[PATCH] IB/mthca: Add SRQ implementation
authorRoland Dreier <roland@eddore.topspincom.com>
Fri, 19 Aug 2005 17:59:31 +0000 (10:59 -0700)
committerRoland Dreier <rolandd@cisco.com>
Sat, 27 Aug 2005 03:37:37 +0000 (20:37 -0700)
Add mthca support for shared receive queues (SRQs),
including userspace SRQs.

Signed-off-by: Roland Dreier <rolandd@cisco.com>
13 files changed:
drivers/infiniband/hw/mthca/Makefile
drivers/infiniband/hw/mthca/mthca_cmd.c
drivers/infiniband/hw/mthca/mthca_cmd.h
drivers/infiniband/hw/mthca/mthca_cq.c
drivers/infiniband/hw/mthca/mthca_dev.h
drivers/infiniband/hw/mthca/mthca_main.c
drivers/infiniband/hw/mthca/mthca_profile.c
drivers/infiniband/hw/mthca/mthca_profile.h
drivers/infiniband/hw/mthca/mthca_provider.c
drivers/infiniband/hw/mthca/mthca_provider.h
drivers/infiniband/hw/mthca/mthca_qp.c
drivers/infiniband/hw/mthca/mthca_srq.c [new file with mode: 0644]
drivers/infiniband/hw/mthca/mthca_user.h

index 5dcbd43073e2454d3944b3fddf1006b325c0d0b5..1eb87408e069e2e824536b4f059ebe0a40cfa2a8 100644 (file)
@@ -9,4 +9,4 @@ obj-$(CONFIG_INFINIBAND_MTHCA) += ib_mthca.o
 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
index c258c1b7022e37af90b2fc06638ea889b0705949..60e4b213635af30091b05115a97866e073881259 100644 (file)
@@ -109,6 +109,7 @@ enum {
        CMD_SW2HW_SRQ       = 0x35,
        CMD_HW2SW_SRQ       = 0x36,
        CMD_QUERY_SRQ       = 0x37,
+       CMD_ARM_SRQ         = 0x40,
 
        /* QP/EE commands */
        CMD_RST2INIT_QPEE   = 0x19,
@@ -1032,6 +1033,8 @@ int mthca_QUERY_DEV_LIM(struct mthca_dev *dev,
 
        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",
@@ -1500,6 +1503,27 @@ int mthca_HW2SW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
                             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,
                    int is_ee, struct mthca_mailbox *mailbox, u32 optmask,
                    u8 *status)
index 11f02a61432af026238c59cce0542e9d9aee7855..ef2a765d69539771a2ec10218403d5d5ac844ba9 100644 (file)
@@ -298,6 +298,11 @@ int mthca_SW2HW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
                   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);
index 5dee908c2f34fac5ce802c5ec3030db2687bcc72..5ece609c2ee0398678c66a3968ad399cfcdefe83 100644 (file)
@@ -224,7 +224,8 @@ void mthca_cq_event(struct mthca_dev *dev, u32 cqn)
        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;
@@ -265,8 +266,11 @@ void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn)
         */
        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),
@@ -455,23 +459,27 @@ static inline int mthca_poll_one(struct mthca_dev *dev,
                             >> 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;
-
-       wq->last_comp = wqe_index;
+       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;
 
-       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,
index cb78b5d07201893a69a1def238e486a0b29589be..7bff5a8425f4e5c5fbd7eb634a78b13eb8a3b88f 100644 (file)
@@ -218,6 +218,13 @@ struct mthca_cq_table {
        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;
@@ -299,6 +306,7 @@ struct mthca_dev {
        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;
@@ -372,6 +380,7 @@ int mthca_init_pd_table(struct mthca_dev *dev);
 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);
@@ -381,6 +390,7 @@ void mthca_cleanup_pd_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);
@@ -431,7 +441,19 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 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);
index 16c5d4a805f0a367f9414517faa42f04f6d9f8e2..3241d6c9dc11d2f2a959224f22cd58070449d0b5 100644 (file)
@@ -253,6 +253,8 @@ static int __devinit mthca_init_tavor(struct mthca_dev *mdev)
        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)
@@ -424,15 +426,29 @@ static int __devinit mthca_init_icm(struct mthca_dev *mdev,
        }
 
        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;
+               }
+       }
+
        /*
         * It's not strictly required, but for simplicity just map the
         * whole multicast group table now.  The table isn't very big
@@ -448,11 +464,15 @@ static int __devinit mthca_init_icm(struct mthca_dev *mdev,
        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);
 
@@ -532,6 +552,8 @@ static int __devinit mthca_init_arbel(struct mthca_dev *mdev)
        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) {
@@ -558,6 +580,8 @@ static int __devinit mthca_init_arbel(struct mthca_dev *mdev)
        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);
@@ -587,6 +611,8 @@ static void mthca_close_hca(struct mthca_dev *mdev)
        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);
@@ -731,11 +757,18 @@ static int __devinit mthca_setup_hca(struct mthca_dev *dev)
                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);
@@ -760,6 +793,9 @@ err_av_table_free:
 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);
 
@@ -1046,6 +1082,7 @@ err_cleanup:
        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);
@@ -1095,6 +1132,7 @@ static void __devexit mthca_remove_one(struct pci_dev *pdev)
                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);
index 9b280661f2a1304e2d8959359bd689fc2cd4c8d5..0576056b34f463742b7979ac9791b5bde002d8d0 100644 (file)
@@ -102,6 +102,7 @@ u64 mthca_make_profile(struct mthca_dev *dev,
        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;
index 0d4f070a3fa1a4fe2ce8df09c3e0f35810b2a5e9..94641808f97f5cff4ef7cf93dd323d1ab4a5f584 100644 (file)
@@ -42,6 +42,7 @@
 struct mthca_profile {
        int num_qp;
        int rdb_per_qp;
+       int num_srq;
        int num_cq;
        int num_mcg;
        int num_mpt;
index 08a7340e19ffdca410aa5fd50a38e8f1df756f0b..23ceb26af8fe5189fc82c4a4c423b98bc4a4913f 100644 (file)
@@ -425,6 +425,77 @@ static int mthca_ah_destroy(struct ib_ah *ah)
        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)
@@ -1003,6 +1074,17 @@ int mthca_register_device(struct mthca_dev *dev)
        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;
index b95249ee46cf2da1959acef7e9b520ce309b57d7..024015678c8a6ff1590b857c18f167187a584330 100644 (file)
@@ -197,6 +197,29 @@ struct mthca_cq {
        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;
@@ -277,6 +300,11 @@ 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)
 {
        return container_of(ibqp, struct mthca_qp, ibqp);
index ebb8f4a3dd80c627d9d9b2e6ea166d5b1a066aa3..7607b980073628debc86979b53c807c60cbc1c5d 100644 (file)
@@ -612,10 +612,13 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
                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 */
@@ -784,6 +787,9 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 
        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);
@@ -806,6 +812,10 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
                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);
        if (status) {
@@ -1260,9 +1270,11 @@ void mthca_free_qp(struct mthca_dev *dev,
         * 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);
@@ -2008,6 +2020,15 @@ int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
 {
        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);
        else
diff --git a/drivers/infiniband/hw/mthca/mthca_srq.c b/drivers/infiniband/hw/mthca/mthca_srq.c
new file mode 100644 (file)
index 0000000..75cd2d8
--- /dev/null
@@ -0,0 +1,591 @@
+/*
+ * 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);
+}
index 3024c1b4547d790e3f9cf0bca4c87bf01c736fb6..41613ec8a04e63c92fdfdbfe5d82e897c63c6a8b 100644 (file)
@@ -69,6 +69,17 @@ struct mthca_create_cq_resp {
        __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;