Re: [PATCH net-next v2 1/2] net/smc: support SMC-R V2 for rdma devices with max_recv_sge equals to 1

[Wenjia Zhang <wenjia@linux.ibm.com>]

On 02.12.24 13:52, Guangguan Wang wrote:
> For SMC-R V2, llc msg can be larger than SMC_WR_BUF_SIZE, thus every
> recv wr has 2 sges, the first sge with length SMC_WR_BUF_SIZE is for
> V1/V2 compatible llc/cdc msg, and the second sge with length
> SMC_WR_BUF_V2_SIZE-SMC_WR_TX_SIZE is for V2 specific llc msg, like
> SMC_LLC_DELETE_RKEY and SMC_LLC_ADD_LINK for SMC-R V2. The memory
> buffer in the second sge is shared by all recv wr in one link and
> all link in one lgr for saving memory usage purpose.
> 
> But not all RDMA devices with max_recv_sge greater than 1. Thus SMC-R
> V2 can not support on such RDMA devices and SMC_CLC_DECL_INTERR fallback
> happens because of the failure of create qp.
> 
> This patch introduce the support for SMC-R V2 on RDMA devices with
> max_recv_sge equals to 1. Every recv wr has only one sge with individual
> buffer whose size is SMC_WR_BUF_V2_SIZE once the RDMA device's max_recv_sge
> equals to 1. It may use more memory, but it is better than
> SMC_CLC_DECL_INTERR fallback.
> 
find good!
> Co-developed-by: Wen Gu <guwen@linux.alibaba.com>
> Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
> Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>
> ---
>   net/smc/smc_core.c |  5 +++++
>   net/smc/smc_core.h | 11 ++++++++++-
>   net/smc/smc_ib.c   |  3 +--
>   net/smc/smc_llc.c  | 21 +++++++++++++++------
>   net/smc/smc_wr.c   | 42 +++++++++++++++++++++---------------------
>   5 files changed, 52 insertions(+), 30 deletions(-)
> 
> diff --git a/net/smc/smc_core.c b/net/smc/smc_core.c
> index 500952c2e67b..ede4d5f3111b 100644
> --- a/net/smc/smc_core.c
> +++ b/net/smc/smc_core.c
> @@ -795,9 +795,14 @@ int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
>   	if (lgr->smc_version == SMC_V2) {
>   		lnk->smcibdev = ini->smcrv2.ib_dev_v2;
>   		lnk->ibport = ini->smcrv2.ib_port_v2;
> +		lnk->wr_rx_sge_cnt = lnk->smcibdev->ibdev->attrs.max_recv_sge < 2 ? 1 : 2;
> +		lnk->wr_rx_buflen = smc_link_shared_v2_rxbuf(lnk) ?
> +			SMC_WR_BUF_SIZE : SMC_WR_BUF_V2_SIZE;
>   	} else {
>   		lnk->smcibdev = ini->ib_dev;
>   		lnk->ibport = ini->ib_port;
> +		lnk->wr_rx_sge_cnt = 1;
> +		lnk->wr_rx_buflen = SMC_WR_BUF_SIZE;
>   	}
>   	get_device(&lnk->smcibdev->ibdev->dev);
>   	atomic_inc(&lnk->smcibdev->lnk_cnt);
> diff --git a/net/smc/smc_core.h b/net/smc/smc_core.h
> index 69b54ecd6503..48a1b1dcb576 100644
> --- a/net/smc/smc_core.h
> +++ b/net/smc/smc_core.h
> @@ -122,10 +122,14 @@ struct smc_link {
>   	} ____cacheline_aligned_in_smp;
>   	struct completion	tx_ref_comp;
>   
> -	struct smc_wr_buf	*wr_rx_bufs;	/* WR recv payload buffers */
> +	u8			*wr_rx_bufs;	/* WR recv payload buffers */
>   	struct ib_recv_wr	*wr_rx_ibs;	/* WR recv meta data */
>   	struct ib_sge		*wr_rx_sges;	/* WR recv scatter meta data */
>   	/* above three vectors have wr_rx_cnt elements and use the same index */
> +	int			wr_rx_sge_cnt; /* rx sge, V1 is 1, V2 is either 2 or 1 */
> +	int			wr_rx_buflen;	/* buffer len for the first sge, len for the
> +						 * second sge is lgr shared if rx sge is 2.
> +						 */
>   	dma_addr_t		wr_rx_dma_addr;	/* DMA address of wr_rx_bufs */
>   	dma_addr_t		wr_rx_v2_dma_addr; /* DMA address of v2 rx buf*/
>   	u64			wr_rx_id;	/* seq # of last recv WR */
> @@ -506,6 +510,11 @@ static inline bool smc_link_active(struct smc_link *lnk)
>   	return lnk->state == SMC_LNK_ACTIVE;
>   }
>   
> +static inline bool smc_link_shared_v2_rxbuf(struct smc_link *lnk)
> +{
> +	return lnk->wr_rx_sge_cnt > 1;
> +}
> +
>   static inline void smc_gid_be16_convert(__u8 *buf, u8 *gid_raw)
>   {
>   	sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x",
> diff --git a/net/smc/smc_ib.c b/net/smc/smc_ib.c
> index 9c563cdbea90..53828833a3f7 100644
> --- a/net/smc/smc_ib.c
> +++ b/net/smc/smc_ib.c
> @@ -662,7 +662,6 @@ void smc_ib_destroy_queue_pair(struct smc_link *lnk)
>   /* create a queue pair within the protection domain for a link */
>   int smc_ib_create_queue_pair(struct smc_link *lnk)
>   {
> -	int sges_per_buf = (lnk->lgr->smc_version == SMC_V2) ? 2 : 1;
>   	struct ib_qp_init_attr qp_attr = {
>   		.event_handler = smc_ib_qp_event_handler,
>   		.qp_context = lnk,
> @@ -676,7 +675,7 @@ int smc_ib_create_queue_pair(struct smc_link *lnk)
>   			.max_send_wr = SMC_WR_BUF_CNT * 3,
>   			.max_recv_wr = SMC_WR_BUF_CNT * 3,
>   			.max_send_sge = SMC_IB_MAX_SEND_SGE,
> -			.max_recv_sge = sges_per_buf,
> +			.max_recv_sge = lnk->wr_rx_sge_cnt,
>   			.max_inline_data = 0,
>   		},
>   		.sq_sig_type = IB_SIGNAL_REQ_WR,
> diff --git a/net/smc/smc_llc.c b/net/smc/smc_llc.c
> index 018ce8133b02..f865c58c3aa7 100644
> --- a/net/smc/smc_llc.c
> +++ b/net/smc/smc_llc.c
> @@ -997,13 +997,14 @@ static int smc_llc_cli_conf_link(struct smc_link *link,
>   }
>   
>   static void smc_llc_save_add_link_rkeys(struct smc_link *link,
> -					struct smc_link *link_new)
> +					struct smc_link *link_new,
> +					u8 *llc_msg)
>   {
>   	struct smc_llc_msg_add_link_v2_ext *ext;
>   	struct smc_link_group *lgr = link->lgr;
>   	int max, i;
>   
> -	ext = (struct smc_llc_msg_add_link_v2_ext *)((u8 *)lgr->wr_rx_buf_v2 +
> +	ext = (struct smc_llc_msg_add_link_v2_ext *)(llc_msg +
>   						     SMC_WR_TX_SIZE);
>   	max = min_t(u8, ext->num_rkeys, SMC_LLC_RKEYS_PER_MSG_V2);
>   	down_write(&lgr->rmbs_lock);
> @@ -1098,7 +1099,9 @@ int smc_llc_cli_add_link(struct smc_link *link, struct smc_llc_qentry *qentry)
>   	if (rc)
>   		goto out_clear_lnk;
>   	if (lgr->smc_version == SMC_V2) {
> -		smc_llc_save_add_link_rkeys(link, lnk_new);
> +		u8 *llc_msg = smc_link_shared_v2_rxbuf(link) ?
> +			(u8 *)lgr->wr_rx_buf_v2 : (u8 *)llc;
> +		smc_llc_save_add_link_rkeys(link, lnk_new, llc_msg);
>   	} else {
>   		rc = smc_llc_cli_rkey_exchange(link, lnk_new);
>   		if (rc) {
> @@ -1498,7 +1501,9 @@ int smc_llc_srv_add_link(struct smc_link *link,
>   	if (rc)
>   		goto out_err;
>   	if (lgr->smc_version == SMC_V2) {
> -		smc_llc_save_add_link_rkeys(link, link_new);
> +		u8 *llc_msg = smc_link_shared_v2_rxbuf(link) ?
> +			(u8 *)lgr->wr_rx_buf_v2 : (u8 *)add_llc;
> +		smc_llc_save_add_link_rkeys(link, link_new, llc_msg);
>   	} else {
>   		rc = smc_llc_srv_rkey_exchange(link, link_new);
>   		if (rc)
> @@ -1807,8 +1812,12 @@ static void smc_llc_rmt_delete_rkey(struct smc_link_group *lgr)
>   	if (lgr->smc_version == SMC_V2) {
>   		struct smc_llc_msg_delete_rkey_v2 *llcv2;
>   
> -		memcpy(lgr->wr_rx_buf_v2, llc, sizeof(*llc));
> -		llcv2 = (struct smc_llc_msg_delete_rkey_v2 *)lgr->wr_rx_buf_v2;
> +		if (smc_link_shared_v2_rxbuf(link)) {
> +			memcpy(lgr->wr_rx_buf_v2, llc, sizeof(*llc));
> +			llcv2 = (struct smc_llc_msg_delete_rkey_v2 *)lgr->wr_rx_buf_v2;
> +		} else {
> +			llcv2 = (struct smc_llc_msg_delete_rkey_v2 *)llc;
> +		}
>   		llcv2->num_inval_rkeys = 0;
>   
>   		max = min_t(u8, llcv2->num_rkeys, SMC_LLC_RKEYS_PER_MSG_V2);
> diff --git a/net/smc/smc_wr.c b/net/smc/smc_wr.c
> index 994c0cd4fddb..b04a21b8c511 100644
> --- a/net/smc/smc_wr.c
> +++ b/net/smc/smc_wr.c
> @@ -439,7 +439,7 @@ static inline void smc_wr_rx_demultiplex(struct ib_wc *wc)
>   		return; /* short message */
>   	temp_wr_id = wc->wr_id;
>   	index = do_div(temp_wr_id, link->wr_rx_cnt);
> -	wr_rx = (struct smc_wr_rx_hdr *)&link->wr_rx_bufs[index];
> +	wr_rx = (struct smc_wr_rx_hdr *)(link->wr_rx_bufs + index * link->wr_rx_buflen);
>   	hash_for_each_possible(smc_wr_rx_hash, handler, list, wr_rx->type) {
>   		if (handler->type == wr_rx->type)
>   			handler->handler(wc, wr_rx);
> @@ -555,7 +555,6 @@ void smc_wr_remember_qp_attr(struct smc_link *lnk)
>   
>   static void smc_wr_init_sge(struct smc_link *lnk)
>   {
> -	int sges_per_buf = (lnk->lgr->smc_version == SMC_V2) ? 2 : 1;
>   	bool send_inline = (lnk->qp_attr.cap.max_inline_data > SMC_WR_TX_SIZE);
>   	u32 i;
>   
> @@ -608,13 +607,14 @@ static void smc_wr_init_sge(struct smc_link *lnk)
>   	 * the larger spillover buffer, allowing easy data mapping.
>   	 */
>   	for (i = 0; i < lnk->wr_rx_cnt; i++) {
> -		int x = i * sges_per_buf;
> +		int x = i * lnk->wr_rx_sge_cnt;
>   
>   		lnk->wr_rx_sges[x].addr =
> -			lnk->wr_rx_dma_addr + i * SMC_WR_BUF_SIZE;
> -		lnk->wr_rx_sges[x].length = SMC_WR_TX_SIZE;
> +			lnk->wr_rx_dma_addr + i * lnk->wr_rx_buflen;
> +		lnk->wr_rx_sges[x].length = smc_link_shared_v2_rxbuf(lnk) ?
> +			SMC_WR_TX_SIZE : lnk->wr_rx_buflen;
>   		lnk->wr_rx_sges[x].lkey = lnk->roce_pd->local_dma_lkey;
> -		if (lnk->lgr->smc_version == SMC_V2) {
> +		if (lnk->lgr->smc_version == SMC_V2 && smc_link_shared_v2_rxbuf(lnk)) {
>   			lnk->wr_rx_sges[x + 1].addr =
>   					lnk->wr_rx_v2_dma_addr + SMC_WR_TX_SIZE;
>   			lnk->wr_rx_sges[x + 1].length =
> @@ -624,7 +624,7 @@ static void smc_wr_init_sge(struct smc_link *lnk)
>   		}
>   		lnk->wr_rx_ibs[i].next = NULL;
>   		lnk->wr_rx_ibs[i].sg_list = &lnk->wr_rx_sges[x];
> -		lnk->wr_rx_ibs[i].num_sge = sges_per_buf;
> +		lnk->wr_rx_ibs[i].num_sge = lnk->wr_rx_sge_cnt;
>   	}
>   	lnk->wr_reg.wr.next = NULL;
>   	lnk->wr_reg.wr.num_sge = 0;
> @@ -655,7 +655,7 @@ void smc_wr_free_link(struct smc_link *lnk)
>   
>   	if (lnk->wr_rx_dma_addr) {
>   		ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
> -				    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
> +				    lnk->wr_rx_buflen * lnk->wr_rx_cnt,
>   				    DMA_FROM_DEVICE);
>   		lnk->wr_rx_dma_addr = 0;
>   	}
> @@ -740,13 +740,11 @@ int smc_wr_alloc_lgr_mem(struct smc_link_group *lgr)
>   
>   int smc_wr_alloc_link_mem(struct smc_link *link)
>   {
> -	int sges_per_buf = link->lgr->smc_version == SMC_V2 ? 2 : 1;
> -
>   	/* allocate link related memory */
>   	link->wr_tx_bufs = kcalloc(SMC_WR_BUF_CNT, SMC_WR_BUF_SIZE, GFP_KERNEL);
>   	if (!link->wr_tx_bufs)
>   		goto no_mem;
> -	link->wr_rx_bufs = kcalloc(SMC_WR_BUF_CNT * 3, SMC_WR_BUF_SIZE,
> +	link->wr_rx_bufs = kcalloc(SMC_WR_BUF_CNT * 3, link->wr_rx_buflen,
>   				   GFP_KERNEL);
>   	if (!link->wr_rx_bufs)
>   		goto no_mem_wr_tx_bufs;
> @@ -774,7 +772,7 @@ int smc_wr_alloc_link_mem(struct smc_link *link)
>   	if (!link->wr_tx_sges)
>   		goto no_mem_wr_tx_rdma_sges;
>   	link->wr_rx_sges = kcalloc(SMC_WR_BUF_CNT * 3,
> -				   sizeof(link->wr_rx_sges[0]) * sges_per_buf,
> +				   sizeof(link->wr_rx_sges[0]) * link->wr_rx_sge_cnt,
>   				   GFP_KERNEL);
>   	if (!link->wr_rx_sges)
>   		goto no_mem_wr_tx_sges;
> @@ -872,7 +870,7 @@ int smc_wr_create_link(struct smc_link *lnk)
>   	smc_wr_tx_set_wr_id(&lnk->wr_tx_id, 0);
>   	lnk->wr_rx_id = 0;
>   	lnk->wr_rx_dma_addr = ib_dma_map_single(
> -		ibdev, lnk->wr_rx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
> +		ibdev, lnk->wr_rx_bufs,	lnk->wr_rx_buflen * lnk->wr_rx_cnt,
>   		DMA_FROM_DEVICE);
>   	if (ib_dma_mapping_error(ibdev, lnk->wr_rx_dma_addr)) {
>   		lnk->wr_rx_dma_addr = 0;
> @@ -880,13 +878,15 @@ int smc_wr_create_link(struct smc_link *lnk)
>   		goto out;
>   	}
>   	if (lnk->lgr->smc_version == SMC_V2) {
> -		lnk->wr_rx_v2_dma_addr = ib_dma_map_single(ibdev,
> -			lnk->lgr->wr_rx_buf_v2, SMC_WR_BUF_V2_SIZE,
> -			DMA_FROM_DEVICE);
> -		if (ib_dma_mapping_error(ibdev, lnk->wr_rx_v2_dma_addr)) {
> -			lnk->wr_rx_v2_dma_addr = 0;
> -			rc = -EIO;
> -			goto dma_unmap;
> +		if (smc_link_shared_v2_rxbuf(lnk)) {
> +			lnk->wr_rx_v2_dma_addr =
> +				ib_dma_map_single(ibdev, lnk->lgr->wr_rx_buf_v2,
> +						  SMC_WR_BUF_V2_SIZE, DMA_FROM_DEVICE);
> +			if (ib_dma_mapping_error(ibdev, lnk->wr_rx_v2_dma_addr)) {
> +				lnk->wr_rx_v2_dma_addr = 0;
> +				rc = -EIO;
> +				goto dma_unmap;
> +			}
>   		}
>   		lnk->wr_tx_v2_dma_addr = ib_dma_map_single(ibdev,
>   			lnk->lgr->wr_tx_buf_v2, SMC_WR_BUF_V2_SIZE,
> @@ -935,7 +935,7 @@ int smc_wr_create_link(struct smc_link *lnk)
>   		lnk->wr_tx_v2_dma_addr = 0;
>   	}
>   	ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
> -			    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
> +			    lnk->wr_rx_buflen * lnk->wr_rx_cnt,
>   			    DMA_FROM_DEVICE);
>   	lnk->wr_rx_dma_addr = 0;
>   out:
It looks good for me!

Reviewed-by: Wenjia Zhang <wenjia@linux.ibm.com>

Thanks,
Wenjia