diff mbox series

[v2,04/26] qedf: Simplify s/g list mapping.

Message ID 20190326073858.26792-5-skashyap@marvell.com (mailing list archive)
State Accepted
Headers show
Series qedf: Misc fixes for the driver. | expand

Commit Message

Saurav Kashyap March 26, 2019, 7:38 a.m. UTC
From: Chad Dupuis <cdupuis@marvell.com>

When mapping the pages from a scatter/gather list from the SCSI layer we
only need to follow these rules:

o Max SGEs for each I/O request is 256
o No size limit on each SGE
o No need to split OS provided SGEs to 4K before sending to firmware.
o Slow SGE is applicable only when:
  - There are > 8 SGEs and any middle SGE is less than a page size (4K)

Make necessary changes so that driver follows these rules.
applicable only for Write requests (not for Read requests)
No need to check SGE address alignment requirements
(first, middle or last) before declaring slow SGE

Signed-off-by: Chad Dupuis <cdupuis@marvell.com>
Signed-off-by: Saurav Kashyap <skashyap@marvell.com>
---
 drivers/scsi/qedf/qedf.h         |   8 +--
 drivers/scsi/qedf/qedf_debugfs.c |   2 -
 drivers/scsi/qedf/qedf_io.c      | 124 ++++++++++++---------------------------
 3 files changed, 40 insertions(+), 94 deletions(-)
diff mbox series

Patch

diff --git a/drivers/scsi/qedf/qedf.h b/drivers/scsi/qedf/qedf.h
index 8e75c21..4080af8 100644
--- a/drivers/scsi/qedf/qedf.h
+++ b/drivers/scsi/qedf/qedf.h
@@ -102,7 +102,6 @@  struct qedf_ioreq {
 	struct list_head link;
 	uint16_t xid;
 	struct scsi_cmnd *sc_cmd;
-	bool use_slowpath; /* Use slow SGL for this I/O */
 #define QEDF_SCSI_CMD		1
 #define QEDF_TASK_MGMT_CMD	2
 #define QEDF_ABTS		3
@@ -152,9 +151,9 @@  struct qedf_ioreq {
 	int fp_idx;
 	unsigned int cpu;
 	unsigned int int_cpu;
-#define QEDF_IOREQ_SLOW_SGE		0
-#define QEDF_IOREQ_SINGLE_SGE		1
-#define QEDF_IOREQ_FAST_SGE		2
+#define QEDF_IOREQ_UNKNOWN_SGE		1
+#define QEDF_IOREQ_SLOW_SGE		2
+#define QEDF_IOREQ_FAST_SGE		3
 	u8 sge_type;
 	struct delayed_work rrq_work;
 
@@ -366,7 +365,6 @@  struct qedf_ctx {
 
 	u32 slow_sge_ios;
 	u32 fast_sge_ios;
-	u32 single_sge_ios;
 
 	uint8_t	*grcdump;
 	uint32_t grcdump_size;
diff --git a/drivers/scsi/qedf/qedf_debugfs.c b/drivers/scsi/qedf/qedf_debugfs.c
index a32d8ee..6ae78dd 100644
--- a/drivers/scsi/qedf/qedf_debugfs.c
+++ b/drivers/scsi/qedf/qedf_debugfs.c
@@ -303,7 +303,6 @@ 
 	seq_printf(s, "cmg_mgr free io_reqs: %d\n",
 	    atomic_read(&qedf->cmd_mgr->free_list_cnt));
 	seq_printf(s, "slow SGEs: %d\n", qedf->slow_sge_ios);
-	seq_printf(s, "single SGEs: %d\n", qedf->single_sge_ios);
 	seq_printf(s, "fast SGEs: %d\n\n", qedf->fast_sge_ios);
 
 	seq_puts(s, "Offloaded ports:\n\n");
@@ -361,7 +360,6 @@ 
 
 	/* Clear stat counters exposed by 'stats' node */
 	qedf->slow_sge_ios = 0;
-	qedf->single_sge_ios = 0;
 	qedf->fast_sge_ios = 0;
 
 	return count;
diff --git a/drivers/scsi/qedf/qedf_io.c b/drivers/scsi/qedf/qedf_io.c
index da728d2..ed38df9 100644
--- a/drivers/scsi/qedf/qedf_io.c
+++ b/drivers/scsi/qedf/qedf_io.c
@@ -428,29 +428,6 @@  void qedf_release_cmd(struct kref *ref)
 	clear_bit(QEDF_CMD_OUTSTANDING, &io_req->flags);
 }
 
-static int qedf_split_bd(struct qedf_ioreq *io_req, u64 addr, int sg_len,
-	int bd_index)
-{
-	struct scsi_sge *bd = io_req->bd_tbl->bd_tbl;
-	int frag_size, sg_frags;
-
-	sg_frags = 0;
-	while (sg_len) {
-		if (sg_len > QEDF_BD_SPLIT_SZ)
-			frag_size = QEDF_BD_SPLIT_SZ;
-		else
-			frag_size = sg_len;
-		bd[bd_index + sg_frags].sge_addr.lo = U64_LO(addr);
-		bd[bd_index + sg_frags].sge_addr.hi = U64_HI(addr);
-		bd[bd_index + sg_frags].sge_len = (uint16_t)frag_size;
-
-		addr += (u64)frag_size;
-		sg_frags++;
-		sg_len -= frag_size;
-	}
-	return sg_frags;
-}
-
 static int qedf_map_sg(struct qedf_ioreq *io_req)
 {
 	struct scsi_cmnd *sc = io_req->sc_cmd;
@@ -462,75 +439,45 @@  static int qedf_map_sg(struct qedf_ioreq *io_req)
 	int byte_count = 0;
 	int sg_count = 0;
 	int bd_count = 0;
-	int sg_frags;
-	unsigned int sg_len;
+	u32 sg_len;
 	u64 addr, end_addr;
-	int i;
+	int i = 0;
 
 	sg_count = dma_map_sg(&qedf->pdev->dev, scsi_sglist(sc),
 	    scsi_sg_count(sc), sc->sc_data_direction);
-
 	sg = scsi_sglist(sc);
 
-	/*
-	 * New condition to send single SGE as cached-SGL with length less
-	 * than 64k.
-	 */
-	if ((sg_count == 1) && (sg_dma_len(sg) <=
-	    QEDF_MAX_SGLEN_FOR_CACHESGL)) {
-		sg_len = sg_dma_len(sg);
-		addr = (u64)sg_dma_address(sg);
-
-		bd[bd_count].sge_addr.lo = (addr & 0xffffffff);
-		bd[bd_count].sge_addr.hi = (addr >> 32);
-		bd[bd_count].sge_len = (u16)sg_len;
+	io_req->sge_type = QEDF_IOREQ_UNKNOWN_SGE;
 
-		return ++bd_count;
-	}
+	if (sg_count <= 8 || io_req->io_req_flags == QEDF_READ)
+		io_req->sge_type = QEDF_IOREQ_FAST_SGE;
 
 	scsi_for_each_sg(sc, sg, sg_count, i) {
-		sg_len = sg_dma_len(sg);
+		sg_len = (u32)sg_dma_len(sg);
 		addr = (u64)sg_dma_address(sg);
 		end_addr = (u64)(addr + sg_len);
 
 		/*
-		 * First s/g element in the list so check if the end_addr
-		 * is paged aligned. Also check to make sure the length is
-		 * at least page size.
-		 */
-		if ((i == 0) && (sg_count > 1) &&
-		    ((end_addr % QEDF_PAGE_SIZE) ||
-		    sg_len < QEDF_PAGE_SIZE))
-			io_req->use_slowpath = true;
-		/*
-		 * Last s/g element so check if the start address is paged
-		 * aligned.
-		 */
-		else if ((i == (sg_count - 1)) && (sg_count > 1) &&
-		    (addr % QEDF_PAGE_SIZE))
-			io_req->use_slowpath = true;
-		/*
 		 * Intermediate s/g element so check if start and end address
-		 * is page aligned.
+		 * is page aligned.  Only required for writes and only if the
+		 * number of scatter/gather elements is 8 or more.
 		 */
-		else if ((i != 0) && (i != (sg_count - 1)) &&
-		    ((addr % QEDF_PAGE_SIZE) || (end_addr % QEDF_PAGE_SIZE)))
-			io_req->use_slowpath = true;
+		if (io_req->sge_type == QEDF_IOREQ_UNKNOWN_SGE && (i) &&
+		    (i != (sg_count - 1)) && sg_len < QEDF_PAGE_SIZE)
+			io_req->sge_type = QEDF_IOREQ_SLOW_SGE;
 
-		if (sg_len > QEDF_MAX_BD_LEN) {
-			sg_frags = qedf_split_bd(io_req, addr, sg_len,
-			    bd_count);
-		} else {
-			sg_frags = 1;
-			bd[bd_count].sge_addr.lo = U64_LO(addr);
-			bd[bd_count].sge_addr.hi  = U64_HI(addr);
-			bd[bd_count].sge_len = (uint16_t)sg_len;
-		}
+		bd[bd_count].sge_addr.lo = cpu_to_le32(U64_LO(addr));
+		bd[bd_count].sge_addr.hi  = cpu_to_le32(U64_HI(addr));
+		bd[bd_count].sge_len = cpu_to_le32(sg_len);
 
-		bd_count += sg_frags;
+		bd_count++;
 		byte_count += sg_len;
 	}
 
+	/* To catch a case where FAST and SLOW nothing is set, set FAST */
+	if (io_req->sge_type == QEDF_IOREQ_UNKNOWN_SGE)
+		io_req->sge_type = QEDF_IOREQ_FAST_SGE;
+
 	if (byte_count != scsi_bufflen(sc))
 		QEDF_ERR(&(qedf->dbg_ctx), "byte_count = %d != "
 			  "scsi_bufflen = %d, task_id = 0x%x.\n", byte_count,
@@ -655,8 +602,10 @@  static void  qedf_init_task(struct qedf_rport *fcport, struct fc_lport *lport,
 		io_req->sgl_task_params->num_sges = bd_count;
 		io_req->sgl_task_params->total_buffer_size =
 		    scsi_bufflen(io_req->sc_cmd);
-		io_req->sgl_task_params->small_mid_sge =
-			io_req->use_slowpath;
+		if (io_req->sge_type == QEDF_IOREQ_SLOW_SGE)
+			io_req->sgl_task_params->small_mid_sge = 1;
+		else
+			io_req->sgl_task_params->small_mid_sge = 0;
 	}
 
 	/* Fill in physical address of sense buffer */
@@ -679,16 +628,10 @@  static void  qedf_init_task(struct qedf_rport *fcport, struct fc_lport *lport,
 				    io_req->task_retry_identifier, fcp_cmnd);
 
 	/* Increment SGL type counters */
-	if (bd_count == 1) {
-		qedf->single_sge_ios++;
-		io_req->sge_type = QEDF_IOREQ_SINGLE_SGE;
-	} else if (io_req->use_slowpath) {
+	if (io_req->sge_type == QEDF_IOREQ_SLOW_SGE)
 		qedf->slow_sge_ios++;
-		io_req->sge_type = QEDF_IOREQ_SLOW_SGE;
-	} else {
+	else
 		qedf->fast_sge_ios++;
-		io_req->sge_type = QEDF_IOREQ_FAST_SGE;
-	}
 }
 
 void qedf_init_mp_task(struct qedf_ioreq *io_req,
@@ -770,9 +713,6 @@  void qedf_init_mp_task(struct qedf_ioreq *io_req,
 						     &task_fc_hdr,
 						     &tx_sgl_task_params,
 						     &rx_sgl_task_params, 0);
-
-	/* Midpath requests always consume 1 SGE */
-	qedf->single_sge_ios++;
 }
 
 /* Presumed that fcport->rport_lock is held */
@@ -872,7 +812,7 @@  int qedf_post_io_req(struct qedf_rport *fcport, struct qedf_ioreq *io_req)
 	/* Initialize rest of io_req fileds */
 	io_req->data_xfer_len = scsi_bufflen(sc_cmd);
 	sc_cmd->SCp.ptr = (char *)io_req;
-	io_req->use_slowpath = false; /* Assume fast SGL by default */
+	io_req->sge_type = QEDF_IOREQ_FAST_SGE; /* Assume fast SGL by default */
 
 	/* Record which cpu this request is associated with */
 	io_req->cpu = smp_processor_id();
@@ -942,7 +882,17 @@  int qedf_post_io_req(struct qedf_rport *fcport, struct qedf_ioreq *io_req)
 	int rc = 0;
 	int rval;
 	unsigned long flags = 0;
-
+	int num_sgs = 0;
+
+	num_sgs = scsi_sg_count(sc_cmd);
+	if (scsi_sg_count(sc_cmd) > QEDF_MAX_BDS_PER_CMD) {
+		QEDF_ERR(&qedf->dbg_ctx,
+			 "Number of SG elements %d exceeds what hardware limitation of %d.\n",
+			 num_sgs, QEDF_MAX_BDS_PER_CMD);
+		sc_cmd->result = DID_ERROR;
+		sc_cmd->scsi_done(sc_cmd);
+		return 0;
+	}
 
 	if (test_bit(QEDF_UNLOADING, &qedf->flags) ||
 	    test_bit(QEDF_DBG_STOP_IO, &qedf->flags)) {