);
 		if (IS_ERR(c_rx->mpa_crc_hd.tfm)) {
@@ -478,6 +512,7 @@ struct ib_qp *siw_create_qp(struct ib_pd *ofa_pd, struct ib_qp_init_attr *attrs,
 			goto remove_qp;
 		}
 	}
+after_crc32c:
 	atomic_set(&qp->tx_ctx.in_use, 0);
 
 	qp->ofa_qp.qp_num = QP_ID(qp);
Signature · 6 lines
1.5.6.5

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#1 lore raw
Quoted 7 lines · Nicholas A. Bellinger

From: Nicholas Bellinger<[email protected]>

This patch updates siw_create_qp() to check for the CONFIG_X86 + cpu_has_xmm4_2 dependent use of the CRC32C instruction offload using libcrypto crc32c-intel.ko. This patch will by default use crc32c-intel when available, and fall back to the legacy slicing by 1x libcrypto crc32c.ko code when the instruction offload is not availabe.

I don't think every caller should handle checks like this. The crypto layer should load the right driver instead and provide the best driver under a generic algorithm name.

Need CPUID module auto probing. I have an older patch that needs some fixes.

-Andi

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#2 lore raw
Quoted 9 lines · Andi Kleen
Quoted 7 lines · Nicholas A. Bellinger

From: Nicholas Bellinger<[email protected]>

This patch updates siw_create_qp() to check for the CONFIG_X86 + cpu_has_xmm4_2 dependent use of the CRC32C instruction offload using libcrypto crc32c-intel.ko. This patch will by default use crc32c-intel when available, and fall back to the legacy slicing by 1x libcrypto crc32c.ko code when the instruction offload is not availabe.

Quoted 4 lines

I don't think every caller should handle checks like this. The crypto layer should load the right driver
instead and provide the best driver under a generic algorithm name.

Indeed, this would clean up the explict RX/TX CRC32C case quite a bit.. Unfortuately I am too busy with other items atm to cook up this patch, but I would be happy to test it if someone wants to take it. ;)

Quoted 2 lines

Need CPUID module auto probing. I have an older patch that needs some fixes.

Hmm, I don't see how that fits in here exactly. Would you mind elaborating a bit..?

Thanks!

--nab

#3 lore raw
Quoted 4 lines · Nicholas A. Bellinger

Indeed, this would clean up the explict RX/TX CRC32C case quite a bit.. Unfortuately I am too busy with other items atm to cook up this patch, but I would be happy to test it if someone wants to take it. ;)

Do you even need to do anything at all? crc32c is provided by crc32c-intel at a higher priority already, so it should be used if it is available..

Quoted 5 lines
Quoted 2 lines

Need CPUID module auto probing. I have an older patch that needs some fixes.

Hmm, I don't see how that fits in here exactly. Would you mind elaborating a bit..?

Unless the module is loaded the optimized algorithm will not be available for automatic use. Maybe your patch causes it to autoload because of the by-name reference? Identifying modules to load by CPUID will let userspace auto load the appropriate ones based on CPU...

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#4 lore raw
Quoted 10 lines · Jason Gunthorpe
Quoted 4 lines · Nicholas A. Bellinger

Indeed, this would clean up the explict RX/TX CRC32C case quite a bit.. Unfortuately I am too busy with other items atm to cook up this patch, but I would be happy to test it if someone wants to take it. ;)

Do you even need to do anything at all? crc32c is provided by crc32c-intel at a higher priority already, so it should be used if it is available..

I believe with the current libcrypto code that consumers are still required to explictly ask for crc32c-intel offload.

Quoted 4 lines
Quoted 3 lines
Quoted 3 lines

Need CPUID module auto probing. I have an older patch that needs some fixes.

Quoted 8 lines
Quoted 2 lines

Hmm, I don't see how that fits in here exactly. Would you mind elaborating a bit..?

Unless the module is loaded the optimized algorithm will not be available for automatic use. Maybe your patch causes it to autoload because of the by-name reference? Identifying modules to load by CPUID will let userspace auto load the appropriate ones based on CPU...

Correct, this patch is so that autoload of crc32c-intel.ko 'just works' and we do the fallback to the legacy slicing by 1x crc32c.ko when the former is not availabe.

But I definately agree with Andi here that we should just add a wrapper around the crypto_alloc_hash() usage of crc32c-intel and crc32c for libcrypto consumers..

Thanks for your comments Jason!

--nab

#5 lore raw

[cc list chopped]

Quoted 8 lines · Nicholas A. Bellinger
Quoted 7 lines · Jason Gunthorpe
Quoted 4 lines · Nicholas A. Bellinger

Indeed, this would clean up the explict RX/TX CRC32C case quite a bit.. Unfortuately I am too busy with other items atm to cook up this patch, but I would be happy to test it if someone wants to take it. ;)

Quoted 6 lines
Quoted 3 lines

Do you even need to do anything at all? crc32c is provided by crc32c-intel at a higher priority already, so it should be used if it is available..

I believe with the current libcrypto code that consumers are still required to explictly ask for crc32c-intel offload.

Really? It all looks OK to me.. What does your /proc/crypto say on a system with crc32c-intel support?

Jason

#6 lore raw
Quoted 15 lines · Jason Gunthorpe

[cc list chopped]

Quoted 12 lines · Nicholas A. Bellinger
Quoted 10 lines · Jason Gunthorpe
Quoted 4 lines · Nicholas A. Bellinger

Indeed, this would clean up the explict RX/TX CRC32C case quite a bit.. Unfortuately I am too busy with other items atm to cook up this patch, but I would be happy to test it if someone wants to take it. ;)

Do you even need to do anything at all? crc32c is provided by crc32c-intel at a higher priority already, so it should be used if it is available..

Quoted 5 lines
Quoted 2 lines

I believe with the current libcrypto code that consumers are still required to explictly ask for crc32c-intel offload.

Really? It all looks OK to me.. What does your /proc/crypto say on a system with crc32c-intel support?

After a fresh boot /proc/crypto looks like:

name : stdrng
driver : krng
module : kernel
priority : 200
refcnt : 1
selftest : passed
type : rng
seedsize : 0

name : crc32c
driver : crc32c-generic
module : kernel
priority : 100
refcnt : 2
selftest : passed
type : shash
blocksize : 1
digestsize : 4

name : sha1
driver : sha1-generic
module : kernel
priority : 0
refcnt : 1
selftest : passed
type : shash
blocksize : 64
digestsize : 20

Once I start up the LIO-Target stack and some iSCSI Initiators login and request crc32c-intel using crypto_alloc_hash() using a method similar to what this patch for Softiwarp does, the following appears in at the top of /proc/crypto:

name : crc32c
driver : crc32c-intel
module : crc32c_intel
priority : 200
refcnt : 5
selftest : passed
type : shash
blocksize : 1
digestsize : 4

So I think the main bit here is the ability to request crc32c-intel.ko first, and then fall back to crc32c.ko when the former is not available on CONFIG_X86.

Best,

--nab

#7 lore raw
Quoted 3 lines · Nicholas A. Bellinger

So I think the main bit here is the ability to request crc32c-intel.ko first, and then fall back to crc32c.ko when the former is not available on CONFIG_X86.

Well, it is what Andi said, everything is working fine but there is no mechanism to autoload the accelerated crypto module. If you did modprobe crc32c_intel prior to loading your driver it would automatically get crc32c-intel when it asks for crc32c since it is loaded and a higher priority.

So, the drivers are correct to just request crc32c .. The work around to limited autoprobing is so trivial (modprob crc32_intel) I'm not sure including extra autoprobing code in the drivers is worthwhile?

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#8 lore raw
Quoted 6 lines · Jason Gunthorpe
Quoted 3 lines · Nicholas A. Bellinger

So I think the main bit here is the ability to request crc32c-intel.ko first, and then fall back to crc32c.ko when the former is not available on CONFIG_X86.

Quoted 6 lines

Well, it is what Andi said, everything is working fine but there is no mechanism to autoload the accelerated crypto module. If you did modprobe crc32c_intel prior to loading your driver it would automatically get crc32c-intel when it asks for crc32c since it is loaded and a higher priority.

Ah, OK. I see what you mean now here wrt to libcrypto priorities and
crc32c + crc32c_intel modules. My apologies for the in-experience with
libcrypto here..

Quoted 4 lines

So, the drivers are correct to just request crc32c .. The work around to limited autoprobing is so trivial (modprob crc32_intel) I'm not sure including extra autoprobing code in the drivers is worthwhile?

Indeed, I am happy to drop this patch if Bernard would be nice enough to add a 'modprobe crc32_intel' into the SIW scripts.

Thanks for your comments Jason!

--nab

#9 lore raw

[email protected] wrote on 09/23/2010 12:36:29 AM:

Quoted 18 lines · Jason Gunthorpe
Quoted 6 lines · Nicholas A. Bellinger
Quoted 4 lines

So I think the main bit here is the ability to request crc32c-intel.ko first, and then fall back to crc32c.ko when the former is not available on CONFIG_X86.

Quoted 6 lines

Well, it is what Andi said, everything is working fine but there is no mechanism to autoload the accelerated crypto module. If you did modprobe crc32c_intel prior to loading your driver it would automatically get crc32c-intel when it asks for crc32c since it is loaded and a higher priority.

Ah, OK. I see what you mean now here wrt to libcrypto priorities and
crc32c + crc32c_intel modules. My apologies for the in-experience with
libcrypto here..

Quoted 8 lines
Quoted 4 lines

So, the drivers are correct to just request crc32c .. The work around to limited autoprobing is so trivial (modprob crc32_intel) I'm not sure including extra autoprobing code in the drivers is worthwhile?

Indeed, I am happy to drop this patch if Bernard would be nice enough to add a 'modprobe crc32_intel' into the SIW scripts.

Ok, thanks for the CRC comments, quite instructive. To sum up, now I'll add
a minimum siw bringup script to the kernel part.

Bernard