Donate to e Foundation | Murena handsets with /e/OS | Own a part of Murena! Learn more

Commit cee2688e authored by yonatanc's avatar yonatanc Committed by Doug Ledford
Browse files

IB/rxe: Offload CRC calculation when possible 



Use CPU ability to perform CRC calculations, by
replacing direct calls to crc32_le() with crypto_shash_updata().

The overall performance gain measured with ib_send_bw tool is 10% and it
was tested on "Intel CPU ES-2660 v2 @ 2.20Ghz" CPU.

ib_send_bw -d rxe0  -x 1 -n 9000 -e  -s $((1024 * 1024 )) -l 100

---------------------------------------------------------------------------------------------
|             | bytes   | iterations | BW peak[MB/sec] | BW average[MB/sec] | MsgRate[Mpps] |
---------------------------------------------------------------------------------------------
| crc32_le    | 1048576 | 9000       | inf             | 497.60             | 0.000498      |
| CRC offload | 1048576 | 9000       | inf             | 546.70             | 0.000547      |
---------------------------------------------------------------------------------------------

Fixes: 8700e3e7 ("Soft RoCE driver")
Signed-off-by: default avatarYonatan Cohen <yonatanc@mellanox.com>
Signed-off-by: default avatarLeon Romanovsky <leon@kernel.org>
Signed-off-by: default avatarDoug Ledford <dledford@redhat.com>
parent 0d38ac8a

drivers/infiniband/sw/rxe/Kconfig

+1 −0
Original line number Diff line number Diff line
@@ -2,6 +2,7 @@ config RDMA_RXE 
	tristate "Software RDMA over Ethernet (RoCE) driver"

	depends on INET && PCI && INFINIBAND

	depends on NET_UDP_TUNNEL

	depends on CRYPTO_CRC32

	select DMA_VIRT_OPS

	---help---

	This driver implements the InfiniBand RDMA transport over

drivers/infiniband/sw/rxe/rxe.c

+2 −0
Original line number Diff line number Diff line
@@ -64,6 +64,8 @@ static void rxe_cleanup(struct rxe_dev *rxe) 
	rxe_pool_cleanup(&rxe->mc_elem_pool);



	rxe_cleanup_ports(rxe);



	crypto_free_shash(rxe->tfm);

}



/* called when all references have been dropped */

drivers/infiniband/sw/rxe/rxe.h

+20 −0
Original line number Diff line number Diff line
@@ -50,6 +50,7 @@ 
#include <rdma/ib_umem.h>

#include <rdma/ib_cache.h>

#include <rdma/ib_addr.h>

#include <crypto/hash.h>



#include "rxe_net.h"

#include "rxe_opcode.h"
@@ -64,6 +65,25 @@ 


#define RXE_ROCE_V2_SPORT		(0xc000)



static inline u32 rxe_crc32(struct rxe_dev *rxe,

			    u32 crc, void *next, size_t len)

{

	int err;



	SHASH_DESC_ON_STACK(shash, rxe->tfm);



	shash->tfm = rxe->tfm;

	shash->flags = 0;

	*(u32 *)shash_desc_ctx(shash) = crc;

	err = crypto_shash_update(shash, next, len);

	if (unlikely(err)) {

		pr_warn_ratelimited("failed crc calculation, err: %d\n", err);

		return crc32_le(crc, next, len);

	}



	return *(u32 *)shash_desc_ctx(shash);

}



int rxe_set_mtu(struct rxe_dev *rxe, unsigned int dev_mtu);



int rxe_add(struct rxe_dev *rxe, unsigned int mtu);

drivers/infiniband/sw/rxe/rxe_icrc.c

+3 −3
Original line number Diff line number Diff line
@@ -87,10 +87,10 @@ u32 rxe_icrc_hdr(struct rxe_pkt_info *pkt, struct sk_buff *skb) 
	bth->qpn |= cpu_to_be32(~BTH_QPN_MASK);



	length = hdr_size + RXE_BTH_BYTES;

	crc = crc32_le(crc, pshdr, length);

	crc = rxe_crc32(pkt->rxe, crc, pshdr, length);



	/* And finish to compute the CRC on the remainder of the headers. */

	crc = crc32_le(crc, pkt->hdr + RXE_BTH_BYTES,

	crc = rxe_crc32(pkt->rxe, crc, pkt->hdr + RXE_BTH_BYTES,

			rxe_opcode[pkt->opcode].length - RXE_BTH_BYTES);

	return crc;

}

drivers/infiniband/sw/rxe/rxe_mr.c

+4 −2
Original line number Diff line number Diff line
@@ -370,7 +370,8 @@ int rxe_mem_copy(struct rxe_mem *mem, u64 iova, void *addr, int length, 
			((void *)(uintptr_t)iova) : addr;



		if (crcp)

			*crcp = crc32_le(*crcp, src, length);

			crc = rxe_crc32(to_rdev(mem->pd->ibpd.device),

					*crcp, src, length);



		memcpy(dest, src, length);
@@ -403,7 +404,8 @@ int rxe_mem_copy(struct rxe_mem *mem, u64 iova, void *addr, int length, 
			bytes = length;



		if (crcp)

			crc = crc32_le(crc, src, bytes);

			crc = rxe_crc32(to_rdev(mem->pd->ibpd.device),

					crc, src, bytes);



		memcpy(dest, src, bytes);