@@ -2738,6 +2738,90 @@ static inline void irdma_get_cqp_reg_info(struct irdma_sc_cqp *cqp, u32 *val,
}
/**
+ * irdma_sc_cqp_def_cmpl_ae_handler - remove completed requests from pending list
+ * @dev: sc device struct
+ * @info: AE entry info
+ * @first: true if this is the first call to this handler for given AEQE
+ * @scratch: (out) scratch entry pointer
+ * @sw_def_info: (in/out) SW ticket value for this AE
+ *
+ * In case of AE_DEF_CMPL event, this function should be called in a loop
+ * until it returns NULL-ptr via scratch.
+ * For each call, it looks for a matching CQP request on pending list,
+ * removes it from the list and returns the pointer to the associated scratch
+ * entry.
+ * If this is the first call to this function for given AEQE, sw_def_info
+ * value is not used to find matching requests. Instead, it is populated
+ * with the value from the first matching cqp_request on the list.
+ * For subsequent calls, ooo_op->sw_def_info need to match the value passed
+ * by a caller.
+ *
+ * Return: scratch entry pointer for cqp_request to be released or NULL
+ * if no matching request is found.
+ */
+void irdma_sc_cqp_def_cmpl_ae_handler(struct irdma_sc_dev *dev,
+ struct irdma_aeqe_info *info,
+ bool first, u64 *scratch,
+ u32 *sw_def_info)
+{
+ struct irdma_ooo_cqp_op *ooo_op;
+ unsigned long flags;
+
+ *scratch = 0;
+
+ spin_lock_irqsave(&dev->cqp->ooo_list_lock, flags);
+ list_for_each_entry(ooo_op, &dev->cqp->ooo_pnd, list_entry) {
+ if (ooo_op->deferred &&
+ ((first && ooo_op->def_info == info->def_info) ||
+ (!first && ooo_op->sw_def_info == *sw_def_info))) {
+ *sw_def_info = ooo_op->sw_def_info;
+ *scratch = ooo_op->scratch;
+
+ list_del(&ooo_op->list_entry);
+ list_add(&ooo_op->list_entry, &dev->cqp->ooo_avail);
+ atomic64_inc(&dev->cqp->completed_ops);
+
+ break;
+ }
+ }
+ spin_unlock_irqrestore(&dev->cqp->ooo_list_lock, flags);
+
+ if (first && !*scratch)
+ ibdev_dbg(to_ibdev(dev),
+ "AEQ: deferred completion with unknown ticket: def_info 0x%x\n",
+ info->def_info);
+}
+
+/**
+ * irdma_sc_cqp_cleanup_handler - remove requests from pending list
+ * @dev: sc device struct
+ *
+ * This function should be called in a loop from irdma_cleanup_pending_cqp_op.
+ * For each call, it returns first CQP request on pending list, removes it
+ * from the list and returns the pointer to the associated scratch entry.
+ *
+ * Return: scratch entry pointer for cqp_request to be released or NULL
+ * if pending list is empty.
+ */
+u64 irdma_sc_cqp_cleanup_handler(struct irdma_sc_dev *dev)
+{
+ struct irdma_ooo_cqp_op *ooo_op;
+ u64 scratch = 0;
+
+ list_for_each_entry(ooo_op, &dev->cqp->ooo_pnd, list_entry) {
+ scratch = ooo_op->scratch;
+
+ list_del(&ooo_op->list_entry);
+ list_add(&ooo_op->list_entry, &dev->cqp->ooo_avail);
+ atomic64_inc(&dev->cqp->completed_ops);
+
+ break;
+ }
+
+ return scratch;
+}
+
+/**
* irdma_cqp_poll_registers - poll cqp registers
* @cqp: struct for cqp hw
* @tail: wqtail register value
@@ -3121,6 +3205,8 @@ void irdma_sc_remove_cq_ctx(struct irdma_sc_ceq *ceq, struct irdma_sc_cq *cq)
int irdma_sc_cqp_init(struct irdma_sc_cqp *cqp,
struct irdma_cqp_init_info *info)
{
+ struct irdma_ooo_cqp_op *ooo_op;
+ u32 num_ooo_ops;
u8 hw_sq_size;
if (info->sq_size > IRDMA_CQP_SW_SQSIZE_2048 ||
@@ -3151,17 +3237,43 @@ int irdma_sc_cqp_init(struct irdma_sc_cqp *cqp,
cqp->rocev2_rto_policy = info->rocev2_rto_policy;
cqp->protocol_used = info->protocol_used;
memcpy(&cqp->dcqcn_params, &info->dcqcn_params, sizeof(cqp->dcqcn_params));
+ if (cqp->dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_3) {
+ cqp->ooisc_blksize = info->ooisc_blksize;
+ cqp->rrsp_blksize = info->rrsp_blksize;
+ cqp->q1_blksize = info->q1_blksize;
+ cqp->xmit_blksize = info->xmit_blksize;
+ cqp->blksizes_valid = info->blksizes_valid;
+ cqp->ts_shift = info->ts_shift;
+ cqp->ts_override = info->ts_override;
+ cqp->en_fine_grained_timers = info->en_fine_grained_timers;
+ cqp->pe_en_vf_cnt = info->pe_en_vf_cnt;
+ cqp->ooo_op_array = info->ooo_op_array;
+ /* initialize the OOO lists */
+ INIT_LIST_HEAD(&cqp->ooo_avail);
+ INIT_LIST_HEAD(&cqp->ooo_pnd);
+ if (cqp->ooo_op_array) {
+ /* Populate avail list entries */
+ for (num_ooo_ops = 0, ooo_op = info->ooo_op_array;
+ num_ooo_ops < cqp->sq_size;
+ num_ooo_ops++, ooo_op++)
+ list_add(&ooo_op->list_entry, &cqp->ooo_avail);
+ }
+ }
info->dev->cqp = cqp;
IRDMA_RING_INIT(cqp->sq_ring, cqp->sq_size);
+ cqp->last_def_cmpl_ticket = 0;
+ cqp->sw_def_cmpl_ticket = 0;
cqp->requested_ops = 0;
atomic64_set(&cqp->completed_ops, 0);
/* for the cqp commands backlog. */
INIT_LIST_HEAD(&cqp->dev->cqp_cmd_head);
writel(0, cqp->dev->hw_regs[IRDMA_CQPTAIL]);
- writel(0, cqp->dev->hw_regs[IRDMA_CQPDB]);
- writel(0, cqp->dev->hw_regs[IRDMA_CCQPSTATUS]);
+ if (cqp->dev->hw_attrs.uk_attrs.hw_rev <= IRDMA_GEN_2) {
+ writel(0, cqp->dev->hw_regs[IRDMA_CQPDB]);
+ writel(0, cqp->dev->hw_regs[IRDMA_CCQPSTATUS]);
+ }
ibdev_dbg(to_ibdev(cqp->dev),
"WQE: sq_size[%04d] hw_sq_size[%04d] sq_base[%p] sq_pa[%pK] cqp[%p] polarity[x%04x]\n",
@@ -3193,6 +3305,7 @@ int irdma_sc_cqp_create(struct irdma_sc_cqp *cqp, u16 *maj_err, u16 *min_err)
return -ENOMEM;
spin_lock_init(&cqp->dev->cqp_lock);
+ spin_lock_init(&cqp->ooo_list_lock);
temp = FIELD_PREP(IRDMA_CQPHC_SQSIZE, cqp->hw_sq_size) |
FIELD_PREP(IRDMA_CQPHC_SVER, cqp->struct_ver) |
@@ -3204,12 +3317,29 @@ int irdma_sc_cqp_create(struct irdma_sc_cqp *cqp, u16 *maj_err, u16 *min_err)
FIELD_PREP(IRDMA_CQPHC_PROTOCOL_USED,
cqp->protocol_used);
}
+ if (hw_rev >= IRDMA_GEN_3)
+ temp |= FIELD_PREP(IRDMA_CQPHC_EN_FINE_GRAINED_TIMERS,
+ cqp->en_fine_grained_timers);
set_64bit_val(cqp->host_ctx, 0, temp);
set_64bit_val(cqp->host_ctx, 8, cqp->sq_pa);
temp = FIELD_PREP(IRDMA_CQPHC_ENABLED_VFS, cqp->ena_vf_count) |
FIELD_PREP(IRDMA_CQPHC_HMC_PROFILE, cqp->hmc_profile);
+
+ if (hw_rev >= IRDMA_GEN_3)
+ temp |= FIELD_PREP(IRDMA_CQPHC_OOISC_BLKSIZE,
+ cqp->ooisc_blksize) |
+ FIELD_PREP(IRDMA_CQPHC_RRSP_BLKSIZE,
+ cqp->rrsp_blksize) |
+ FIELD_PREP(IRDMA_CQPHC_Q1_BLKSIZE, cqp->q1_blksize) |
+ FIELD_PREP(IRDMA_CQPHC_XMIT_BLKSIZE,
+ cqp->xmit_blksize) |
+ FIELD_PREP(IRDMA_CQPHC_BLKSIZES_VALID,
+ cqp->blksizes_valid) |
+ FIELD_PREP(IRDMA_CQPHC_TIMESTAMP_OVERRIDE,
+ cqp->ts_override) |
+ FIELD_PREP(IRDMA_CQPHC_TS_SHIFT, cqp->ts_shift);
set_64bit_val(cqp->host_ctx, 16, temp);
set_64bit_val(cqp->host_ctx, 24, (uintptr_t)cqp);
temp = FIELD_PREP(IRDMA_CQPHC_HW_MAJVER, cqp->hw_maj_ver) |
@@ -3371,6 +3501,87 @@ void irdma_sc_ccq_arm(struct irdma_sc_cq *ccq)
}
/**
+ * irdma_sc_process_def_cmpl - process deferred or pending completion
+ * @cqp: CQP sc struct
+ * @info: CQP CQE info
+ * @wqe_idx: CQP WQE descriptor index
+ * @def_info: deferred op ticket value or out-of-order completion id
+ * @def_cmpl: true for deferred completion, false for pending (RCA)
+ */
+static void irdma_sc_process_def_cmpl(struct irdma_sc_cqp *cqp,
+ struct irdma_ccq_cqe_info *info,
+ u32 wqe_idx, u32 def_info, bool def_cmpl)
+{
+ struct irdma_ooo_cqp_op *ooo_op;
+ unsigned long flags;
+
+ /* Deferred and out-of-order completions share the same list of pending
+ * completions. Since the list can be also accessed from AE handler,
+ * it must be protected by a lock.
+ */
+ spin_lock_irqsave(&cqp->ooo_list_lock, flags);
+
+ /* For deferred completions bump up SW completion ticket value. */
+ if (def_cmpl) {
+ cqp->last_def_cmpl_ticket = def_info;
+ cqp->sw_def_cmpl_ticket++;
+ }
+ if (!list_empty(&cqp->ooo_avail)) {
+ ooo_op = (struct irdma_ooo_cqp_op *)
+ list_entry(cqp->ooo_avail.next,
+ struct irdma_ooo_cqp_op, list_entry);
+
+ list_del(&ooo_op->list_entry);
+ ooo_op->scratch = info->scratch;
+ ooo_op->def_info = def_info;
+ ooo_op->sw_def_info = cqp->sw_def_cmpl_ticket;
+ ooo_op->deferred = def_cmpl;
+ ooo_op->wqe_idx = wqe_idx;
+ /* Pending completions must be chronologically ordered,
+ * so adding at the end of list.
+ */
+ list_add_tail(&ooo_op->list_entry, &cqp->ooo_pnd);
+ }
+ spin_unlock_irqrestore(&cqp->ooo_list_lock, flags);
+
+ info->pending = true;
+}
+
+/**
+ * irdma_sc_process_ooo_cmpl - process out-of-order (final) completion
+ * @cqp: CQP sc struct
+ * @info: CQP CQE info
+ * @def_info: out-of-order completion id
+ */
+static void irdma_sc_process_ooo_cmpl(struct irdma_sc_cqp *cqp,
+ struct irdma_ccq_cqe_info *info,
+ u32 def_info)
+{
+ struct irdma_ooo_cqp_op *ooo_op_tmp;
+ struct irdma_ooo_cqp_op *ooo_op;
+ unsigned long flags;
+
+ info->scratch = 0;
+
+ spin_lock_irqsave(&cqp->ooo_list_lock, flags);
+ list_for_each_entry_safe(ooo_op, ooo_op_tmp, &cqp->ooo_pnd,
+ list_entry) {
+ if (!ooo_op->deferred && ooo_op->def_info == def_info) {
+ list_del(&ooo_op->list_entry);
+ info->scratch = ooo_op->scratch;
+ list_add(&ooo_op->list_entry, &cqp->ooo_avail);
+ break;
+ }
+ }
+ spin_unlock_irqrestore(&cqp->ooo_list_lock, flags);
+
+ if (!info->scratch)
+ ibdev_dbg(to_ibdev(cqp->dev),
+ "CQP: DEBUG_FW_OOO out-of-order completion with unknown def_info = 0x%x\n",
+ def_info);
+}
+
+/**
* irdma_sc_ccq_get_cqe_info - get ccq's cq entry
* @ccq: ccq sc struct
* @info: completion q entry to return
@@ -3378,6 +3589,10 @@ void irdma_sc_ccq_arm(struct irdma_sc_cq *ccq)
int irdma_sc_ccq_get_cqe_info(struct irdma_sc_cq *ccq,
struct irdma_ccq_cqe_info *info)
{
+ u32 def_info;
+ bool def_cmpl = false;
+ bool pend_cmpl = false;
+ bool ooo_final_cmpl = false;
u64 qp_ctx, temp, temp1;
__le64 *cqe;
struct irdma_sc_cqp *cqp;
@@ -3385,6 +3600,7 @@ int irdma_sc_ccq_get_cqe_info(struct irdma_sc_cq *ccq,
u32 error;
u8 polarity;
int ret_code = 0;
+ unsigned long flags;
if (ccq->cq_uk.avoid_mem_cflct)
cqe = IRDMA_GET_CURRENT_EXTENDED_CQ_ELEM(&ccq->cq_uk);
@@ -3416,6 +3632,25 @@ int irdma_sc_ccq_get_cqe_info(struct irdma_sc_cq *ccq,
get_64bit_val(cqe, 16, &temp1);
info->op_ret_val = (u32)FIELD_GET(IRDMA_CCQ_OPRETVAL, temp1);
+ if (cqp->dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_3) {
+ def_cmpl = info->maj_err_code == IRDMA_CQPSQ_MAJ_NO_ERROR &&
+ info->min_err_code == IRDMA_CQPSQ_MIN_DEF_CMPL;
+ def_info = (u32)FIELD_GET(IRDMA_CCQ_DEFINFO, temp1);
+
+ pend_cmpl = info->maj_err_code == IRDMA_CQPSQ_MAJ_NO_ERROR &&
+ info->min_err_code == IRDMA_CQPSQ_MIN_OOO_CMPL;
+
+ ooo_final_cmpl = (bool)FIELD_GET(IRDMA_OOO_CMPL, temp);
+
+ if (def_cmpl || pend_cmpl || ooo_final_cmpl) {
+ if (ooo_final_cmpl)
+ irdma_sc_process_ooo_cmpl(cqp, info, def_info);
+ else
+ irdma_sc_process_def_cmpl(cqp, info, wqe_idx,
+ def_info, def_cmpl);
+ }
+ }
+
get_64bit_val(cqp->sq_base[wqe_idx].elem, 24, &temp1);
info->op_code = (u8)FIELD_GET(IRDMA_CQPSQ_OPCODE, temp1);
info->cqp = cqp;
@@ -3432,7 +3667,16 @@ int irdma_sc_ccq_get_cqe_info(struct irdma_sc_cq *ccq,
dma_wmb(); /* make sure shadow area is updated before moving tail */
- IRDMA_RING_MOVE_TAIL(cqp->sq_ring);
+ spin_lock_irqsave(&cqp->dev->cqp_lock, flags);
+ if (!ooo_final_cmpl)
+ IRDMA_RING_MOVE_TAIL(cqp->sq_ring);
+ spin_unlock_irqrestore(&cqp->dev->cqp_lock, flags);
+
+ /* Do not increment completed_ops counter on pending or deferred
+ * completions.
+ */
+ if (pend_cmpl || def_cmpl)
+ return ret_code;
atomic64_inc(&cqp->completed_ops);
return ret_code;
@@ -4118,6 +4362,10 @@ int irdma_sc_get_next_aeqe(struct irdma_sc_aeq *aeq,
info->compl_ctx = compl_ctx << 1;
ae_src = IRDMA_AE_SOURCE_RSVD;
break;
+ case IRDMA_AE_CQP_DEFERRED_COMPLETE:
+ info->def_info = info->wqe_idx;
+ ae_src = IRDMA_AE_SOURCE_RSVD;
+ break;
case IRDMA_AE_ROCE_EMPTY_MCG:
case IRDMA_AE_ROCE_BAD_MC_IP_ADDR:
case IRDMA_AE_ROCE_BAD_MC_QPID:
@@ -367,6 +367,7 @@ enum irdma_cqp_op_type {
#define IRDMA_AE_LCE_FUNCTION_CATASTROPHIC 0x0701
#define IRDMA_AE_LCE_CQ_CATASTROPHIC 0x0702
#define IRDMA_AE_QP_SUSPEND_COMPLETE 0x0900
+#define IRDMA_AE_CQP_DEFERRED_COMPLETE 0x0901
#define FLD_LS_64(dev, val, field) \
(((u64)(val) << (dev)->hw_shifts[field ## _S]) & (dev)->hw_masks[field ## _M])
@@ -465,6 +466,16 @@ enum irdma_cqp_op_type {
#define IRDMA_CQPHC_SVER GENMASK_ULL(31, 24)
#define IRDMA_CQPHC_SQBASE GENMASK_ULL(63, 9)
+#define IRDMA_CQPHC_TIMESTAMP_OVERRIDE BIT_ULL(5)
+#define IRDMA_CQPHC_TS_SHIFT GENMASK_ULL(12, 8)
+#define IRDMA_CQPHC_EN_FINE_GRAINED_TIMERS BIT_ULL(0)
+
+#define IRDMA_CQPHC_OOISC_BLKSIZE GENMASK_ULL(63, 60)
+#define IRDMA_CQPHC_RRSP_BLKSIZE GENMASK_ULL(59, 56)
+#define IRDMA_CQPHC_Q1_BLKSIZE GENMASK_ULL(55, 52)
+#define IRDMA_CQPHC_XMIT_BLKSIZE GENMASK_ULL(51, 48)
+#define IRDMA_CQPHC_BLKSIZES_VALID BIT_ULL(4)
+
#define IRDMA_CQPHC_QPCTX GENMASK_ULL(63, 0)
#define IRDMA_QP_DBSA_HW_SQ_TAIL GENMASK_ULL(14, 0)
#define IRDMA_CQ_DBSA_CQEIDX GENMASK_ULL(19, 0)
@@ -478,6 +489,8 @@ enum irdma_cqp_op_type {
#define IRDMA_CCQ_OPRETVAL GENMASK_ULL(31, 0)
+#define IRDMA_CCQ_DEFINFO GENMASK_ULL(63, 32)
+
#define IRDMA_CQ_MINERR GENMASK_ULL(15, 0)
#define IRDMA_CQ_MAJERR GENMASK_ULL(31, 16)
#define IRDMA_CQ_WQEIDX GENMASK_ULL(46, 32)
@@ -715,6 +728,8 @@ enum irdma_cqp_op_type {
#define IRDMA_CQPSQ_MIN_STAG_INVALID 0x0001
#define IRDMA_CQPSQ_MIN_SUSPEND_PND 0x0005
+#define IRDMA_CQPSQ_MIN_DEF_CMPL 0x0006
+#define IRDMA_CQPSQ_MIN_OOO_CMPL 0x0007
#define IRDMA_CQPSQ_MAJ_NO_ERROR 0x0000
#define IRDMA_CQPSQ_MAJ_OBJCACHE_ERROR 0xF000
@@ -208,6 +208,51 @@ static void irdma_set_flush_fields(struct irdma_sc_qp *qp,
}
/**
+ * irdma_complete_cqp_request - perform post-completion cleanup
+ * @cqp: device CQP
+ * @cqp_request: CQP request
+ *
+ * Mark CQP request as done, wake up waiting thread or invoke
+ * callback function and release/free CQP request.
+ */
+static void irdma_complete_cqp_request(struct irdma_cqp *cqp,
+ struct irdma_cqp_request *cqp_request)
+{
+ if (cqp_request->waiting) {
+ WRITE_ONCE(cqp_request->request_done, true);
+ wake_up(&cqp_request->waitq);
+ } else if (cqp_request->callback_fcn) {
+ cqp_request->callback_fcn(cqp_request);
+ }
+ irdma_put_cqp_request(cqp, cqp_request);
+}
+
+/**
+ * irdma_process_ae_def_cmpl - handle IRDMA_AE_CQP_DEFERRED_COMPLETE event
+ * @rf: RDMA PCI function
+ * @info: AEQ entry info
+ */
+static void irdma_process_ae_def_cmpl(struct irdma_pci_f *rf,
+ struct irdma_aeqe_info *info)
+{
+ u32 sw_def_info;
+ u64 scratch;
+
+ irdma_cqp_ce_handler(rf, &rf->ccq.sc_cq);
+
+ irdma_sc_cqp_def_cmpl_ae_handler(&rf->sc_dev, info, true,
+ &scratch, &sw_def_info);
+ while (scratch) {
+ struct irdma_cqp_request *cqp_request =
+ (struct irdma_cqp_request *)(uintptr_t)scratch;
+
+ irdma_complete_cqp_request(&rf->cqp, cqp_request);
+ irdma_sc_cqp_def_cmpl_ae_handler(&rf->sc_dev, info, false,
+ &scratch, &sw_def_info);
+ }
+}
+
+/**
* irdma_process_aeq - handle aeq events
* @rf: RDMA PCI function
*/
@@ -269,7 +314,8 @@ static void irdma_process_aeq(struct irdma_pci_f *rf)
spin_unlock_irqrestore(&iwqp->lock, flags);
ctx_info = &iwqp->ctx_info;
} else {
- if (info->ae_id != IRDMA_AE_CQ_OPERATION_ERROR)
+ if (info->ae_id != IRDMA_AE_CQ_OPERATION_ERROR &&
+ info->ae_id != IRDMA_AE_CQP_DEFERRED_COMPLETE)
continue;
}
@@ -364,6 +410,12 @@ static void irdma_process_aeq(struct irdma_pci_f *rf)
}
irdma_cq_rem_ref(&iwcq->ibcq);
break;
+ case IRDMA_AE_CQP_DEFERRED_COMPLETE:
+ /* Remove completed CQP requests from pending list
+ * and notify about those CQP ops completion.
+ */
+ irdma_process_ae_def_cmpl(rf, info);
+ break;
case IRDMA_AE_RESET_NOT_SENT:
case IRDMA_AE_LLP_DOUBT_REACHABILITY:
case IRDMA_AE_RESOURCE_EXHAUSTION:
@@ -602,6 +654,8 @@ static void irdma_destroy_cqp(struct irdma_pci_f *rf)
dma_free_coherent(dev->hw->device, cqp->sq.size, cqp->sq.va,
cqp->sq.pa);
cqp->sq.va = NULL;
+ kfree(cqp->oop_op_array);
+ cqp->oop_op_array = NULL;
kfree(cqp->scratch_array);
cqp->scratch_array = NULL;
kfree(cqp->cqp_requests);
@@ -945,6 +999,13 @@ static int irdma_create_cqp(struct irdma_pci_f *rf)
goto err_scratch;
}
+ cqp->oop_op_array = kcalloc(sqsize, sizeof(*cqp->oop_op_array),
+ GFP_KERNEL);
+ if (!cqp->oop_op_array) {
+ status = -ENOMEM;
+ goto err_oop;
+ }
+ cqp_init_info.ooo_op_array = cqp->oop_op_array;
dev->cqp = &cqp->sc_cqp;
dev->cqp->dev = dev;
cqp->sq.size = ALIGN(sizeof(struct irdma_cqp_sq_wqe) * sqsize,
@@ -981,6 +1042,10 @@ static int irdma_create_cqp(struct irdma_pci_f *rf)
case IRDMA_GEN_2:
cqp_init_info.hw_maj_ver = IRDMA_CQPHC_HW_MAJVER_GEN_2;
break;
+ case IRDMA_GEN_3:
+ cqp_init_info.hw_maj_ver = IRDMA_CQPHC_HW_MAJVER_GEN_3;
+ cqp_init_info.ts_override = 1;
+ break;
}
status = irdma_sc_cqp_init(dev->cqp, &cqp_init_info);
if (status) {
@@ -1015,6 +1080,9 @@ static int irdma_create_cqp(struct irdma_pci_f *rf)
cqp->sq.va, cqp->sq.pa);
cqp->sq.va = NULL;
err_sq:
+ kfree(cqp->oop_op_array);
+ cqp->oop_op_array = NULL;
+err_oop:
kfree(cqp->scratch_array);
cqp->scratch_array = NULL;
err_scratch:
@@ -2106,15 +2174,16 @@ void irdma_cqp_ce_handler(struct irdma_pci_f *rf, struct irdma_sc_cq *cq)
cqp_request->compl_info.op_ret_val = info.op_ret_val;
cqp_request->compl_info.error = info.error;
- if (cqp_request->waiting) {
- WRITE_ONCE(cqp_request->request_done, true);
- wake_up(&cqp_request->waitq);
- irdma_put_cqp_request(&rf->cqp, cqp_request);
- } else {
- if (cqp_request->callback_fcn)
- cqp_request->callback_fcn(cqp_request);
- irdma_put_cqp_request(&rf->cqp, cqp_request);
- }
+ /*
+ * If this is deferred or pending completion, then mark
+ * CQP request as pending to not block the CQ, but don't
+ * release CQP request, as it is still on the OOO list.
+ */
+ if (info.pending)
+ cqp_request->pending = true;
+ else
+ irdma_complete_cqp_request(&rf->cqp,
+ cqp_request);
}
cqe_count++;
@@ -167,6 +167,7 @@ struct irdma_cqp_request {
bool request_done; /* READ/WRITE_ONCE macros operate on it */
bool waiting:1;
bool dynamic:1;
+ bool pending:1;
};
struct irdma_cqp {
@@ -179,6 +180,7 @@ struct irdma_cqp {
struct irdma_dma_mem host_ctx;
u64 *scratch_array;
struct irdma_cqp_request *cqp_requests;
+ struct irdma_ooo_cqp_op *oop_op_array;
struct list_head cqp_avail_reqs;
struct list_head cqp_pending_reqs;
};
@@ -10,6 +10,7 @@
#define ALL_TC2PFC 0xff
#define CQP_COMPL_WAIT_TIME_MS 10
#define CQP_TIMEOUT_THRESHOLD 500
+#define CQP_DEF_CMPL_TIMEOUT_THRESHOLD 2500
/* init operations */
int irdma_sc_dev_init(enum irdma_vers ver, struct irdma_sc_dev *dev,
@@ -262,12 +262,22 @@ struct irdma_cqp_init_info {
__le64 *host_ctx;
u64 *scratch_array;
u32 sq_size;
+ struct irdma_ooo_cqp_op *ooo_op_array;
+ u32 pe_en_vf_cnt;
u16 hw_maj_ver;
u16 hw_min_ver;
u8 struct_ver;
u8 hmc_profile;
u8 ena_vf_count;
u8 ceqs_per_vf;
+ u8 ooisc_blksize;
+ u8 rrsp_blksize;
+ u8 q1_blksize;
+ u8 xmit_blksize;
+ u8 ts_override;
+ u8 ts_shift;
+ u8 en_fine_grained_timers;
+ u8 blksizes_valid;
bool en_datacenter_tcp:1;
bool disable_packed:1;
bool rocev2_rto_policy:1;
@@ -392,7 +402,21 @@ struct irdma_cqp_quanta {
__le64 elem[IRDMA_CQP_WQE_SIZE];
};
+struct irdma_ooo_cqp_op {
+ struct list_head list_entry;
+ u64 scratch;
+ u32 def_info;
+ u32 sw_def_info;
+ u32 wqe_idx;
+ bool deferred:1;
+};
+
struct irdma_sc_cqp {
+ spinlock_t ooo_list_lock; /* protects list of pending completions */
+ struct list_head ooo_avail;
+ struct list_head ooo_pnd;
+ u32 last_def_cmpl_ticket;
+ u32 sw_def_cmpl_ticket;
u32 size;
u64 sq_pa;
u64 host_ctx_pa;
@@ -408,8 +432,10 @@ struct irdma_sc_cqp {
u64 *scratch_array;
u64 requested_ops;
atomic64_t completed_ops;
+ struct irdma_ooo_cqp_op *ooo_op_array;
u32 cqp_id;
u32 sq_size;
+ u32 pe_en_vf_cnt;
u32 hw_sq_size;
u16 hw_maj_ver;
u16 hw_min_ver;
@@ -419,6 +445,14 @@ struct irdma_sc_cqp {
u8 ena_vf_count;
u8 timeout_count;
u8 ceqs_per_vf;
+ u8 ooisc_blksize;
+ u8 rrsp_blksize;
+ u8 q1_blksize;
+ u8 xmit_blksize;
+ u8 ts_override;
+ u8 ts_shift;
+ u8 en_fine_grained_timers;
+ u8 blksizes_valid;
bool en_datacenter_tcp:1;
bool disable_packed:1;
bool rocev2_rto_policy:1;
@@ -723,7 +757,8 @@ struct irdma_ccq_cqe_info {
u16 maj_err_code;
u16 min_err_code;
u8 op_code;
- bool error;
+ bool error:1;
+ bool pending:1;
};
struct irdma_dcb_app_info {
@@ -998,6 +1033,7 @@ struct irdma_qp_host_ctx_info {
struct irdma_aeqe_info {
u64 compl_ctx;
u32 qp_cq_id;
+ u32 def_info; /* only valid for DEF_CMPL */
u16 ae_id;
u16 wqe_idx;
u8 tcp_state;
@@ -1242,6 +1278,11 @@ void irdma_sc_pd_init(struct irdma_sc_dev *dev, struct irdma_sc_pd *pd, u32 pd_i
void irdma_cfg_aeq(struct irdma_sc_dev *dev, u32 idx, bool enable);
void irdma_check_cqp_progress(struct irdma_cqp_timeout *cqp_timeout,
struct irdma_sc_dev *dev);
+void irdma_sc_cqp_def_cmpl_ae_handler(struct irdma_sc_dev *dev,
+ struct irdma_aeqe_info *info,
+ bool first, u64 *scratch,
+ u32 *sw_def_info);
+u64 irdma_sc_cqp_cleanup_handler(struct irdma_sc_dev *dev);
int irdma_sc_cqp_create(struct irdma_sc_cqp *cqp, u16 *maj_err, u16 *min_err);
int irdma_sc_cqp_destroy(struct irdma_sc_cqp *cqp);
int irdma_sc_cqp_init(struct irdma_sc_cqp *cqp,
@@ -484,6 +484,7 @@ void irdma_free_cqp_request(struct irdma_cqp *cqp,
WRITE_ONCE(cqp_request->request_done, false);
cqp_request->callback_fcn = NULL;
cqp_request->waiting = false;
+ cqp_request->pending = false;
spin_lock_irqsave(&cqp->req_lock, flags);
list_add_tail(&cqp_request->list, &cqp->cqp_avail_reqs);
@@ -524,6 +525,22 @@ void irdma_put_cqp_request(struct irdma_cqp *cqp,
}
/**
+ * irdma_cleanup_deferred_cqp_ops - clean-up cqp with no completions
+ * @dev: sc_dev
+ * @cqp: cqp
+ */
+static void irdma_cleanup_deferred_cqp_ops(struct irdma_sc_dev *dev,
+ struct irdma_cqp *cqp)
+{
+ u64 scratch;
+
+ /* process all CQP requests with deferred/pending completions */
+ while ((scratch = irdma_sc_cqp_cleanup_handler(dev)))
+ irdma_free_pending_cqp_request(cqp, (struct irdma_cqp_request *)
+ (uintptr_t)scratch);
+}
+
+/**
* irdma_cleanup_pending_cqp_op - clean-up cqp with no
* completions
* @rf: RDMA PCI function
@@ -536,6 +553,8 @@ void irdma_cleanup_pending_cqp_op(struct irdma_pci_f *rf)
struct cqp_cmds_info *pcmdinfo = NULL;
u32 i, pending_work, wqe_idx;
+ if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_3)
+ irdma_cleanup_deferred_cqp_ops(dev, cqp);
pending_work = IRDMA_RING_USED_QUANTA(cqp->sc_cqp.sq_ring);
wqe_idx = IRDMA_RING_CURRENT_TAIL(cqp->sc_cqp.sq_ring);
for (i = 0; i < pending_work; i++) {
@@ -555,6 +574,26 @@ void irdma_cleanup_pending_cqp_op(struct irdma_pci_f *rf)
}
}
+static int irdma_get_timeout_threshold(struct irdma_sc_dev *dev)
+{
+ u16 time_s = dev->vc_caps.cqp_timeout_s;
+
+ if (!time_s)
+ return CQP_TIMEOUT_THRESHOLD;
+
+ return time_s * 1000 / dev->hw_attrs.max_cqp_compl_wait_time_ms;
+}
+
+static int irdma_get_def_timeout_threshold(struct irdma_sc_dev *dev)
+{
+ u16 time_s = dev->vc_caps.cqp_def_timeout_s;
+
+ if (!time_s)
+ return CQP_DEF_CMPL_TIMEOUT_THRESHOLD;
+
+ return time_s * 1000 / dev->hw_attrs.max_cqp_compl_wait_time_ms;
+}
+
/**
* irdma_wait_event - wait for completion
* @rf: RDMA PCI function
@@ -564,6 +603,7 @@ static int irdma_wait_event(struct irdma_pci_f *rf,
struct irdma_cqp_request *cqp_request)
{
struct irdma_cqp_timeout cqp_timeout = {};
+ int timeout_threshold = irdma_get_timeout_threshold(&rf->sc_dev);
bool cqp_error = false;
int err_code = 0;
@@ -575,9 +615,17 @@ static int irdma_wait_event(struct irdma_pci_f *rf,
msecs_to_jiffies(CQP_COMPL_WAIT_TIME_MS)))
break;
+ if (cqp_request->pending)
+ /* There was a deferred or pending completion
+ * received for this CQP request, so we need
+ * to wait longer than usual.
+ */
+ timeout_threshold =
+ irdma_get_def_timeout_threshold(&rf->sc_dev);
+
irdma_check_cqp_progress(&cqp_timeout, &rf->sc_dev);
- if (cqp_timeout.count < CQP_TIMEOUT_THRESHOLD)
+ if (cqp_timeout.count < timeout_threshold)
continue;
if (!rf->reset) {