Commit 82d8dfd2 authored by Quinn Tran's avatar Quinn Tran Committed by Martin K. Petersen

scsi: qla2xxx: edif: Fix performance dip due to lock contention

User experienced performance dip on measuring IOPS while EDIF
enabled. During I/O time, driver uses dma_pool_zalloc() call to allocate a
chunk of memory. This call contains a lock behind the scene which
contribute to lock contention. Save the allocated memory for reuse and
avoid the lock.
Signed-off-by: default avatarQuinn Tran <qutran@marvell.com>
Signed-off-by: default avatarNilesh Javali <njavali@marvell.com>
Reviewed-by: default avatarHimanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: default avatarMartin K. Petersen <martin.petersen@oracle.com>
parent 430eef03
......@@ -384,6 +384,13 @@ struct els_reject {
struct req_que;
struct qla_tgt_sess;
struct qla_buf_dsc {
u16 tag;
#define TAG_FREED 0xffff
void *buf;
dma_addr_t buf_dma;
};
/*
* SCSI Request Block
*/
......@@ -392,14 +399,16 @@ struct srb_cmd {
uint32_t request_sense_length;
uint32_t fw_sense_length;
uint8_t *request_sense_ptr;
struct ct6_dsd *ct6_ctx;
struct crc_context *crc_ctx;
struct ct6_dsd ct6_ctx;
struct qla_buf_dsc buf_dsc;
};
/*
* SRB flag definitions
*/
#define SRB_DMA_VALID BIT_0 /* Command sent to ISP */
#define SRB_GOT_BUF BIT_1
#define SRB_FCP_CMND_DMA_VALID BIT_12 /* DIF: DSD List valid */
#define SRB_CRC_CTX_DMA_VALID BIT_2 /* DIF: context DMA valid */
#define SRB_CRC_PROT_DMA_VALID BIT_4 /* DIF: prot DMA valid */
......@@ -3722,6 +3731,16 @@ struct qla_fw_resources {
#define QLA_IOCB_PCT_LIMIT 95
struct qla_buf_pool {
u16 num_bufs;
u16 num_active;
u16 max_used;
u16 reserved;
unsigned long *buf_map;
void **buf_array;
dma_addr_t *dma_array;
};
/*Queue pair data structure */
struct qla_qpair {
spinlock_t qp_lock;
......@@ -3775,6 +3794,7 @@ struct qla_qpair {
struct qla_tgt_counters tgt_counters;
uint16_t cpuid;
struct qla_fw_resources fwres ____cacheline_aligned;
struct qla_buf_pool buf_pool;
u32 cmd_cnt;
u32 cmd_completion_cnt;
u32 prev_completion_cnt;
......
......@@ -3007,26 +3007,16 @@ qla28xx_start_scsi_edif(srb_t *sp)
goto queuing_error;
}
ctx = sp->u.scmd.ct6_ctx =
mempool_alloc(ha->ctx_mempool, GFP_ATOMIC);
if (!ctx) {
ql_log(ql_log_fatal, vha, 0x3010,
"Failed to allocate ctx for cmd=%p.\n", cmd);
goto queuing_error;
}
memset(ctx, 0, sizeof(struct ct6_dsd));
ctx->fcp_cmnd = dma_pool_zalloc(ha->fcp_cmnd_dma_pool,
GFP_ATOMIC, &ctx->fcp_cmnd_dma);
if (!ctx->fcp_cmnd) {
if (qla_get_buf(vha, sp->qpair, &sp->u.scmd.buf_dsc)) {
ql_log(ql_log_fatal, vha, 0x3011,
"Failed to allocate fcp_cmnd for cmd=%p.\n", cmd);
"Failed to allocate buf for fcp_cmnd for cmd=%p.\n", cmd);
goto queuing_error;
}
/* Initialize the DSD list and dma handle */
INIT_LIST_HEAD(&ctx->dsd_list);
ctx->dsd_use_cnt = 0;
sp->flags |= SRB_GOT_BUF;
ctx = &sp->u.scmd.ct6_ctx;
ctx->fcp_cmnd = sp->u.scmd.buf_dsc.buf;
ctx->fcp_cmnd_dma = sp->u.scmd.buf_dsc.buf_dma;
if (cmd->cmd_len > 16) {
additional_cdb_len = cmd->cmd_len - 16;
......@@ -3145,7 +3135,6 @@ qla28xx_start_scsi_edif(srb_t *sp)
cmd_pkt->fcp_cmnd_dseg_len = cpu_to_le16(ctx->fcp_cmnd_len);
put_unaligned_le64(ctx->fcp_cmnd_dma, &cmd_pkt->fcp_cmnd_dseg_address);
sp->flags |= SRB_FCP_CMND_DMA_VALID;
cmd_pkt->byte_count = cpu_to_le32((uint32_t)scsi_bufflen(cmd));
/* Set total data segment count. */
cmd_pkt->entry_count = (uint8_t)req_cnt;
......@@ -3177,15 +3166,11 @@ qla28xx_start_scsi_edif(srb_t *sp)
return QLA_SUCCESS;
queuing_error_fcp_cmnd:
dma_pool_free(ha->fcp_cmnd_dma_pool, ctx->fcp_cmnd, ctx->fcp_cmnd_dma);
queuing_error:
if (tot_dsds)
scsi_dma_unmap(cmd);
if (sp->u.scmd.ct6_ctx) {
mempool_free(sp->u.scmd.ct6_ctx, ha->ctx_mempool);
sp->u.scmd.ct6_ctx = NULL;
}
qla_put_buf(sp->qpair, &sp->u.scmd.buf_dsc);
qla_put_fw_resources(sp->qpair, &sp->iores);
spin_unlock_irqrestore(lock, flags);
......
......@@ -1015,5 +1015,8 @@ int qla2xxx_enable_port(struct Scsi_Host *shost);
uint64_t qla2x00_get_num_tgts(scsi_qla_host_t *vha);
uint64_t qla2x00_count_set_bits(u32 num);
int qla_create_buf_pool(struct scsi_qla_host *, struct qla_qpair *);
void qla_free_buf_pool(struct qla_qpair *);
int qla_get_buf(struct scsi_qla_host *, struct qla_qpair *, struct qla_buf_dsc *);
void qla_put_buf(struct qla_qpair *, struct qla_buf_dsc *);
#endif /* _QLA_GBL_H */
......@@ -9442,6 +9442,13 @@ struct qla_qpair *qla2xxx_create_qpair(struct scsi_qla_host *vha, int qos,
goto fail_mempool;
}
if (qla_create_buf_pool(vha, qpair)) {
ql_log(ql_log_warn, vha, 0xd036,
"Failed to initialize buf pool for qpair %d\n",
qpair->id);
goto fail_bufpool;
}
/* Mark as online */
qpair->online = 1;
......@@ -9457,7 +9464,10 @@ struct qla_qpair *qla2xxx_create_qpair(struct scsi_qla_host *vha, int qos,
}
return qpair;
fail_bufpool:
mempool_destroy(qpair->srb_mempool);
fail_mempool:
qla25xx_delete_req_que(vha, qpair->req);
fail_req:
qla25xx_delete_rsp_que(vha, qpair->rsp);
fail_rsp:
......@@ -9483,6 +9493,8 @@ int qla2xxx_delete_qpair(struct scsi_qla_host *vha, struct qla_qpair *qpair)
qpair->delete_in_progress = 1;
qla_free_buf_pool(qpair);
ret = qla25xx_delete_req_que(vha, qpair->req);
if (ret != QLA_SUCCESS)
goto fail;
......
......@@ -623,7 +623,7 @@ qla24xx_build_scsi_type_6_iocbs(srb_t *sp, struct cmd_type_6 *cmd_pkt,
}
cur_seg = scsi_sglist(cmd);
ctx = sp->u.scmd.ct6_ctx;
ctx = &sp->u.scmd.ct6_ctx;
while (tot_dsds) {
avail_dsds = (tot_dsds > QLA_DSDS_PER_IOCB) ?
......@@ -3459,13 +3459,7 @@ qla82xx_start_scsi(srb_t *sp)
goto queuing_error;
}
ctx = sp->u.scmd.ct6_ctx =
mempool_alloc(ha->ctx_mempool, GFP_ATOMIC);
if (!ctx) {
ql_log(ql_log_fatal, vha, 0x3010,
"Failed to allocate ctx for cmd=%p.\n", cmd);
goto queuing_error;
}
ctx = &sp->u.scmd.ct6_ctx;
memset(ctx, 0, sizeof(struct ct6_dsd));
ctx->fcp_cmnd = dma_pool_zalloc(ha->fcp_cmnd_dma_pool,
......
......@@ -1080,3 +1080,119 @@ void qla_update_host_map(struct scsi_qla_host *vha, port_id_t id)
qla_update_vp_map(vha, SET_AL_PA);
}
}
int qla_create_buf_pool(struct scsi_qla_host *vha, struct qla_qpair *qp)
{
int sz;
qp->buf_pool.num_bufs = qp->req->length;
sz = BITS_TO_LONGS(qp->req->length);
qp->buf_pool.buf_map = kcalloc(sz, sizeof(long), GFP_KERNEL);
if (!qp->buf_pool.buf_map) {
ql_log(ql_log_warn, vha, 0x0186,
"Failed to allocate buf_map(%ld).\n", sz * sizeof(unsigned long));
return -ENOMEM;
}
sz = qp->req->length * sizeof(void *);
qp->buf_pool.buf_array = kcalloc(qp->req->length, sizeof(void *), GFP_KERNEL);
if (!qp->buf_pool.buf_array) {
ql_log(ql_log_warn, vha, 0x0186,
"Failed to allocate buf_array(%d).\n", sz);
kfree(qp->buf_pool.buf_map);
return -ENOMEM;
}
sz = qp->req->length * sizeof(dma_addr_t);
qp->buf_pool.dma_array = kcalloc(qp->req->length, sizeof(dma_addr_t), GFP_KERNEL);
if (!qp->buf_pool.dma_array) {
ql_log(ql_log_warn, vha, 0x0186,
"Failed to allocate dma_array(%d).\n", sz);
kfree(qp->buf_pool.buf_map);
kfree(qp->buf_pool.buf_array);
return -ENOMEM;
}
set_bit(0, qp->buf_pool.buf_map);
return 0;
}
void qla_free_buf_pool(struct qla_qpair *qp)
{
int i;
struct qla_hw_data *ha = qp->vha->hw;
for (i = 0; i < qp->buf_pool.num_bufs; i++) {
if (qp->buf_pool.buf_array[i] && qp->buf_pool.dma_array[i])
dma_pool_free(ha->fcp_cmnd_dma_pool, qp->buf_pool.buf_array[i],
qp->buf_pool.dma_array[i]);
qp->buf_pool.buf_array[i] = NULL;
qp->buf_pool.dma_array[i] = 0;
}
kfree(qp->buf_pool.dma_array);
kfree(qp->buf_pool.buf_array);
kfree(qp->buf_pool.buf_map);
}
/* it is assume qp->qp_lock is held at this point */
int qla_get_buf(struct scsi_qla_host *vha, struct qla_qpair *qp, struct qla_buf_dsc *dsc)
{
u16 tag, i = 0;
void *buf;
dma_addr_t buf_dma;
struct qla_hw_data *ha = vha->hw;
dsc->tag = TAG_FREED;
again:
tag = find_first_zero_bit(qp->buf_pool.buf_map, qp->buf_pool.num_bufs);
if (tag >= qp->buf_pool.num_bufs) {
ql_dbg(ql_dbg_io, vha, 0x00e2,
"qp(%d) ran out of buf resource.\n", qp->id);
return -EIO;
}
if (tag == 0) {
set_bit(0, qp->buf_pool.buf_map);
i++;
if (i == 5) {
ql_dbg(ql_dbg_io, vha, 0x00e3,
"qp(%d) unable to get tag.\n", qp->id);
return -EIO;
}
goto again;
}
if (!qp->buf_pool.buf_array[tag]) {
buf = dma_pool_zalloc(ha->fcp_cmnd_dma_pool, GFP_ATOMIC, &buf_dma);
if (!buf) {
ql_log(ql_log_fatal, vha, 0x13b1,
"Failed to allocate buf.\n");
return -ENOMEM;
}
dsc->buf = qp->buf_pool.buf_array[tag] = buf;
dsc->buf_dma = qp->buf_pool.dma_array[tag] = buf_dma;
} else {
dsc->buf = qp->buf_pool.buf_array[tag];
dsc->buf_dma = qp->buf_pool.dma_array[tag];
memset(dsc->buf, 0, FCP_CMND_DMA_POOL_SIZE);
}
qp->buf_pool.num_active++;
if (qp->buf_pool.num_active > qp->buf_pool.max_used)
qp->buf_pool.max_used = qp->buf_pool.num_active;
dsc->tag = tag;
set_bit(tag, qp->buf_pool.buf_map);
return 0;
}
/* it is assume qp->qp_lock is held at this point */
void qla_put_buf(struct qla_qpair *qp, struct qla_buf_dsc *dsc)
{
if (dsc->tag == TAG_FREED)
return;
clear_bit(dsc->tag, qp->buf_pool.buf_map);
qp->buf_pool.num_active--;
dsc->tag = TAG_FREED;
}
......@@ -733,15 +733,17 @@ void qla2x00_sp_free_dma(srb_t *sp)
}
if (sp->flags & SRB_FCP_CMND_DMA_VALID) {
struct ct6_dsd *ctx1 = sp->u.scmd.ct6_ctx;
struct ct6_dsd *ctx1 = &sp->u.scmd.ct6_ctx;
dma_pool_free(ha->fcp_cmnd_dma_pool, ctx1->fcp_cmnd,
ctx1->fcp_cmnd_dma);
list_splice(&ctx1->dsd_list, &ha->gbl_dsd_list);
ha->gbl_dsd_inuse -= ctx1->dsd_use_cnt;
ha->gbl_dsd_avail += ctx1->dsd_use_cnt;
mempool_free(ctx1, ha->ctx_mempool);
}
if (sp->flags & SRB_GOT_BUF)
qla_put_buf(sp->qpair, &sp->u.scmd.buf_dsc);
}
void qla2x00_sp_compl(srb_t *sp, int res)
......@@ -817,14 +819,13 @@ void qla2xxx_qpair_sp_free_dma(srb_t *sp)
}
if (sp->flags & SRB_FCP_CMND_DMA_VALID) {
struct ct6_dsd *ctx1 = sp->u.scmd.ct6_ctx;
struct ct6_dsd *ctx1 = &sp->u.scmd.ct6_ctx;
dma_pool_free(ha->fcp_cmnd_dma_pool, ctx1->fcp_cmnd,
ctx1->fcp_cmnd_dma);
list_splice(&ctx1->dsd_list, &ha->gbl_dsd_list);
ha->gbl_dsd_inuse -= ctx1->dsd_use_cnt;
ha->gbl_dsd_avail += ctx1->dsd_use_cnt;
mempool_free(ctx1, ha->ctx_mempool);
sp->flags &= ~SRB_FCP_CMND_DMA_VALID;
}
......@@ -834,6 +835,9 @@ void qla2xxx_qpair_sp_free_dma(srb_t *sp)
dma_pool_free(ha->dl_dma_pool, ctx0, ctx0->crc_ctx_dma);
sp->flags &= ~SRB_CRC_CTX_DMA_VALID;
}
if (sp->flags & SRB_GOT_BUF)
qla_put_buf(sp->qpair, &sp->u.scmd.buf_dsc);
}
void qla2xxx_qpair_sp_compl(srb_t *sp, int res)
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment