[PATCH] IB/mthca: Add SRQ implementation

Add mthca support for shared receive queues (SRQs),
including userspace SRQs.

Signed-off-by: Roland Dreier <[e-mail]>

drivers/infiniband/hw/mthca/Makefile

···
 ib_mthca-y :=	mthca_main.o mthca_cmd.o mthca_profile.o mthca_reset.o \
 		mthca_allocator.o mthca_eq.o mthca_pd.o mthca_cq.o \
 		mthca_mr.o mthca_qp.o mthca_av.o mthca_mcg.o mthca_mad.o \
-		mthca_provider.o mthca_memfree.o mthca_uar.o
+		mthca_provider.o mthca_memfree.o mthca_uar.o mthca_srq.o

drivers/infiniband/hw/mthca/mthca_cmd.c

···
 	CMD_SW2HW_SRQ 	    = 0x35,
 	CMD_HW2SW_SRQ 	    = 0x36,
 	CMD_QUERY_SRQ       = 0x37,
+	CMD_ARM_SRQ         = 0x40,
 
 	/* QP/EE commands */
 	CMD_RST2INIT_QPEE   = 0x19,
···
 
 	mthca_dbg(dev, "Max QPs: %d, reserved QPs: %d, entry size: %d\n",
 		  dev_lim->max_qps, dev_lim->reserved_qps, dev_lim->qpc_entry_sz);
+	mthca_dbg(dev, "Max SRQs: %d, reserved SRQs: %d, entry size: %d\n",
+		  dev_lim->max_srqs, dev_lim->reserved_srqs, dev_lim->srq_entry_sz);
 	mthca_dbg(dev, "Max CQs: %d, reserved CQs: %d, entry size: %d\n",
 		  dev_lim->max_cqs, dev_lim->reserved_cqs, dev_lim->cqc_entry_sz);
 	mthca_dbg(dev, "Max EQs: %d, reserved EQs: %d, entry size: %d\n",
···
 	return mthca_cmd_box(dev, 0, mailbox->dma, cq_num, 0,
 			     CMD_HW2SW_CQ,
 			     CMD_TIME_CLASS_A, status);
+}
+
+int mthca_SW2HW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int srq_num, u8 *status)
+{
+	return mthca_cmd(dev, mailbox->dma, srq_num, 0, CMD_SW2HW_SRQ,
+			CMD_TIME_CLASS_A, status);
+}
+
+int mthca_HW2SW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int srq_num, u8 *status)
+{
+	return mthca_cmd_box(dev, 0, mailbox->dma, srq_num, 0,
+			     CMD_HW2SW_SRQ,
+			     CMD_TIME_CLASS_A, status);
+}
+
+int mthca_ARM_SRQ(struct mthca_dev *dev, int srq_num, int limit, u8 *status)
+{
+	return mthca_cmd(dev, limit, srq_num, 0, CMD_ARM_SRQ,
+			 CMD_TIME_CLASS_B, status);
 }
 
 int mthca_MODIFY_QP(struct mthca_dev *dev, int trans, u32 num,

drivers/infiniband/hw/mthca/mthca_cmd.h

···
 		   int cq_num, u8 *status);
 int mthca_HW2SW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int cq_num, u8 *status);
+int mthca_SW2HW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int srq_num, u8 *status);
+int mthca_HW2SW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int srq_num, u8 *status);
+int mthca_ARM_SRQ(struct mthca_dev *dev, int srq_num, int limit, u8 *status);
 int mthca_MODIFY_QP(struct mthca_dev *dev, int trans, u32 num,
 		    int is_ee, struct mthca_mailbox *mailbox, u32 optmask,
 		    u8 *status);

drivers/infiniband/hw/mthca/mthca_cq.c

···
 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
 }
 
-void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn)
+void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
+		    struct mthca_srq *srq)
 {
 	struct mthca_cq *cq;
 	struct mthca_cqe *cqe;
···
 	 */
 	while (prod_index > cq->cons_index) {
 		cqe = get_cqe(cq, (prod_index - 1) & cq->ibcq.cqe);
-		if (cqe->my_qpn == cpu_to_be32(qpn))
+		if (cqe->my_qpn == cpu_to_be32(qpn)) {
+			if (srq)
+				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
 			++nfreed;
+		}
 		else if (nfreed)
 			memcpy(get_cqe(cq, (prod_index - 1 + nfreed) &
 				       cq->ibcq.cqe),
···
 			     >> wq->wqe_shift);
 		entry->wr_id = (*cur_qp)->wrid[wqe_index +
 					       (*cur_qp)->rq.max];
+	} else if ((*cur_qp)->ibqp.srq) {
+		struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
+		u32 wqe = be32_to_cpu(cqe->wqe);
+		wq = NULL;
+		wqe_index = wqe >> srq->wqe_shift;
+		entry->wr_id = srq->wrid[wqe_index];
+		mthca_free_srq_wqe(srq, wqe);
 	} else {
 		wq = &(*cur_qp)->rq;
 		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
 		entry->wr_id = (*cur_qp)->wrid[wqe_index];
 	}
 
-	if (wq->last_comp < wqe_index)
-		wq->tail += wqe_index - wq->last_comp;
-	else
-		wq->tail += wqe_index + wq->max - wq->last_comp;
+	if (wq) {
+		if (wq->last_comp < wqe_index)
+			wq->tail += wqe_index - wq->last_comp;
+		else
+			wq->tail += wqe_index + wq->max - wq->last_comp;
 
-	wq->last_comp = wqe_index;
-
-	if (0)
-		mthca_dbg(dev, "%s completion for QP %06x, index %d (nr %d)\n",
-			  is_send ? "Send" : "Receive",
-			  (*cur_qp)->qpn, wqe_index, wq->max);
+		wq->last_comp = wqe_index;
+	}
 
 	if (is_error) {
 		err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,

drivers/infiniband/hw/mthca/mthca_dev.h

···
 	struct mthca_icm_table *table;
 };
 
+struct mthca_srq_table {
+	struct mthca_alloc 	alloc;
+	spinlock_t         	lock;
+	struct mthca_array      srq;
+	struct mthca_icm_table *table;
+};
+
 struct mthca_qp_table {
 	struct mthca_alloc     	alloc;
 	u32                    	rdb_base;
···
 	struct mthca_mr_table  mr_table;
 	struct mthca_eq_table  eq_table;
 	struct mthca_cq_table  cq_table;
+	struct mthca_srq_table srq_table;
 	struct mthca_qp_table  qp_table;
 	struct mthca_av_table  av_table;
 	struct mthca_mcg_table mcg_table;
···
 int mthca_init_mr_table(struct mthca_dev *dev);
 int mthca_init_eq_table(struct mthca_dev *dev);
 int mthca_init_cq_table(struct mthca_dev *dev);
+int mthca_init_srq_table(struct mthca_dev *dev);
 int mthca_init_qp_table(struct mthca_dev *dev);
 int mthca_init_av_table(struct mthca_dev *dev);
 int mthca_init_mcg_table(struct mthca_dev *dev);
···
 void mthca_cleanup_mr_table(struct mthca_dev *dev);
 void mthca_cleanup_eq_table(struct mthca_dev *dev);
 void mthca_cleanup_cq_table(struct mthca_dev *dev);
+void mthca_cleanup_srq_table(struct mthca_dev *dev);
 void mthca_cleanup_qp_table(struct mthca_dev *dev);
 void mthca_cleanup_av_table(struct mthca_dev *dev);
 void mthca_cleanup_mcg_table(struct mthca_dev *dev);
···
 void mthca_free_cq(struct mthca_dev *dev,
 		   struct mthca_cq *cq);
 void mthca_cq_event(struct mthca_dev *dev, u32 cqn);
-void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn);
+void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
+		    struct mthca_srq *srq);
+
+int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
+		    struct ib_srq_attr *attr, struct mthca_srq *srq);
+void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq);
+void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
+		     enum ib_event_type event_type);
+void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr);
+int mthca_tavor_post_srq_recv(struct ib_srq *srq, struct ib_recv_wr *wr,
+			      struct ib_recv_wr **bad_wr);
+int mthca_arbel_post_srq_recv(struct ib_srq *srq, struct ib_recv_wr *wr,
+			      struct ib_recv_wr **bad_wr);
 
 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
 		    enum ib_event_type event_type);

drivers/infiniband/hw/mthca/mthca_main.c

···
 	profile = default_profile;
 	profile.num_uar   = dev_lim.uar_size / PAGE_SIZE;
 	profile.uarc_size = 0;
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+		profile.num_srq = dev_lim.max_srqs;
 
 	err = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
 	if (err < 0)
···
 	}
 
        mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
-						     dev_lim->cqc_entry_sz,
-						     mdev->limits.num_cqs,
-						     mdev->limits.reserved_cqs, 0);
+						    dev_lim->cqc_entry_sz,
+						    mdev->limits.num_cqs,
+						    mdev->limits.reserved_cqs, 0);
 	if (!mdev->cq_table.table) {
 		mthca_err(mdev, "Failed to map CQ context memory, aborting.\n");
 		err = -ENOMEM;
 		goto err_unmap_rdb;
+	}
+
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ) {
+		mdev->srq_table.table =
+			mthca_alloc_icm_table(mdev, init_hca->srqc_base,
+					      dev_lim->srq_entry_sz,
+					      mdev->limits.num_srqs,
+					      mdev->limits.reserved_srqs, 0);
+		if (!mdev->srq_table.table) {
+			mthca_err(mdev, "Failed to map SRQ context memory, "
+				  "aborting.\n");
+			err = -ENOMEM;
+			goto err_unmap_cq;
+		}
 	}
 
 	/*
···
 	if (!mdev->mcg_table.table) {
 		mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
 		err = -ENOMEM;
-		goto err_unmap_cq;
+		goto err_unmap_srq;
 	}
 
 	return 0;
+
+err_unmap_srq:
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+		mthca_free_icm_table(mdev, mdev->srq_table.table);
 
 err_unmap_cq:
 	mthca_free_icm_table(mdev, mdev->cq_table.table);
···
 	profile = default_profile;
 	profile.num_uar  = dev_lim.uar_size / PAGE_SIZE;
 	profile.num_udav = 0;
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+		profile.num_srq = dev_lim.max_srqs;
 
 	icm_size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
 	if ((int) icm_size < 0) {
···
 	return 0;
 
 err_free_icm:
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+		mthca_free_icm_table(mdev, mdev->srq_table.table);
 	mthca_free_icm_table(mdev, mdev->cq_table.table);
 	mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
 	mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
···
 	mthca_CLOSE_HCA(mdev, 0, &status);
 
 	if (mthca_is_memfree(mdev)) {
+		if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+			mthca_free_icm_table(mdev, mdev->srq_table.table);
 		mthca_free_icm_table(mdev, mdev->cq_table.table);
 		mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
 		mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
···
 		goto err_cmd_poll;
 	}
 
+	err = mthca_init_srq_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "shared receive queue table, aborting.\n");
+		goto err_cq_table_free;
+	}
+
 	err = mthca_init_qp_table(dev);
 	if (err) {
 		mthca_err(dev, "Failed to initialize "
 			  "queue pair table, aborting.\n");
-		goto err_cq_table_free;
+		goto err_srq_table_free;
 	}
 
 	err = mthca_init_av_table(dev);
···
 
 err_qp_table_free:
 	mthca_cleanup_qp_table(dev);
+
+err_srq_table_free:
+	mthca_cleanup_srq_table(dev);
 
 err_cq_table_free:
 	mthca_cleanup_cq_table(dev);
···
 	mthca_cleanup_mcg_table(mdev);
 	mthca_cleanup_av_table(mdev);
 	mthca_cleanup_qp_table(mdev);
+	mthca_cleanup_srq_table(mdev);
 	mthca_cleanup_cq_table(mdev);
 	mthca_cmd_use_polling(mdev);
 	mthca_cleanup_eq_table(mdev);
···
 		mthca_cleanup_mcg_table(mdev);
 		mthca_cleanup_av_table(mdev);
 		mthca_cleanup_qp_table(mdev);
+		mthca_cleanup_srq_table(mdev);
 		mthca_cleanup_cq_table(mdev);
 		mthca_cmd_use_polling(mdev);
 		mthca_cleanup_eq_table(mdev);

drivers/infiniband/hw/mthca/mthca_profile.c

···
 	profile[MTHCA_RES_UARC].size = request->uarc_size;
 
 	profile[MTHCA_RES_QP].num    = request->num_qp;
+	profile[MTHCA_RES_SRQ].num   = request->num_srq;
 	profile[MTHCA_RES_EQP].num   = request->num_qp;
 	profile[MTHCA_RES_RDB].num   = request->num_qp * request->rdb_per_qp;
 	profile[MTHCA_RES_CQ].num    = request->num_cq;

drivers/infiniband/hw/mthca/mthca_profile.h

···
 struct mthca_profile {
 	int num_qp;
 	int rdb_per_qp;
+	int num_srq;
 	int num_cq;
 	int num_mcg;
 	int num_mpt;

drivers/infiniband/hw/mthca/mthca_provider.c

···
 	return 0;
 }
 
+static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
+				       struct ib_srq_init_attr *init_attr,
+				       struct ib_udata *udata)
+{
+	struct mthca_create_srq ucmd;
+	struct mthca_ucontext *context = NULL;
+	struct mthca_srq *srq;
+	int err;
+
+	srq = kmalloc(sizeof *srq, GFP_KERNEL);
+	if (!srq)
+		return ERR_PTR(-ENOMEM);
+
+	if (pd->uobject) {
+		context = to_mucontext(pd->uobject->context);
+
+		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
+			return ERR_PTR(-EFAULT);
+
+		err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
+					context->db_tab, ucmd.db_index,
+					ucmd.db_page);
+
+		if (err)
+			goto err_free;
+
+		srq->mr.ibmr.lkey = ucmd.lkey;
+		srq->db_index     = ucmd.db_index;
+	}
+
+	err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
+			      &init_attr->attr, srq);
+
+	if (err && pd->uobject)
+		mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
+				    context->db_tab, ucmd.db_index);
+
+	if (err)
+		goto err_free;
+
+	if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
+		mthca_free_srq(to_mdev(pd->device), srq);
+		err = -EFAULT;
+		goto err_free;
+	}
+
+	return &srq->ibsrq;
+
+err_free:
+	kfree(srq);
+
+	return ERR_PTR(err);
+}
+
+static int mthca_destroy_srq(struct ib_srq *srq)
+{
+	struct mthca_ucontext *context;
+
+	if (srq->uobject) {
+		context = to_mucontext(srq->uobject->context);
+
+		mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
+				    context->db_tab, to_msrq(srq)->db_index);
+	}
+
+	mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
+	kfree(srq);
+
+	return 0;
+}
+
 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
 				     struct ib_qp_init_attr *init_attr,
 				     struct ib_udata *udata)
···
 	dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
 	dev->ib_dev.create_ah            = mthca_ah_create;
 	dev->ib_dev.destroy_ah           = mthca_ah_destroy;
+
+	if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
+		dev->ib_dev.create_srq           = mthca_create_srq;
+		dev->ib_dev.destroy_srq          = mthca_destroy_srq;
+
+		if (mthca_is_memfree(dev))
+			dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
+		else
+			dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
+	}
+
 	dev->ib_dev.create_qp            = mthca_create_qp;
 	dev->ib_dev.modify_qp            = mthca_modify_qp;
 	dev->ib_dev.destroy_qp           = mthca_destroy_qp;

drivers/infiniband/hw/mthca/mthca_provider.h

···
 	wait_queue_head_t      wait;
 };
 
+struct mthca_srq {
+	struct ib_srq		ibsrq;
+	spinlock_t		lock;
+	atomic_t		refcount;
+	int			srqn;
+	int			max;
+	int			max_gs;
+	int			wqe_shift;
+	int			first_free;
+	int			last_free;
+	u16			counter;  /* Arbel only */
+	int			db_index; /* Arbel only */
+	__be32		       *db;       /* Arbel only */
+	void		       *last;
+
+	int			is_direct;
+	u64		       *wrid;
+	union mthca_buf		queue;
+	struct mthca_mr		mr;
+
+	wait_queue_head_t	wait;
+};
+
 struct mthca_wq {
 	spinlock_t lock;
 	int        max;
···
 static inline struct mthca_cq *to_mcq(struct ib_cq *ibcq)
 {
 	return container_of(ibcq, struct mthca_cq, ibcq);
+}
+
+static inline struct mthca_srq *to_msrq(struct ib_srq *ibsrq)
+{
+	return container_of(ibsrq, struct mthca_srq, ibsrq);
 }
 
 static inline struct mthca_qp *to_mqp(struct ib_qp *ibqp)

drivers/infiniband/hw/mthca/mthca_qp.c

···
 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
 
 	if (mthca_is_memfree(dev)) {
-		qp_context->rq_size_stride =
-			((ffs(qp->rq.max) - 1) << 3) | (qp->rq.wqe_shift - 4);
-		qp_context->sq_size_stride =
-			((ffs(qp->sq.max) - 1) << 3) | (qp->sq.wqe_shift - 4);
+		if (qp->rq.max)
+			qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
+		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
+
+		if (qp->sq.max)
+			qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
+		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
 	}
 
 	/* leave arbel_sched_queue as 0 */
···
 
 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
 
+	if (ibqp->srq)
+		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
+
 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
···
 		qp_context->qkey = cpu_to_be32(attr->qkey);
 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
 	}
+
+	if (ibqp->srq)
+		qp_context->srqn = cpu_to_be32(1 << 24 |
+					       to_msrq(ibqp->srq)->srqn);
 
 	err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
 			      qp->qpn, 0, mailbox, 0, &status);
···
 	 * unref the mem-free tables and free the QPN in our table.
 	 */
 	if (!qp->ibqp.uobject) {
-		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn);
+		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
+			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
-			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn);
+			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
+				       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 
 		mthca_free_memfree(dev, qp);
 		mthca_free_wqe_buf(dev, qp);
···
 		       int index, int *dbd, __be32 *new_wqe)
 {
 	struct mthca_next_seg *next;
+
+	/*
+	 * For SRQs, all WQEs generate a CQE, so we're always at the
+	 * end of the doorbell chain.
+	 */
+	if (qp->ibqp.srq) {
+		*new_wqe = 0;
+		return 0;
+	}
 
 	if (is_send)
 		next = get_send_wqe(qp, index);

drivers/infiniband/hw/mthca/mthca_srq.c

···
+/*
+ * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * 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 "mthca_dev.h"
+#include "mthca_cmd.h"
+#include "mthca_memfree.h"
+#include "mthca_wqe.h"
+
+enum {
+	MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
+};
+
+struct mthca_tavor_srq_context {
+	__be64 wqe_base_ds;	/* low 6 bits is descriptor size */
+	__be32 state_pd;
+	__be32 lkey;
+	__be32 uar;
+	__be32 wqe_cnt;
+	u32    reserved[2];
+};
+
+struct mthca_arbel_srq_context {
+	__be32 state_logsize_srqn;
+	__be32 lkey;
+	__be32 db_index;
+	__be32 logstride_usrpage;
+	__be64 wqe_base;
+	__be32 eq_pd;
+	__be16 limit_watermark;
+	__be16 wqe_cnt;
+	u16    reserved1;
+	__be16 wqe_counter;
+	u32    reserved2[3];
+};
+
+static void *get_wqe(struct mthca_srq *srq, int n)
+{
+	if (srq->is_direct)
+		return srq->queue.direct.buf + (n << srq->wqe_shift);
+	else
+		return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
+			((n << srq->wqe_shift) & (PAGE_SIZE - 1));
+}
+
+/*
+ * Return a pointer to the location within a WQE that we're using as a
+ * link when the WQE is in the free list.  We use an offset of 4
+ * because in the Tavor case, posting a WQE may overwrite the first
+ * four bytes of the previous WQE.  The offset avoids corrupting our
+ * free list if the WQE has already completed and been put on the free
+ * list when we post the next WQE.
+ */
+static inline int *wqe_to_link(void *wqe)
+{
+	return (int *) (wqe + 4);
+}
+
+static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
+					 struct mthca_pd *pd,
+					 struct mthca_srq *srq,
+					 struct mthca_tavor_srq_context *context)
+{
+	memset(context, 0, sizeof *context);
+
+	context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
+	context->state_pd    = cpu_to_be32(pd->pd_num);
+	context->lkey        = cpu_to_be32(srq->mr.ibmr.lkey);
+
+	if (pd->ibpd.uobject)
+		context->uar =
+			cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
+	else
+		context->uar = cpu_to_be32(dev->driver_uar.index);
+}
+
+static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
+					 struct mthca_pd *pd,
+					 struct mthca_srq *srq,
+					 struct mthca_arbel_srq_context *context)
+{
+	int logsize;
+
+	memset(context, 0, sizeof *context);
+
+	logsize = long_log2(srq->max) + srq->wqe_shift;
+	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);
+	context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
+	if (pd->ibpd.uobject)
+		context->logstride_usrpage |=
+			cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
+	else
+		context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
+	context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
+}
+
+static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
+{
+	mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
+		       srq->is_direct, &srq->mr);
+	kfree(srq->wrid);
+}
+
+static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
+			       struct mthca_srq *srq)
+{
+	struct mthca_data_seg *scatter;
+	void *wqe;
+	int err;
+	int i;
+
+	if (pd->ibpd.uobject)
+		return 0;
+
+	srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
+	if (!srq->wrid)
+		return -ENOMEM;
+
+	err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
+			      MTHCA_MAX_DIRECT_SRQ_SIZE,
+			      &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
+	if (err) {
+		kfree(srq->wrid);
+		return err;
+	}
+
+	/*
+	 * Now initialize the SRQ buffer so that all of the WQEs are
+	 * linked into the list of free WQEs.  In addition, set the
+	 * scatter list L_Keys to the sentry value of 0x100.
+	 */
+	for (i = 0; i < srq->max; ++i) {
+		wqe = get_wqe(srq, i);
+
+		*wqe_to_link(wqe) = i < srq->max - 1 ? i + 1 : -1;
+
+		for (scatter = wqe + sizeof (struct mthca_next_seg);
+		     (void *) scatter < wqe + (1 << srq->wqe_shift);
+		     ++scatter)
+			scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
+	}
+
+	return 0;
+}
+
+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 > 16 << 20 || attr->max_sge > 64)
+		return -EINVAL;
+
+	srq->max      = attr->max_wr;
+	srq->max_gs   = attr->max_sge;
+	srq->last     = NULL;
+	srq->counter  = 0;
+
+	if (mthca_is_memfree(dev))
+		srq->max = roundup_pow_of_two(srq->max + 1);
+
+	ds = min(64UL,
+		 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
+				    srq->max_gs * sizeof (struct mthca_data_seg)));
+	srq->wqe_shift = long_log2(ds);
+
+	srq->srqn = mthca_alloc(&dev->srq_table.alloc);
+	if (srq->srqn == -1)
+		return -ENOMEM;
+
+	if (mthca_is_memfree(dev)) {
+		err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
+		if (err)
+			goto err_out;
+
+		if (!pd->ibpd.uobject) {
+			srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
+						       srq->srqn, &srq->db);
+			if (srq->db_index < 0) {
+				err = -ENOMEM;
+				goto err_out_icm;
+			}
+		}
+	}
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
+		goto err_out_db;
+	}
+
+	err = mthca_alloc_srq_buf(dev, pd, srq);
+	if (err)
+		goto err_out_mailbox;
+
+	spin_lock_init(&srq->lock);
+	atomic_set(&srq->refcount, 1);
+	init_waitqueue_head(&srq->wait);
+
+	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);
+
+	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,
+			    srq->srqn & (dev->limits.num_srqs - 1),
+			    srq)) {
+		spin_unlock_irq(&dev->srq_table.lock);
+		goto err_out_free_srq;
+	}
+	spin_unlock_irq(&dev->srq_table.lock);
+
+	mthca_free_mailbox(dev, mailbox);
+
+	srq->first_free = 0;
+	srq->last_free  = srq->max - 1;
+
+	return 0;
+
+err_out_free_srq:
+	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
+	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)
+		mthca_free_srq_buf(dev, srq);
+
+err_out_mailbox:
+	mthca_free_mailbox(dev, mailbox);
+
+err_out_db:
+	if (!pd->ibpd.uobject && mthca_is_memfree(dev))
+		mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
+
+err_out_icm:
+	mthca_table_put(dev, dev->srq_table.table, srq->srqn);
+
+err_out:
+	mthca_free(&dev->srq_table.alloc, srq->srqn);
+
+	return err;
+}
+
+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)) {
+		mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
+		return;
+	}
+
+	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
+	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));
+	spin_unlock_irq(&dev->srq_table.lock);
+
+	atomic_dec(&srq->refcount);
+	wait_event(srq->wait, !atomic_read(&srq->refcount));
+
+	if (!srq->ibsrq.uobject) {
+		mthca_free_srq_buf(dev, srq);
+		if (mthca_is_memfree(dev))
+			mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
+	}
+
+	mthca_table_put(dev, dev->srq_table.table, srq->srqn);
+	mthca_free(&dev->srq_table.alloc, srq->srqn);
+	mthca_free_mailbox(dev, mailbox);
+}
+
+void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
+		     enum ib_event_type event_type)
+{
+	struct mthca_srq *srq;
+	struct ib_event event;
+
+	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);
+	spin_unlock(&dev->srq_table.lock);
+
+	if (!srq) {
+		mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
+		return;
+	}
+
+	if (!srq->ibsrq.event_handler)
+		goto out;
+
+	event.device      = &dev->ib_dev;
+	event.event       = event_type;
+	event.element.srq  = &srq->ibsrq;
+	srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
+
+out:
+	if (atomic_dec_and_test(&srq->refcount))
+		wake_up(&srq->wait);
+}
+
+/*
+ * This function must be called with IRQs disabled.
+ */
+void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
+{
+	int ind;
+
+	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;
+
+	*wqe_to_link(get_wqe(srq, ind)) = -1;
+	srq->last_free = ind;
+
+	spin_unlock(&srq->lock);
+}
+
+int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
+			      struct ib_recv_wr **bad_wr)
+{
+	struct mthca_dev *dev = to_mdev(ibsrq->device);
+	struct mthca_srq *srq = to_msrq(ibsrq);
+	unsigned long flags;
+	int err = 0;
+	int first_ind;
+	int ind;
+	int next_ind;
+	int nreq;
+	int i;
+	void *wqe;
+	void *prev_wqe;
+
+	spin_lock_irqsave(&srq->lock, flags);
+
+	first_ind = srq->first_free;
+
+	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;
+			return nreq;
+		}
+
+		wqe       = get_wqe(srq, ind);
+		next_ind  = *wqe_to_link(wqe);
+		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 */
+
+		wqe += sizeof (struct mthca_next_seg);
+
+		if (unlikely(wr->num_sge > srq->max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			srq->last = prev_wqe;
+			return nreq;
+		}
+
+		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);
+			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 (likely(prev_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;
+	}
+
+	return nreq;
+
+	if (likely(nreq)) {
+		__be32 doorbell[2];
+
+		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,
+			      dev->kar + MTHCA_RECEIVE_DOORBELL,
+			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+	}
+
+	spin_unlock_irqrestore(&srq->lock, flags);
+	return err;
+}
+
+int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
+			      struct ib_recv_wr **bad_wr)
+{
+	struct mthca_dev *dev = to_mdev(ibsrq->device);
+	struct mthca_srq *srq = to_msrq(ibsrq);
+	unsigned long flags;
+	int err = 0;
+	int ind;
+	int next_ind;
+	int nreq;
+	int i;
+	void *wqe;
+
+	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;
+			return nreq;
+		}
+
+		wqe       = get_wqe(srq, ind);
+		next_ind  = *wqe_to_link(wqe);
+
+		((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 */
+
+		wqe += sizeof (struct mthca_next_seg);
+
+		if (unlikely(wr->num_sge > srq->max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			return nreq;
+		}
+
+		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);
+			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;
+		}
+
+		srq->wrid[ind]  = wr->wr_id;
+		srq->first_free = next_ind;
+	}
+
+	if (likely(nreq)) {
+		srq->counter += nreq;
+
+		/*
+		 * Make sure that descriptors are written before
+		 * we write doorbell record.
+		 */
+		wmb();
+		*srq->db = cpu_to_be32(srq->counter);
+	}
+
+	spin_unlock_irqrestore(&srq->lock, flags);
+	return err;
+}
+
+int __devinit mthca_init_srq_table(struct mthca_dev *dev)
+{
+	int err;
+
+	if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
+		return 0;
+
+	spin_lock_init(&dev->srq_table.lock);
+
+	err = mthca_alloc_init(&dev->srq_table.alloc,
+			       dev->limits.num_srqs,
+			       dev->limits.num_srqs - 1,
+			       dev->limits.reserved_srqs);
+	if (err)
+		return err;
+
+	err = mthca_array_init(&dev->srq_table.srq,
+			       dev->limits.num_srqs);
+	if (err)
+		mthca_alloc_cleanup(&dev->srq_table.alloc);
+
+	return err;
+}
+
+void __devexit mthca_cleanup_srq_table(struct mthca_dev *dev)
+{
+	if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
+		return;
+
+	mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
+	mthca_alloc_cleanup(&dev->srq_table.alloc);
+}

drivers/infiniband/hw/mthca/mthca_user.h

···
 	__u32 reserved;
 };
 
+struct mthca_create_srq {
+	__u32 lkey;
+	__u32 db_index;
+	__u64 db_page;
+};
+
+struct mthca_create_srq_resp {
+	__u32 srqn;
+	__u32 reserved;
+};
+
 struct mthca_create_qp {
 	__u32 lkey;
 	__u32 reserved;