scsi: elx: efct: Hardware queues processing

Add driver definitions for:

 - Routines for EQ, CQ, WQ and RQ processing.

 - Routines for I/O object pool allocation and deallocation.

Link: https://lore.kernel.org/r/20210601235512.20104-23-jsmart2021@gmail.com
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Daniel Wagner <dwagner@suse.de>
Co-developed-by: Ram Vegesna <ram.vegesna@broadcom.com>
Signed-off-by: Ram Vegesna <ram.vegesna@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

drivers/scsi/elx/efct/efct_hw.c

@@ -2131,3 +2131,363 @@ efct_hw_reqtag_get_instance(struct efct_hw *hw, u32 instance_index)
 
 	return wqcb;
 }
+
+int
+efct_hw_queue_hash_find(struct efct_queue_hash *hash, u16 id)
+{
+	int index = -1;
+	int i = id & (EFCT_HW_Q_HASH_SIZE - 1);
+
+	/*
+	 * Since the hash is always bigger than the maximum number of Qs, then
+	 * we never have to worry about an infinite loop. We will always find
+	 * an unused entry.
+	 */
+	do {
+		if (hash[i].in_use && hash[i].id == id)
+			index = hash[i].index;
+		else
+			i = (i + 1) & (EFCT_HW_Q_HASH_SIZE - 1);
+	} while (index == -1 && hash[i].in_use);
+
+	return index;
+}
+
+int
+efct_hw_process(struct efct_hw *hw, u32 vector,
+		u32 max_isr_time_msec)
+{
+	struct hw_eq *eq;
+
+	/*
+	 * The caller should disable interrupts if they wish to prevent us
+	 * from processing during a shutdown. The following states are defined:
+	 *   EFCT_HW_STATE_UNINITIALIZED - No queues allocated
+	 *   EFCT_HW_STATE_QUEUES_ALLOCATED - The state after a chip reset,
+	 *                                    queues are cleared.
+	 *   EFCT_HW_STATE_ACTIVE - Chip and queues are operational
+	 *   EFCT_HW_STATE_RESET_IN_PROGRESS - reset, we still want completions
+	 *   EFCT_HW_STATE_TEARDOWN_IN_PROGRESS - We still want mailbox
+	 *                                        completions.
+	 */
+	if (hw->state == EFCT_HW_STATE_UNINITIALIZED)
+		return 0;
+
+	/* Get pointer to struct hw_eq */
+	eq = hw->hw_eq[vector];
+	if (!eq)
+		return 0;
+
+	eq->use_count++;
+
+	return efct_hw_eq_process(hw, eq, max_isr_time_msec);
+}
+
+int
+efct_hw_eq_process(struct efct_hw *hw, struct hw_eq *eq,
+		   u32 max_isr_time_msec)
+{
+	u8 eqe[sizeof(struct sli4_eqe)] = { 0 };
+	u32 tcheck_count;
+	u64 tstart;
+	u64 telapsed;
+	bool done = false;
+
+	tcheck_count = EFCT_HW_TIMECHECK_ITERATIONS;
+	tstart = jiffies_to_msecs(jiffies);
+
+	while (!done && !sli_eq_read(&hw->sli, eq->queue, eqe)) {
+		u16 cq_id = 0;
+		int rc;
+
+		rc = sli_eq_parse(&hw->sli, eqe, &cq_id);
+		if (unlikely(rc)) {
+			if (rc == SLI4_EQE_STATUS_EQ_FULL) {
+				u32 i;
+
+				/*
+				 * Received a sentinel EQE indicating the
+				 * EQ is full. Process all CQs
+				 */
+				for (i = 0; i < hw->cq_count; i++)
+					efct_hw_cq_process(hw, hw->hw_cq[i]);
+				continue;
+			} else {
+				return rc;
+			}
+		} else {
+			int index;
+
+			index  = efct_hw_queue_hash_find(hw->cq_hash, cq_id);
+
+			if (likely(index >= 0))
+				efct_hw_cq_process(hw, hw->hw_cq[index]);
+			else
+				efc_log_err(hw->os, "bad CQ_ID %#06x\n", cq_id);
+		}
+
+		if (eq->queue->n_posted > eq->queue->posted_limit)
+			sli_queue_arm(&hw->sli, eq->queue, false);
+
+		if (tcheck_count && (--tcheck_count == 0)) {
+			tcheck_count = EFCT_HW_TIMECHECK_ITERATIONS;
+			telapsed = jiffies_to_msecs(jiffies) - tstart;
+			if (telapsed >= max_isr_time_msec)
+				done = true;
+		}
+	}
+	sli_queue_eq_arm(&hw->sli, eq->queue, true);
+
+	return 0;
+}
+
+static int
+_efct_hw_wq_write(struct hw_wq *wq, struct efct_hw_wqe *wqe)
+{
+	int queue_rc;
+
+	/* Every so often, set the wqec bit to generate comsummed completions */
+	if (wq->wqec_count)
+		wq->wqec_count--;
+
+	if (wq->wqec_count == 0) {
+		struct sli4_generic_wqe *genwqe = (void *)wqe->wqebuf;
+
+		genwqe->cmdtype_wqec_byte |= SLI4_GEN_WQE_WQEC;
+		wq->wqec_count = wq->wqec_set_count;
+	}
+
+	/* Decrement WQ free count */
+	wq->free_count--;
+
+	queue_rc = sli_wq_write(&wq->hw->sli, wq->queue, wqe->wqebuf);
+
+	return (queue_rc < 0) ? -EIO : 0;
+}
+
+static void
+hw_wq_submit_pending(struct hw_wq *wq, u32 update_free_count)
+{
+	struct efct_hw_wqe *wqe;
+	unsigned long flags = 0;
+
+	spin_lock_irqsave(&wq->queue->lock, flags);
+
+	/* Update free count with value passed in */
+	wq->free_count += update_free_count;
+
+	while ((wq->free_count > 0) && (!list_empty(&wq->pending_list))) {
+		wqe = list_first_entry(&wq->pending_list,
+				       struct efct_hw_wqe, list_entry);
+		list_del_init(&wqe->list_entry);
+		_efct_hw_wq_write(wq, wqe);
+
+		if (wqe->abort_wqe_submit_needed) {
+			wqe->abort_wqe_submit_needed = false;
+			efct_hw_fill_abort_wqe(wq->hw, wqe);
+			INIT_LIST_HEAD(&wqe->list_entry);
+			list_add_tail(&wqe->list_entry, &wq->pending_list);
+			wq->wq_pending_count++;
+		}
+	}
+
+	spin_unlock_irqrestore(&wq->queue->lock, flags);
+}
+
+void
+efct_hw_cq_process(struct efct_hw *hw, struct hw_cq *cq)
+{
+	u8 cqe[sizeof(struct sli4_mcqe)];
+	u16 rid = U16_MAX;
+	/* completion type */
+	enum sli4_qentry ctype;
+	u32 n_processed = 0;
+	u32 tstart, telapsed;
+
+	tstart = jiffies_to_msecs(jiffies);
+
+	while (!sli_cq_read(&hw->sli, cq->queue, cqe)) {
+		int status;
+
+		status = sli_cq_parse(&hw->sli, cq->queue, cqe, &ctype, &rid);
+		/*
+		 * The sign of status is significant. If status is:
+		 * == 0 : call completed correctly and
+		 * the CQE indicated success
+		 * > 0 : call completed correctly and
+		 * the CQE indicated an error
+		 * < 0 : call failed and no information is available about the
+		 * CQE
+		 */
+		if (status < 0) {
+			if (status == SLI4_MCQE_STATUS_NOT_COMPLETED)
+				/*
+				 * Notification that an entry was consumed,
+				 * but not completed
+				 */
+				continue;
+
+			break;
+		}
+
+		switch (ctype) {
+		case SLI4_QENTRY_ASYNC:
+			sli_cqe_async(&hw->sli, cqe);
+			break;
+		case SLI4_QENTRY_MQ:
+			/*
+			 * Process MQ entry. Note there is no way to determine
+			 * the MQ_ID from the completion entry.
+			 */
+			efct_hw_mq_process(hw, status, hw->mq);
+			break;
+		case SLI4_QENTRY_WQ:
+			efct_hw_wq_process(hw, cq, cqe, status, rid);
+			break;
+		case SLI4_QENTRY_WQ_RELEASE: {
+			u32 wq_id = rid;
+			int index;
+			struct hw_wq *wq = NULL;
+
+			index = efct_hw_queue_hash_find(hw->wq_hash, wq_id);
+
+			if (likely(index >= 0)) {
+				wq = hw->hw_wq[index];
+			} else {
+				efc_log_err(hw->os, "bad WQ_ID %#06x\n", wq_id);
+				break;
+			}
+			/* Submit any HW IOs that are on the WQ pending list */
+			hw_wq_submit_pending(wq, wq->wqec_set_count);
+
+			break;
+		}
+
+		case SLI4_QENTRY_RQ:
+			efct_hw_rqpair_process_rq(hw, cq, cqe);
+			break;
+		case SLI4_QENTRY_XABT: {
+			efct_hw_xabt_process(hw, cq, cqe, rid);
+			break;
+		}
+		default:
+			efc_log_debug(hw->os, "unhandled ctype=%#x rid=%#x\n",
+				      ctype, rid);
+			break;
+		}
+
+		n_processed++;
+		if (n_processed == cq->queue->proc_limit)
+			break;
+
+		if (cq->queue->n_posted >= cq->queue->posted_limit)
+			sli_queue_arm(&hw->sli, cq->queue, false);
+	}
+
+	sli_queue_arm(&hw->sli, cq->queue, true);
+
+	if (n_processed > cq->queue->max_num_processed)
+		cq->queue->max_num_processed = n_processed;
+	telapsed = jiffies_to_msecs(jiffies) - tstart;
+	if (telapsed > cq->queue->max_process_time)
+		cq->queue->max_process_time = telapsed;
+}
+
+void
+efct_hw_wq_process(struct efct_hw *hw, struct hw_cq *cq,
+		   u8 *cqe, int status, u16 rid)
+{
+	struct hw_wq_callback *wqcb;
+
+	if (rid == EFCT_HW_REQUE_XRI_REGTAG) {
+		if (status)
+			efc_log_err(hw->os, "reque xri failed, status = %d\n",
+				    status);
+		return;
+	}
+
+	wqcb = efct_hw_reqtag_get_instance(hw, rid);
+	if (!wqcb) {
+		efc_log_err(hw->os, "invalid request tag: x%x\n", rid);
+		return;
+	}
+
+	if (!wqcb->callback) {
+		efc_log_err(hw->os, "wqcb callback is NULL\n");
+		return;
+	}
+
+	(*wqcb->callback)(wqcb->arg, cqe, status);
+}
+
+void
+efct_hw_xabt_process(struct efct_hw *hw, struct hw_cq *cq,
+		     u8 *cqe, u16 rid)
+{
+	/* search IOs wait free list */
+	struct efct_hw_io *io = NULL;
+	unsigned long flags = 0;
+
+	io = efct_hw_io_lookup(hw, rid);
+	if (!io) {
+		/* IO lookup failure should never happen */
+		efc_log_err(hw->os, "xabt io lookup failed rid=%#x\n", rid);
+		return;
+	}
+
+	if (!io->xbusy)
+		efc_log_debug(hw->os, "xabt io not busy rid=%#x\n", rid);
+	else
+		/* mark IO as no longer busy */
+		io->xbusy = false;
+
+	/*
+	 * For IOs that were aborted internally, we need to issue any pending
+	 * callback here.
+	 */
+	if (io->done) {
+		efct_hw_done_t done = io->done;
+		void		*arg = io->arg;
+
+		/*
+		 * Use latched status as this is always saved for an internal
+		 * abort
+		 */
+		int status = io->saved_status;
+		u32 len = io->saved_len;
+		u32 ext = io->saved_ext;
+
+		io->done = NULL;
+		io->status_saved = false;
+
+		done(io, len, status, ext, arg);
+	}
+
+	spin_lock_irqsave(&hw->io_lock, flags);
+	if (io->state == EFCT_HW_IO_STATE_INUSE ||
+	    io->state == EFCT_HW_IO_STATE_WAIT_FREE) {
+		/* if on wait_free list, caller has already freed IO;
+		 * remove from wait_free list and add to free list.
+		 * if on in-use list, already marked as no longer busy;
+		 * just leave there and wait for caller to free.
+		 */
+		if (io->state == EFCT_HW_IO_STATE_WAIT_FREE) {
+			io->state = EFCT_HW_IO_STATE_FREE;
+			list_del_init(&io->list_entry);
+			efct_hw_io_free_move_correct_list(hw, io);
+		}
+	}
+	spin_unlock_irqrestore(&hw->io_lock, flags);
+}
+
+static int
+efct_hw_flush(struct efct_hw *hw)
+{
+	u32 i = 0;
+
+	/* Process any remaining completions */
+	for (i = 0; i < hw->eq_count; i++)
+		efct_hw_process(hw, i, ~0);
+
+	return 0;
+}