1 files changed, 165 insertions, 118 deletions

diff --git a/drivers/infiniband/hw/mthca/mthca_srq.c b/drivers/infiniband/hw/mthca/mthca_srq.c
index e7e153d9c4c..d22f970480c 100644
--- a/drivers/infiniband/hw/mthca/mthca_srq.c
+++ b/drivers/infiniband/hw/mthca/mthca_srq.c
@@ -28,12 +28,13 @@
  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
- *
- * $Id: mthca_srq.c 3047 2005-08-10 03:59:35Z roland $
  */
 
 #include <linux/slab.h>
 #include <linux/string.h>
+#include <linux/sched.h>
+
+#include <asm/io.h>
 
 #include "mthca_dev.h"
 #include "mthca_cmd.h"
@@ -49,7 +50,8 @@ struct mthca_tavor_srq_context {
 	__be32 state_pd;
 	__be32 lkey;
 	__be32 uar;
-	__be32 wqe_cnt;
+	__be16 limit_watermark;
+	__be16 wqe_cnt;
 	u32    reserved[2];
 };
 
@@ -113,11 +115,16 @@ static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
 					 struct mthca_srq *srq,
 					 struct mthca_arbel_srq_context *context)
 {
-	int logsize;
+	int logsize, max;
 
 	memset(context, 0, sizeof *context);
 
-	logsize = long_log2(srq->max) + srq->wqe_shift;
+	/*
+	 * Put max in a temporary variable to work around gcc bug
+	 * triggered by ilog2() on sparc64.
+	 */
+	max = srq->max;
+	logsize = ilog2(max);
 	context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
 	context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
 	context->db_index = cpu_to_be32(srq->db_index);
@@ -166,9 +173,17 @@ static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
 	 * scatter list L_Keys to the sentry value of 0x100.
 	 */
 	for (i = 0; i < srq->max; ++i) {
-		wqe = get_wqe(srq, i);
+		struct mthca_next_seg *next;
+
+		next = wqe = get_wqe(srq, i);
 
-		*wqe_to_link(wqe) = i < srq->max - 1 ? i + 1 : -1;
+		if (i < srq->max - 1) {
+			*wqe_to_link(wqe) = i + 1;
+			next->nda_op = htonl(((i + 1) << srq->wqe_shift) | 1);
+		} else {
+			*wqe_to_link(wqe) = -1;
+			next->nda_op = 0;
+		}
 
 		for (scatter = wqe + sizeof (struct mthca_next_seg);
 		     (void *) scatter < wqe + (1 << srq->wqe_shift);
@@ -185,13 +200,12 @@ int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
 		    struct ib_srq_attr *attr, struct mthca_srq *srq)
 {
 	struct mthca_mailbox *mailbox;
-	u8 status;
 	int ds;
 	int err;
 
 	/* Sanity check SRQ size before proceeding */
 	if (attr->max_wr  > dev->limits.max_srq_wqes ||
-	    attr->max_sge > dev->limits.max_sg)
+	    attr->max_sge > dev->limits.max_srq_sge)
 		return -EINVAL;
 
 	srq->max      = attr->max_wr;
@@ -200,11 +214,17 @@ int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
 
 	if (mthca_is_memfree(dev))
 		srq->max = roundup_pow_of_two(srq->max + 1);
+	else
+		srq->max = srq->max + 1;
 
 	ds = max(64UL,
 		 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
 				    srq->max_gs * sizeof (struct mthca_data_seg)));
-	srq->wqe_shift = long_log2(ds);
+
+	if (!mthca_is_memfree(dev) && (ds > dev->limits.max_desc_sz))
+		return -EINVAL;
+
+	srq->wqe_shift = ilog2(ds);
 
 	srq->srqn = mthca_alloc(&dev->srq_table.alloc);
 	if (srq->srqn == -1)
@@ -236,26 +256,21 @@ int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
 		goto err_out_mailbox;
 
 	spin_lock_init(&srq->lock);
-	atomic_set(&srq->refcount, 1);
+	srq->refcount = 1;
 	init_waitqueue_head(&srq->wait);
+	mutex_init(&srq->mutex);
 
 	if (mthca_is_memfree(dev))
 		mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
 	else
 		mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
 
-	err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn, &status);
+	err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn);
 
 	if (err) {
 		mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
 		goto err_out_free_buf;
 	}
-	if (status) {
-		mthca_warn(dev, "SW2HW_SRQ returned status 0x%02x\n",
-			   status);
-		err = -EINVAL;
-		goto err_out_free_buf;
-	}
 
 	spin_lock_irq(&dev->srq_table.lock);
 	if (mthca_array_set(&dev->srq_table.srq,
@@ -271,14 +286,15 @@ int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
 	srq->first_free = 0;
 	srq->last_free  = srq->max - 1;
 
+	attr->max_wr    = srq->max - 1;
+	attr->max_sge   = srq->max_gs;
+
 	return 0;
 
 err_out_free_srq:
-	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
+	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
 	if (err)
 		mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
-	else if (status)
-		mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
 
 err_out_free_buf:
 	if (!pd->ibpd.uobject)
@@ -300,11 +316,21 @@ err_out:
 	return err;
 }
 
+static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
+{
+	int c;
+
+	spin_lock_irq(&dev->srq_table.lock);
+	c = srq->refcount;
+	spin_unlock_irq(&dev->srq_table.lock);
+
+	return c;
+}
+
 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
 {
 	struct mthca_mailbox *mailbox;
 	int err;
-	u8 status;
 
 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
 	if (IS_ERR(mailbox)) {
@@ -312,19 +338,17 @@ void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
 		return;
 	}
 
-	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
+	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
 	if (err)
 		mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
-	else if (status)
-		mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
 
 	spin_lock_irq(&dev->srq_table.lock);
 	mthca_array_clear(&dev->srq_table.srq,
 			  srq->srqn & (dev->limits.num_srqs - 1));
+	--srq->refcount;
 	spin_unlock_irq(&dev->srq_table.lock);
 
-	atomic_dec(&srq->refcount);
-	wait_event(srq->wait, !atomic_read(&srq->refcount));
+	wait_event(srq->wait, !get_srq_refcount(dev, srq));
 
 	if (!srq->ibsrq.uobject) {
 		mthca_free_srq_buf(dev, srq);
@@ -338,26 +362,61 @@ void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
 }
 
 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
-		     enum ib_srq_attr_mask attr_mask)
-{	
+		     enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
+{
 	struct mthca_dev *dev = to_mdev(ibsrq->device);
 	struct mthca_srq *srq = to_msrq(ibsrq);
-	int ret;
-	u8 status;
+	int ret = 0;
 
 	/* We don't support resizing SRQs (yet?) */
 	if (attr_mask & IB_SRQ_MAX_WR)
 		return -EINVAL;
 
 	if (attr_mask & IB_SRQ_LIMIT) {
-		ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit, &status);
-		if (ret)
-			return ret;
-		if (status)
+		u32 max_wr = mthca_is_memfree(dev) ? srq->max - 1 : srq->max;
+		if (attr->srq_limit > max_wr)
 			return -EINVAL;
+
+		mutex_lock(&srq->mutex);
+		ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit);
+		mutex_unlock(&srq->mutex);
 	}
 
-	return 0;
+	return ret;
+}
+
+int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
+{
+	struct mthca_dev *dev = to_mdev(ibsrq->device);
+	struct mthca_srq *srq = to_msrq(ibsrq);
+	struct mthca_mailbox *mailbox;
+	struct mthca_arbel_srq_context *arbel_ctx;
+	struct mthca_tavor_srq_context *tavor_ctx;
+	int err;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+
+	err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox);
+	if (err)
+		goto out;
+
+	if (mthca_is_memfree(dev)) {
+		arbel_ctx = mailbox->buf;
+		srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
+	} else {
+		tavor_ctx = mailbox->buf;
+		srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
+	}
+
+	srq_attr->max_wr  = srq->max - 1;
+	srq_attr->max_sge = srq->max_gs;
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+
+	return err;
 }
 
 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
@@ -369,7 +428,7 @@ void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
 	spin_lock(&dev->srq_table.lock);
 	srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
 	if (srq)
-		atomic_inc(&srq->refcount);
+		++srq->refcount;
 	spin_unlock(&dev->srq_table.lock);
 
 	if (!srq) {
@@ -386,8 +445,10 @@ void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
 	srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
 
 out:
-	if (atomic_dec_and_test(&srq->refcount))
+	spin_lock(&dev->srq_table.lock);
+	if (!--srq->refcount)
 		wake_up(&srq->wait);
+	spin_unlock(&dev->srq_table.lock);
 }
 
 /*
@@ -396,16 +457,15 @@ out:
 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
 {
 	int ind;
+	struct mthca_next_seg *last_free;
 
 	ind = wqe_addr >> srq->wqe_shift;
 
 	spin_lock(&srq->lock);
 
-	if (likely(srq->first_free >= 0))
-		*wqe_to_link(get_wqe(srq, srq->last_free)) = ind;
-	else
-		srq->first_free = ind;
-
+	last_free = get_wqe(srq, srq->last_free);
+	*wqe_to_link(last_free) = ind;
+	last_free->nda_op = htonl((ind << srq->wqe_shift) | 1);
 	*wqe_to_link(get_wqe(srq, ind)) = -1;
 	srq->last_free = ind;
 
@@ -417,7 +477,6 @@ int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 {
 	struct mthca_dev *dev = to_mdev(ibsrq->device);
 	struct mthca_srq *srq = to_msrq(ibsrq);
-	__be32 doorbell[2];
 	unsigned long flags;
 	int err = 0;
 	int first_ind;
@@ -432,39 +491,12 @@ int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 
 	first_ind = srq->first_free;
 
-	for (nreq = 0; wr; ++nreq, wr = wr->next) {
-		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
-			nreq = 0;
-
-			doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
-			doorbell[1] = cpu_to_be32(srq->srqn << 8);
-
-			/*
-			 * Make sure that descriptors are written
-			 * before doorbell is rung.
-			 */
-			wmb();
-
-			mthca_write64(doorbell,
-				      dev->kar + MTHCA_RECEIVE_DOORBELL,
-				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
-
-			first_ind = srq->first_free;
-		}
-
-		ind = srq->first_free;
-
-		if (ind < 0) {
-			mthca_err(dev, "SRQ %06x full\n", srq->srqn);
-			err = -ENOMEM;
-			*bad_wr = wr;
-			break;
-		}
-
+	for (nreq = 0; wr; wr = wr->next) {
+		ind       = srq->first_free;
 		wqe       = get_wqe(srq, ind);
 		next_ind  = *wqe_to_link(wqe);
 
-		if (next_ind < 0) {
+		if (unlikely(next_ind < 0)) {
 			mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 			err = -ENOMEM;
 			*bad_wr = wr;
@@ -474,7 +506,6 @@ int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 		prev_wqe  = srq->last;
 		srq->last = wqe;
 
-		((struct mthca_next_seg *) wqe)->nda_op = 0;
 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
 		/* flags field will always remain 0 */
 
@@ -488,46 +519,55 @@ int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 		}
 
 		for (i = 0; i < wr->num_sge; ++i) {
-			((struct mthca_data_seg *) wqe)->byte_count =
-				cpu_to_be32(wr->sg_list[i].length);
-			((struct mthca_data_seg *) wqe)->lkey =
-				cpu_to_be32(wr->sg_list[i].lkey);
-			((struct mthca_data_seg *) wqe)->addr =
-				cpu_to_be64(wr->sg_list[i].addr);
+			mthca_set_data_seg(wqe, wr->sg_list + i);
 			wqe += sizeof (struct mthca_data_seg);
 		}
 
-		if (i < srq->max_gs) {
-			((struct mthca_data_seg *) wqe)->byte_count = 0;
-			((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
-			((struct mthca_data_seg *) wqe)->addr = 0;
-		}
+		if (i < srq->max_gs)
+			mthca_set_data_seg_inval(wqe);
 
-		((struct mthca_next_seg *) prev_wqe)->nda_op =
-			cpu_to_be32((ind << srq->wqe_shift) | 1);
-		wmb();
 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
 			cpu_to_be32(MTHCA_NEXT_DBD);
 
 		srq->wrid[ind]  = wr->wr_id;
 		srq->first_free = next_ind;
+
+		++nreq;
+		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
+			nreq = 0;
+
+			/*
+			 * Make sure that descriptors are written
+			 * before doorbell is rung.
+			 */
+			wmb();
+
+			mthca_write64(first_ind << srq->wqe_shift, srq->srqn << 8,
+				      dev->kar + MTHCA_RECEIVE_DOORBELL,
+				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+
+			first_ind = srq->first_free;
+		}
 	}
 
 	if (likely(nreq)) {
-		doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
-		doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
-
 		/*
 		 * Make sure that descriptors are written before
 		 * doorbell is rung.
 		 */
 		wmb();
 
-		mthca_write64(doorbell,
+		mthca_write64(first_ind << srq->wqe_shift, (srq->srqn << 8) | nreq,
 			      dev->kar + MTHCA_RECEIVE_DOORBELL,
 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
 	}
 
+	/*
+	 * Make sure doorbells don't leak out of SRQ spinlock and
+	 * reach the HCA out of order:
+	 */
+	mmiowb();
+
 	spin_unlock_irqrestore(&srq->lock, flags);
 	return err;
 }
@@ -548,27 +588,17 @@ int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 	spin_lock_irqsave(&srq->lock, flags);
 
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
-		ind = srq->first_free;
-
-		if (ind < 0) {
-			mthca_err(dev, "SRQ %06x full\n", srq->srqn);
-			err = -ENOMEM;
-			*bad_wr = wr;
-			break;
-		}
-
+		ind       = srq->first_free;
 		wqe       = get_wqe(srq, ind);
 		next_ind  = *wqe_to_link(wqe);
 
-		if (next_ind < 0) {
+		if (unlikely(next_ind < 0)) {
 			mthca_err(dev, "SRQ %06x full\n", srq->srqn);
 			err = -ENOMEM;
 			*bad_wr = wr;
 			break;
 		}
 
-		((struct mthca_next_seg *) wqe)->nda_op =
-			cpu_to_be32((next_ind << srq->wqe_shift) | 1);
 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
 		/* flags field will always remain 0 */
 
@@ -581,20 +611,12 @@ int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 		}
 
 		for (i = 0; i < wr->num_sge; ++i) {
-			((struct mthca_data_seg *) wqe)->byte_count =
-				cpu_to_be32(wr->sg_list[i].length);
-			((struct mthca_data_seg *) wqe)->lkey =
-				cpu_to_be32(wr->sg_list[i].lkey);
-			((struct mthca_data_seg *) wqe)->addr =
-				cpu_to_be64(wr->sg_list[i].addr);
+			mthca_set_data_seg(wqe, wr->sg_list + i);
 			wqe += sizeof (struct mthca_data_seg);
 		}
 
-		if (i < srq->max_gs) {
-			((struct mthca_data_seg *) wqe)->byte_count = 0;
-			((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
-			((struct mthca_data_seg *) wqe)->addr = 0;
-		}
+		if (i < srq->max_gs)
+			mthca_set_data_seg_inval(wqe);
 
 		srq->wrid[ind]  = wr->wr_id;
 		srq->first_free = next_ind;
@@ -615,7 +637,32 @@ int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 	return err;
 }
 
-int __devinit mthca_init_srq_table(struct mthca_dev *dev)
+int mthca_max_srq_sge(struct mthca_dev *dev)
+{
+	if (mthca_is_memfree(dev))
+		return dev->limits.max_sg;
+
+	/*
+	 * SRQ allocations are based on powers of 2 for Tavor,
+	 * (although they only need to be multiples of 16 bytes).
+	 *
+	 * Therefore, we need to base the max number of sg entries on
+	 * the largest power of 2 descriptor size that is <= to the
+	 * actual max WQE descriptor size, rather than return the
+	 * max_sg value given by the firmware (which is based on WQE
+	 * sizes as multiples of 16, not powers of 2).
+	 *
+	 * If SRQ implementation is changed for Tavor to be based on
+	 * multiples of 16, the calculation below can be deleted and
+	 * the FW max_sg value returned.
+	 */
+	return min_t(int, dev->limits.max_sg,
+		     ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
+		      sizeof (struct mthca_next_seg)) /
+		     sizeof (struct mthca_data_seg));
+}
+
+int mthca_init_srq_table(struct mthca_dev *dev)
 {
 	int err;
 
@@ -639,7 +686,7 @@ int __devinit mthca_init_srq_table(struct mthca_dev *dev)
 	return err;
 }
 
-void __devexit mthca_cleanup_srq_table(struct mthca_dev *dev)
+void mthca_cleanup_srq_table(struct mthca_dev *dev)
 {
 	if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
 		return;