X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;ds=sidebyside;f=drivers%2Finfiniband%2Fhw%2Fmthca%2Fmthca_cq.c;h=1159c8a0f2c5211d3b9b6f9f2a2b59ea8de6872b;hb=refs%2Fheads%2Fvserver;hp=96f1a86bf04950f54df57b36522ae008bfff00d5;hpb=76828883507a47dae78837ab5dec5a5b4513c667;p=linux-2.6.git

diff --git a/drivers/infiniband/hw/mthca/mthca_cq.c b/drivers/infiniband/hw/mthca/mthca_cq.c
index 96f1a86bf..1159c8a0f 100644
--- a/drivers/infiniband/hw/mthca/mthca_cq.c
+++ b/drivers/infiniband/hw/mthca/mthca_cq.c
@@ -1,7 +1,7 @@
 /*
  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
- * Copyright (c) 2005 Cisco Systems, Inc. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems, Inc. All rights reserved.
  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
  *
@@ -36,9 +36,10 @@
  * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
  */
 
-#include <linux/init.h>
 #include <linux/hardirq.h>
 
+#include <asm/io.h>
+
 #include <rdma/ib_pack.h>
 
 #include "mthca_dev.h"
@@ -53,6 +54,10 @@ enum {
 	MTHCA_CQ_ENTRY_SIZE = 0x20
 };
 
+enum {
+	MTHCA_ATOMIC_BYTE_LEN = 8
+};
+
 /*
  * Must be packed because start is 64 bits but only aligned to 32 bits.
  */
@@ -150,24 +155,29 @@ struct mthca_err_cqe {
 #define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
 
-static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
+static inline struct mthca_cqe *get_cqe_from_buf(struct mthca_cq_buf *buf,
+						 int entry)
 {
-	if (cq->is_direct)
-		return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
+	if (buf->is_direct)
+		return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
 	else
-		return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
+		return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
 			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
 }
 
-static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
+static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
+{
+	return get_cqe_from_buf(&cq->buf, entry);
+}
+
+static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
 {
-	struct mthca_cqe *cqe = get_cqe(cq, i);
 	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
 }
 
 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
 {
-	return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
+	return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
 }
 
 static inline void set_cqe_hw(struct mthca_cqe *cqe)
@@ -205,6 +215,11 @@ static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
 		mthca_write64(doorbell,
 			      dev->kar + MTHCA_CQ_DOORBELL,
 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+		/*
+		 * Make sure doorbells don't leak out of CQ spinlock
+		 * and reach the HCA out of order:
+		 */
+		mmiowb();
 	}
 }
 
@@ -233,9 +248,9 @@ void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
 	spin_lock(&dev->cq_table.lock);
 
 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
-
 	if (cq)
-		atomic_inc(&cq->refcount);
+		++cq->refcount;
+
 	spin_unlock(&dev->cq_table.lock);
 
 	if (!cq) {
@@ -249,8 +264,10 @@ void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
 	if (cq->ibcq.event_handler)
 		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
 
-	if (atomic_dec_and_test(&cq->refcount))
+	spin_lock(&dev->cq_table.lock);
+	if (!--cq->refcount)
 		wake_up(&cq->wait);
+	spin_unlock(&dev->cq_table.lock);
 }
 
 static inline int is_recv_cqe(struct mthca_cqe *cqe)
@@ -262,23 +279,13 @@ static inline int is_recv_cqe(struct mthca_cqe *cqe)
 		return !(cqe->is_send & 0x80);
 }
 
-void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
+void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
 		    struct mthca_srq *srq)
 {
-	struct mthca_cq *cq;
 	struct mthca_cqe *cqe;
 	u32 prod_index;
 	int nfreed = 0;
 
-	spin_lock_irq(&dev->cq_table.lock);
-	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
-	if (cq)
-		atomic_inc(&cq->refcount);
-	spin_unlock_irq(&dev->cq_table.lock);
-
-	if (!cq)
-		return;
-
 	spin_lock_irq(&cq->lock);
 
 	/*
@@ -289,14 +296,14 @@ void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
 	 * from our QP and therefore don't need to be checked.
 	 */
 	for (prod_index = cq->cons_index;
-	     cqe_sw(cq, prod_index & cq->ibcq.cqe);
+	     cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
 	     ++prod_index)
 		if (prod_index == cq->cons_index + cq->ibcq.cqe)
 			break;
 
 	if (0)
 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
-			  qpn, cqn, cq->cons_index, prod_index);
+			  qpn, cq->cqn, cq->cons_index, prod_index);
 
 	/*
 	 * Now sweep backwards through the CQ, removing CQ entries
@@ -320,16 +327,60 @@ void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
 	}
 
 	spin_unlock_irq(&cq->lock);
-	if (atomic_dec_and_test(&cq->refcount))
-		wake_up(&cq->wait);
 }
 
-static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
-			    struct mthca_qp *qp, int wqe_index, int is_send,
-			    struct mthca_err_cqe *cqe,
-			    struct ib_wc *entry, int *free_cqe)
+void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
+{
+	int i;
+
+	/*
+	 * In Tavor mode, the hardware keeps the consumer and producer
+	 * indices mod the CQ size.  Since we might be making the CQ
+	 * bigger, we need to deal with the case where the producer
+	 * index wrapped around before the CQ was resized.
+	 */
+	if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
+	    cq->ibcq.cqe < cq->resize_buf->cqe) {
+		cq->cons_index &= cq->ibcq.cqe;
+		if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
+			cq->cons_index -= cq->ibcq.cqe + 1;
+	}
+
+	for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
+		memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
+					i & cq->resize_buf->cqe),
+		       get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
+}
+
+int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
+{
+	int ret;
+	int i;
+
+	ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
+			      MTHCA_MAX_DIRECT_CQ_SIZE,
+			      &buf->queue, &buf->is_direct,
+			      &dev->driver_pd, 1, &buf->mr);
+	if (ret)
+		return ret;
+
+	for (i = 0; i < nent; ++i)
+		set_cqe_hw(get_cqe_from_buf(buf, i));
+
+	return 0;
+}
+
+void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
+{
+	mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
+		       buf->is_direct, &buf->mr);
+}
+
+static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
+			     struct mthca_qp *qp, int wqe_index, int is_send,
+			     struct mthca_err_cqe *cqe,
+			     struct ib_wc *entry, int *free_cqe)
 {
-	int err;
 	int dbd;
 	__be32 new_wqe;
 
@@ -412,11 +463,9 @@ static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
 	 * error case, so we don't have to check the doorbell count, etc.
 	 */
 	if (mthca_is_memfree(dev))
-		return 0;
+		return;
 
-	err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
-	if (err)
-		return err;
+	mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
 
 	/*
 	 * If we're at the end of the WQE chain, or we've used up our
@@ -424,15 +473,13 @@ static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
 	 * the next poll operation.
 	 */
 	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
-		return 0;
+		return;
 
 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
 	cqe->wqe      = new_wqe;
 	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
 
 	*free_cqe = 0;
-
-	return 0;
 }
 
 static inline int mthca_poll_one(struct mthca_dev *dev,
@@ -503,8 +550,17 @@ static inline int mthca_poll_one(struct mthca_dev *dev,
 		entry->wr_id = srq->wrid[wqe_index];
 		mthca_free_srq_wqe(srq, wqe);
 	} else {
+		s32 wqe;
 		wq = &(*cur_qp)->rq;
-		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
+		wqe = be32_to_cpu(cqe->wqe);
+		wqe_index = wqe >> wq->wqe_shift;
+		/*
+		 * WQE addr == base - 1 might be reported in receive completion
+		 * with error instead of (rq size - 1) by Sinai FW 1.0.800 and
+		 * Arbel FW 5.1.400.  This bug should be fixed in later FW revs.
+		 */
+		if (unlikely(wqe_index < 0))
+			wqe_index = wq->max - 1;
 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
 	}
 
@@ -518,9 +574,9 @@ static inline int mthca_poll_one(struct mthca_dev *dev,
 	}
 
 	if (is_error) {
-		err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
-				       (struct mthca_err_cqe *) cqe,
-				       entry, &free_cqe);
+		handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
+				 (struct mthca_err_cqe *) cqe,
+				 entry, &free_cqe);
 		goto out;
 	}
 
@@ -547,11 +603,11 @@ static inline int mthca_poll_one(struct mthca_dev *dev,
 			break;
 		case MTHCA_OPCODE_ATOMIC_CS:
 			entry->opcode    = IB_WC_COMP_SWAP;
-			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
+			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
 			break;
 		case MTHCA_OPCODE_ATOMIC_FA:
 			entry->opcode    = IB_WC_FETCH_ADD;
-			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
+			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
 			break;
 		case MTHCA_OPCODE_BIND_MW:
 			entry->opcode    = IB_WC_BIND_MW;
@@ -614,11 +670,14 @@ int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
 
 	spin_lock_irqsave(&cq->lock, flags);
 
-	for (npolled = 0; npolled < num_entries; ++npolled) {
+	npolled = 0;
+repoll:
+	while (npolled < num_entries) {
 		err = mthca_poll_one(dev, cq, &qp,
 				     &freed, entry + npolled);
 		if (err)
 			break;
+		++npolled;
 	}
 
 	if (freed) {
@@ -626,6 +685,42 @@ int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
 		update_cons_index(dev, cq, freed);
 	}
 
+	/*
+	 * If a CQ resize is in progress and we discovered that the
+	 * old buffer is empty, then peek in the new buffer, and if
+	 * it's not empty, switch to the new buffer and continue
+	 * polling there.
+	 */
+	if (unlikely(err == -EAGAIN && cq->resize_buf &&
+		     cq->resize_buf->state == CQ_RESIZE_READY)) {
+		/*
+		 * In Tavor mode, the hardware keeps the producer
+		 * index modulo the CQ size.  Since we might be making
+		 * the CQ bigger, we need to mask our consumer index
+		 * using the size of the old CQ buffer before looking
+		 * in the new CQ buffer.
+		 */
+		if (!mthca_is_memfree(dev))
+			cq->cons_index &= cq->ibcq.cqe;
+
+		if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
+					    cq->cons_index & cq->resize_buf->cqe))) {
+			struct mthca_cq_buf tbuf;
+			int tcqe;
+
+			tbuf         = cq->buf;
+			tcqe         = cq->ibcq.cqe;
+			cq->buf      = cq->resize_buf->buf;
+			cq->ibcq.cqe = cq->resize_buf->cqe;
+
+			cq->resize_buf->buf   = tbuf;
+			cq->resize_buf->cqe   = tcqe;
+			cq->resize_buf->state = CQ_RESIZE_SWAPPED;
+
+			goto repoll;
+		}
+	}
+
 	spin_unlock_irqrestore(&cq->lock, flags);
 
 	return err == 0 || err == -EAGAIN ? npolled : err;
@@ -684,24 +779,14 @@ int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
 	return 0;
 }
 
-static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
-{
-	mthca_buf_free(dev, (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
-		       &cq->queue, cq->is_direct, &cq->mr);
-}
-
 int mthca_init_cq(struct mthca_dev *dev, int nent,
 		  struct mthca_ucontext *ctx, u32 pdn,
 		  struct mthca_cq *cq)
 {
-	int size = nent * MTHCA_CQ_ENTRY_SIZE;
 	struct mthca_mailbox *mailbox;
 	struct mthca_cq_context *cq_context;
 	int err = -ENOMEM;
 	u8 status;
-	int i;
-
-	might_sleep();
 
 	cq->ibcq.cqe  = nent - 1;
 	cq->is_kernel = !ctx;
@@ -739,19 +824,15 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 	cq_context = mailbox->buf;
 
 	if (cq->is_kernel) {
-		err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_CQ_SIZE,
-				      &cq->queue, &cq->is_direct,
-				      &dev->driver_pd, 1, &cq->mr);
+		err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
 		if (err)
 			goto err_out_mailbox;
-
-		for (i = 0; i < nent; ++i)
-			set_cqe_hw(get_cqe(cq, i));
 	}
 
 	spin_lock_init(&cq->lock);
-	atomic_set(&cq->refcount, 1);
+	cq->refcount = 1;
 	init_waitqueue_head(&cq->wait);
+	mutex_init(&cq->mutex);
 
 	memset(cq_context, 0, sizeof *cq_context);
 	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
@@ -765,7 +846,7 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
 	cq_context->pd              = cpu_to_be32(pdn);
-	cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
+	cq_context->lkey            = cpu_to_be32(cq->buf.mr.ibmr.lkey);
 	cq_context->cqn             = cpu_to_be32(cq->cqn);
 
 	if (mthca_is_memfree(dev)) {
@@ -803,7 +884,7 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 
 err_out_free_mr:
 	if (cq->is_kernel)
-		mthca_free_cq_buf(dev, cq);
+		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
 
 err_out_mailbox:
 	mthca_free_mailbox(dev, mailbox);
@@ -825,6 +906,17 @@ err_out:
 	return err;
 }
 
+static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
+{
+	int c;
+
+	spin_lock_irq(&dev->cq_table.lock);
+	c = cq->refcount;
+	spin_unlock_irq(&dev->cq_table.lock);
+
+	return c;
+}
+
 void mthca_free_cq(struct mthca_dev *dev,
 		   struct mthca_cq *cq)
 {
@@ -832,8 +924,6 @@ void mthca_free_cq(struct mthca_dev *dev,
 	int err;
 	u8 status;
 
-	might_sleep();
-
 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 	if (IS_ERR(mailbox)) {
 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
@@ -860,6 +950,7 @@ void mthca_free_cq(struct mthca_dev *dev,
 	spin_lock_irq(&dev->cq_table.lock);
 	mthca_array_clear(&dev->cq_table.cq,
 			  cq->cqn & (dev->limits.num_cqs - 1));
+	--cq->refcount;
 	spin_unlock_irq(&dev->cq_table.lock);
 
 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
@@ -867,11 +958,10 @@ void mthca_free_cq(struct mthca_dev *dev,
 	else
 		synchronize_irq(dev->pdev->irq);
 
-	atomic_dec(&cq->refcount);
-	wait_event(cq->wait, !atomic_read(&cq->refcount));
+	wait_event(cq->wait, !get_cq_refcount(dev, cq));
 
 	if (cq->is_kernel) {
-		mthca_free_cq_buf(dev, cq);
+		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
 		if (mthca_is_memfree(dev)) {
 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
 			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
@@ -883,7 +973,7 @@ void mthca_free_cq(struct mthca_dev *dev,
 	mthca_free_mailbox(dev, mailbox);
 }
 
-int __devinit mthca_init_cq_table(struct mthca_dev *dev)
+int mthca_init_cq_table(struct mthca_dev *dev)
 {
 	int err;
 
@@ -904,7 +994,7 @@ int __devinit mthca_init_cq_table(struct mthca_dev *dev)
 	return err;
 }
 
-void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
+void mthca_cleanup_cq_table(struct mthca_dev *dev)
 {
 	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
 	mthca_alloc_cleanup(&dev->cq_table.alloc);