/*
* Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
* Copyright (c) 2005 Mellanox Technologies. All rights reserved.
- * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ * Copyright (c) 2005, 2006 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
CMD_TIME_CLASS_A, status);
}
+int mthca_RESIZE_CQ(struct mthca_dev *dev, int cq_num, u32 lkey, u8 log_size,
+ u8 *status)
+{
+ struct mthca_mailbox *mailbox;
+ __be32 *inbox;
+ int err;
+
+#define RESIZE_CQ_IN_SIZE 0x40
+#define RESIZE_CQ_LOG_SIZE_OFFSET 0x0c
+#define RESIZE_CQ_LKEY_OFFSET 0x1c
+
+ mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+ if (IS_ERR(mailbox))
+ return PTR_ERR(mailbox);
+ inbox = mailbox->buf;
+
+ memset(inbox, 0, RESIZE_CQ_IN_SIZE);
+ /*
+ * Leave start address fields zeroed out -- mthca assumes that
+ * MRs for CQs always start at virtual address 0.
+ */
+ MTHCA_PUT(inbox, log_size, RESIZE_CQ_LOG_SIZE_OFFSET);
+ MTHCA_PUT(inbox, lkey, RESIZE_CQ_LKEY_OFFSET);
+
+ err = mthca_cmd(dev, mailbox->dma, cq_num, 1, CMD_RESIZE_CQ,
+ CMD_TIME_CLASS_B, status);
+
+ mthca_free_mailbox(dev, mailbox);
+ return err;
+}
+
int mthca_SW2HW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
int srq_num, u8 *status)
{
/*
* Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
* Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2006 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
int cq_num, u8 *status);
int mthca_HW2SW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
int cq_num, u8 *status);
+int mthca_RESIZE_CQ(struct mthca_dev *dev, int cq_num, u32 lkey, u8 log_size,
+ 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,
/*
* Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
* Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
- * Copyright (c) 2005 Cisco Systems, Inc. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2005 Mellanox Technologies. All rights reserved.
* Copyright (c) 2004 Voltaire, Inc. All rights reserved.
*
#define MTHCA_ARBEL_CQ_DB_REQ_NOT (2 << 24)
#define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
-static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
+static inline struct mthca_cqe *get_cqe_from_buf(struct mthca_cq_buf *buf,
+ int entry)
{
- if (cq->is_direct)
- return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
+ if (buf->is_direct)
+ return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
else
- return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
+ return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
}
-static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
+static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
+{
+ return get_cqe_from_buf(&cq->buf, entry);
+}
+
+static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
{
- struct mthca_cqe *cqe = get_cqe(cq, i);
return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
}
static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
{
- return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
+ return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
}
static inline void set_cqe_hw(struct mthca_cqe *cqe)
* from our QP and therefore don't need to be checked.
*/
for (prod_index = cq->cons_index;
- cqe_sw(cq, prod_index & cq->ibcq.cqe);
+ cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
++prod_index)
if (prod_index == cq->cons_index + cq->ibcq.cqe)
break;
wake_up(&cq->wait);
}
+void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
+{
+ int i;
+
+ /*
+ * In Tavor mode, the hardware keeps the consumer and producer
+ * indices mod the CQ size. Since we might be making the CQ
+ * bigger, we need to deal with the case where the producer
+ * index wrapped around before the CQ was resized.
+ */
+ if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
+ cq->ibcq.cqe < cq->resize_buf->cqe) {
+ cq->cons_index &= cq->ibcq.cqe;
+ if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
+ cq->cons_index -= cq->ibcq.cqe + 1;
+ }
+
+ for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
+ memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
+ i & cq->resize_buf->cqe),
+ get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
+}
+
+int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
+{
+ int ret;
+ int i;
+
+ ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
+ MTHCA_MAX_DIRECT_CQ_SIZE,
+ &buf->queue, &buf->is_direct,
+ &dev->driver_pd, 1, &buf->mr);
+ if (ret)
+ return ret;
+
+ for (i = 0; i < nent; ++i)
+ set_cqe_hw(get_cqe_from_buf(buf, i));
+
+ return 0;
+}
+
+void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
+{
+ mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
+ buf->is_direct, &buf->mr);
+}
+
static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
struct mthca_qp *qp, int wqe_index, int is_send,
struct mthca_err_cqe *cqe,
spin_lock_irqsave(&cq->lock, flags);
- for (npolled = 0; npolled < num_entries; ++npolled) {
+ npolled = 0;
+repoll:
+ while (npolled < num_entries) {
err = mthca_poll_one(dev, cq, &qp,
&freed, entry + npolled);
if (err)
break;
+ ++npolled;
}
if (freed) {
update_cons_index(dev, cq, freed);
}
+ /*
+ * If a CQ resize is in progress and we discovered that the
+ * old buffer is empty, then peek in the new buffer, and if
+ * it's not empty, switch to the new buffer and continue
+ * polling there.
+ */
+ if (unlikely(err == -EAGAIN && cq->resize_buf &&
+ cq->resize_buf->state == CQ_RESIZE_READY)) {
+ /*
+ * In Tavor mode, the hardware keeps the producer
+ * index modulo the CQ size. Since we might be making
+ * the CQ bigger, we need to mask our consumer index
+ * using the size of the old CQ buffer before looking
+ * in the new CQ buffer.
+ */
+ if (!mthca_is_memfree(dev))
+ cq->cons_index &= cq->ibcq.cqe;
+
+ if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
+ cq->cons_index & cq->resize_buf->cqe))) {
+ struct mthca_cq_buf tbuf;
+ int tcqe;
+
+ tbuf = cq->buf;
+ tcqe = cq->ibcq.cqe;
+ cq->buf = cq->resize_buf->buf;
+ cq->ibcq.cqe = cq->resize_buf->cqe;
+
+ cq->resize_buf->buf = tbuf;
+ cq->resize_buf->cqe = tcqe;
+ cq->resize_buf->state = CQ_RESIZE_SWAPPED;
+
+ goto repoll;
+ }
+ }
+
spin_unlock_irqrestore(&cq->lock, flags);
return err == 0 || err == -EAGAIN ? npolled : err;
return 0;
}
-static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
-{
- mthca_buf_free(dev, (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
- &cq->queue, cq->is_direct, &cq->mr);
-}
-
int mthca_init_cq(struct mthca_dev *dev, int nent,
struct mthca_ucontext *ctx, u32 pdn,
struct mthca_cq *cq)
{
- int size = nent * MTHCA_CQ_ENTRY_SIZE;
struct mthca_mailbox *mailbox;
struct mthca_cq_context *cq_context;
int err = -ENOMEM;
u8 status;
- int i;
cq->ibcq.cqe = nent - 1;
cq->is_kernel = !ctx;
cq_context = mailbox->buf;
if (cq->is_kernel) {
- err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_CQ_SIZE,
- &cq->queue, &cq->is_direct,
- &dev->driver_pd, 1, &cq->mr);
+ err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
if (err)
goto err_out_mailbox;
-
- for (i = 0; i < nent; ++i)
- set_cqe_hw(get_cqe(cq, i));
}
spin_lock_init(&cq->lock);
cq_context->error_eqn = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
cq_context->comp_eqn = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
cq_context->pd = cpu_to_be32(pdn);
- cq_context->lkey = cpu_to_be32(cq->mr.ibmr.lkey);
+ cq_context->lkey = cpu_to_be32(cq->buf.mr.ibmr.lkey);
cq_context->cqn = cpu_to_be32(cq->cqn);
if (mthca_is_memfree(dev)) {
err_out_free_mr:
if (cq->is_kernel)
- mthca_free_cq_buf(dev, cq);
+ mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
err_out_mailbox:
mthca_free_mailbox(dev, mailbox);
wait_event(cq->wait, !atomic_read(&cq->refcount));
if (cq->is_kernel) {
- mthca_free_cq_buf(dev, cq);
+ mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
if (mthca_is_memfree(dev)) {
mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
/*
* Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
* Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
- * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
* Copyright (c) 2005 Mellanox Technologies. All rights reserved.
* Copyright (c) 2004 Voltaire, Inc. All rights reserved.
*
enum ib_event_type event_type);
void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
struct mthca_srq *srq);
+void mthca_cq_resize_copy_cqes(struct mthca_cq *cq);
+int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent);
+void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe);
int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
struct ib_srq_attr *attr, struct mthca_srq *srq);
/*
* Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
* Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
- * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
* Copyright (c) 2005 Mellanox Technologies. All rights reserved.
* Copyright (c) 2004 Voltaire, Inc. All rights reserved.
*
}
if (context) {
- cq->mr.ibmr.lkey = ucmd.lkey;
- cq->set_ci_db_index = ucmd.set_db_index;
- cq->arm_db_index = ucmd.arm_db_index;
+ cq->buf.mr.ibmr.lkey = ucmd.lkey;
+ cq->set_ci_db_index = ucmd.set_db_index;
+ cq->arm_db_index = ucmd.arm_db_index;
}
for (nent = 1; nent <= entries; nent <<= 1)
goto err_free;
}
+ cq->resize_buf = NULL;
+
return &cq->ibcq;
err_free:
return ERR_PTR(err);
}
+static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
+ int entries)
+{
+ int ret;
+
+ spin_lock_irq(&cq->lock);
+ if (cq->resize_buf) {
+ ret = -EBUSY;
+ goto unlock;
+ }
+
+ cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
+ if (!cq->resize_buf) {
+ ret = -ENOMEM;
+ goto unlock;
+ }
+
+ cq->resize_buf->state = CQ_RESIZE_ALLOC;
+
+ ret = 0;
+
+unlock:
+ spin_unlock_irq(&cq->lock);
+
+ if (ret)
+ return ret;
+
+ ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
+ if (ret) {
+ spin_lock_irq(&cq->lock);
+ kfree(cq->resize_buf);
+ cq->resize_buf = NULL;
+ spin_unlock_irq(&cq->lock);
+ return ret;
+ }
+
+ cq->resize_buf->cqe = entries - 1;
+
+ spin_lock_irq(&cq->lock);
+ cq->resize_buf->state = CQ_RESIZE_READY;
+ spin_unlock_irq(&cq->lock);
+
+ return 0;
+}
+
+static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
+{
+ struct mthca_dev *dev = to_mdev(ibcq->device);
+ struct mthca_cq *cq = to_mcq(ibcq);
+ struct mthca_resize_cq ucmd;
+ u32 lkey;
+ u8 status;
+ int ret;
+
+ if (entries < 1 || entries > dev->limits.max_cqes)
+ return -EINVAL;
+
+ entries = roundup_pow_of_two(entries + 1);
+ if (entries == ibcq->cqe + 1)
+ return 0;
+
+ if (cq->is_kernel) {
+ ret = mthca_alloc_resize_buf(dev, cq, entries);
+ if (ret)
+ return ret;
+ lkey = cq->resize_buf->buf.mr.ibmr.lkey;
+ } else {
+ if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
+ return -EFAULT;
+ lkey = ucmd.lkey;
+ }
+
+ ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, long_log2(entries), &status);
+ if (status)
+ ret = -EINVAL;
+
+ if (ret) {
+ if (cq->resize_buf) {
+ mthca_free_cq_buf(dev, &cq->resize_buf->buf,
+ cq->resize_buf->cqe);
+ kfree(cq->resize_buf);
+ spin_lock_irq(&cq->lock);
+ cq->resize_buf = NULL;
+ spin_unlock_irq(&cq->lock);
+ }
+ return ret;
+ }
+
+ if (cq->is_kernel) {
+ struct mthca_cq_buf tbuf;
+ int tcqe;
+
+ spin_lock_irq(&cq->lock);
+ if (cq->resize_buf->state == CQ_RESIZE_READY) {
+ mthca_cq_resize_copy_cqes(cq);
+ tbuf = cq->buf;
+ tcqe = cq->ibcq.cqe;
+ cq->buf = cq->resize_buf->buf;
+ cq->ibcq.cqe = cq->resize_buf->cqe;
+ } else {
+ tbuf = cq->resize_buf->buf;
+ tcqe = cq->resize_buf->cqe;
+ }
+
+ kfree(cq->resize_buf);
+ cq->resize_buf = NULL;
+ spin_unlock_irq(&cq->lock);
+
+ mthca_free_cq_buf(dev, &tbuf, tcqe);
+ } else
+ ibcq->cqe = entries - 1;
+
+ return 0;
+}
+
static int mthca_destroy_cq(struct ib_cq *cq)
{
if (cq->uobject) {
(1ull << IB_USER_VERBS_CMD_DEREG_MR) |
(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
(1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
+ (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
(1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
(1ull << IB_USER_VERBS_CMD_CREATE_QP) |
(1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
dev->ib_dev.modify_qp = mthca_modify_qp;
dev->ib_dev.destroy_qp = mthca_destroy_qp;
dev->ib_dev.create_cq = mthca_create_cq;
+ dev->ib_dev.resize_cq = mthca_resize_cq;
dev->ib_dev.destroy_cq = mthca_destroy_cq;
dev->ib_dev.poll_cq = mthca_poll_cq;
dev->ib_dev.get_dma_mr = mthca_get_dma_mr;
/*
* Copyright (c) 2004 Topspin Communications. All rights reserved.
- * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
* Copyright (c) 2005 Mellanox Technologies. All rights reserved.
*
* This software is available to you under a choice of one of two
* - wait_event until ref count is zero
*
* It is the consumer's responsibilty to make sure that no QP
- * operations (WQE posting or state modification) are pending when the
+ * operations (WQE posting or state modification) are pending when a
* QP is destroyed. Also, the consumer must make sure that calls to
- * qp_modify are serialized.
+ * qp_modify are serialized. Similarly, the consumer is responsible
+ * for ensuring that no CQ resize operations are pending when a CQ
+ * is destroyed.
*
* Possible optimizations (wait for profile data to see if/where we
* have locks bouncing between CPUs):
* send queue and one for the receive queue)
*/
+struct mthca_cq_buf {
+ union mthca_buf queue;
+ struct mthca_mr mr;
+ int is_direct;
+};
+
+struct mthca_cq_resize {
+ struct mthca_cq_buf buf;
+ int cqe;
+ enum {
+ CQ_RESIZE_ALLOC,
+ CQ_RESIZE_READY,
+ CQ_RESIZE_SWAPPED
+ } state;
+};
+
struct mthca_cq {
- struct ib_cq ibcq;
- spinlock_t lock;
- atomic_t refcount;
- int cqn;
- u32 cons_index;
- int is_direct;
- int is_kernel;
+ struct ib_cq ibcq;
+ spinlock_t lock;
+ atomic_t refcount;
+ int cqn;
+ u32 cons_index;
+ struct mthca_cq_buf buf;
+ struct mthca_cq_resize *resize_buf;
+ int is_kernel;
/* Next fields are Arbel only */
- int set_ci_db_index;
- __be32 *set_ci_db;
- int arm_db_index;
- __be32 *arm_db;
- int arm_sn;
+ int set_ci_db_index;
+ __be32 *set_ci_db;
+ int arm_db_index;
+ __be32 *arm_db;
+ int arm_sn;
- union mthca_buf queue;
- struct mthca_mr mr;
- wait_queue_head_t wait;
+ wait_queue_head_t wait;
};
struct mthca_srq {
/*
* Copyright (c) 2005 Topspin Communications. All rights reserved.
- * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ * Copyright (c) 2005, 2006 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
__u32 reserved;
};
+struct mthca_resize_cq {
+ __u32 lkey;
+ __u32 reserved;
+};
+
struct mthca_create_srq {
__u32 lkey;
__u32 db_index;