[ofa-general] Re: [PATCH 1/1 v1] MLX4: Added resize_cq capability.
Roland Dreier
rdreier at cisco.com
Fri Apr 4 14:21:30 PDT 2008
Thanks, I applied this with a lot of changes. Some comments:
> entries = roundup_pow_of_two(entries + 1);
your patch was corrupted in a very strange way... the context lines had
two spaces instead of one at the beginning. I just deleted the extra
space by hand.
> + err = mlx4_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
> + if (err) {
> + spin_lock_irq(&cq->lock);
> + kfree(cq->resize_buf);
> + cq->resize_buf = NULL;
> + spin_unlock_irq(&cq->lock);
> + goto out;
> + }
> +err_buf:
> + if (cq->resize_buf) {
> + if (!ibcq->uobject)
> + mlx4_free_cq_buf(dev, &cq->resize_buf->buf,
> + cq->resize_buf->cqe);
> +
> + spin_lock_irq(&cq->lock);
> + kfree(cq->resize_buf);
> + cq->resize_buf = NULL;
> + spin_unlock_irq(&cq->lock);
> + }
Why do we need the spinlock in these places? There's no way for this to
race with mlx4_ib_poll_one() is there, since that should never see the
RESIZE CQE? (If there is such a race, then we're in trouble even with
the lock, since we're aborting the resize, and the poll code shouldn't
swap the buffers)
Also I got rid of the duplicated code to allocate buffers and get
userspace buffers, so that the allocate and resize paths use the same
code. And I cleaned up some other stuff.
So please review/test my work to make sure I didn't break your patch...
---
drivers/infiniband/hw/mlx4/cq.c | 292 ++++++++++++++++++++++++++++++----
drivers/infiniband/hw/mlx4/main.c | 2 +
drivers/infiniband/hw/mlx4/mlx4_ib.h | 9 +
drivers/net/mlx4/cq.c | 28 ++++
include/linux/mlx4/cq.h | 2 +
5 files changed, 300 insertions(+), 33 deletions(-)
diff --git a/drivers/infiniband/hw/mlx4/cq.c b/drivers/infiniband/hw/mlx4/cq.c
index e4fb64b..3557e7e 100644
--- a/drivers/infiniband/hw/mlx4/cq.c
+++ b/drivers/infiniband/hw/mlx4/cq.c
@@ -93,6 +93,74 @@ int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
}
+static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
+{
+ int err;
+
+ err = mlx4_buf_alloc(dev->dev, nent * sizeof(struct mlx4_cqe),
+ PAGE_SIZE * 2, &buf->buf);
+
+ if (err)
+ goto out;
+
+ err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
+ &buf->mtt);
+ if (err)
+ goto err_buf;
+
+ err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
+ if (err)
+ goto err_mtt;
+
+ return 0;
+
+err_mtt:
+ mlx4_mtt_cleanup(dev->dev, &buf->mtt);
+
+err_buf:
+ mlx4_buf_free(dev->dev, nent * sizeof(struct mlx4_cqe),
+ &buf->buf);
+
+out:
+ return err;
+}
+
+static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
+{
+ mlx4_buf_free(dev->dev, (cqe + 1) * sizeof(struct mlx4_cqe), &buf->buf);
+}
+
+static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
+ struct mlx4_ib_cq_buf *buf, struct ib_umem **umem,
+ u64 buf_addr, int cqe)
+{
+ int err;
+
+ *umem = ib_umem_get(context, buf_addr, cqe * sizeof (struct mlx4_cqe),
+ IB_ACCESS_LOCAL_WRITE);
+ if (IS_ERR(*umem))
+ return PTR_ERR(*umem);
+
+ err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem),
+ ilog2((*umem)->page_size), &buf->mtt);
+ if (err)
+ goto err_buf;
+
+ err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
+ if (err)
+ goto err_mtt;
+
+ return 0;
+
+err_mtt:
+ mlx4_mtt_cleanup(dev->dev, &buf->mtt);
+
+err_buf:
+ ib_umem_release(*umem);
+
+ return err;
+}
+
struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector,
struct ib_ucontext *context,
struct ib_udata *udata)
@@ -100,7 +168,6 @@ struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector
struct mlx4_ib_dev *dev = to_mdev(ibdev);
struct mlx4_ib_cq *cq;
struct mlx4_uar *uar;
- int buf_size;
int err;
if (entries < 1 || entries > dev->dev->caps.max_cqes)
@@ -112,8 +179,10 @@ struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector
entries = roundup_pow_of_two(entries + 1);
cq->ibcq.cqe = entries - 1;
- buf_size = entries * sizeof (struct mlx4_cqe);
+ mutex_init(&cq->resize_mutex);
spin_lock_init(&cq->lock);
+ cq->resize_buf = NULL;
+ cq->resize_umem = NULL;
if (context) {
struct mlx4_ib_create_cq ucmd;
@@ -123,21 +192,10 @@ struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector
goto err_cq;
}
- cq->umem = ib_umem_get(context, ucmd.buf_addr, buf_size,
- IB_ACCESS_LOCAL_WRITE);
- if (IS_ERR(cq->umem)) {
- err = PTR_ERR(cq->umem);
- goto err_cq;
- }
-
- err = mlx4_mtt_init(dev->dev, ib_umem_page_count(cq->umem),
- ilog2(cq->umem->page_size), &cq->buf.mtt);
+ err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
+ ucmd.buf_addr, entries);
if (err)
- goto err_buf;
-
- err = mlx4_ib_umem_write_mtt(dev, &cq->buf.mtt, cq->umem);
- if (err)
- goto err_mtt;
+ goto err_cq;
err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
&cq->db);
@@ -155,19 +213,9 @@ struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector
*cq->mcq.set_ci_db = 0;
*cq->mcq.arm_db = 0;
- if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &cq->buf.buf)) {
- err = -ENOMEM;
- goto err_db;
- }
-
- err = mlx4_mtt_init(dev->dev, cq->buf.buf.npages, cq->buf.buf.page_shift,
- &cq->buf.mtt);
+ err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
if (err)
- goto err_buf;
-
- err = mlx4_buf_write_mtt(dev->dev, &cq->buf.mtt, &cq->buf.buf);
- if (err)
- goto err_mtt;
+ goto err_db;
uar = &dev->priv_uar;
}
@@ -195,12 +243,10 @@ err_dbmap:
err_mtt:
mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
-err_buf:
if (context)
ib_umem_release(cq->umem);
else
- mlx4_buf_free(dev->dev, entries * sizeof (struct mlx4_cqe),
- &cq->buf.buf);
+ mlx4_ib_free_cq_buf(dev, &cq->buf, entries);
err_db:
if (!context)
@@ -212,6 +258,170 @@ err_cq:
return ERR_PTR(err);
}
+static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
+ int entries)
+{
+ int err;
+
+ if (cq->resize_buf)
+ return -EBUSY;
+
+ cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
+ if (!cq->resize_buf)
+ return -ENOMEM;
+
+ err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
+ if (err) {
+ kfree(cq->resize_buf);
+ cq->resize_buf = NULL;
+ return err;
+ }
+
+ cq->resize_buf->cqe = entries - 1;
+
+ return 0;
+}
+
+static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
+ int entries, struct ib_udata *udata)
+{
+ struct mlx4_ib_resize_cq ucmd;
+ int err;
+
+ if (cq->resize_umem)
+ return -EBUSY;
+
+ if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
+ return -EFAULT;
+
+ cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
+ if (!cq->resize_buf)
+ return -ENOMEM;
+
+ err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
+ &cq->resize_umem, ucmd.buf_addr, entries);
+ if (err) {
+ kfree(cq->resize_buf);
+ cq->resize_buf = NULL;
+ return err;
+ }
+
+ cq->resize_buf->cqe = entries - 1;
+
+ return 0;
+}
+
+static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
+{
+ u32 i;
+
+ i = cq->mcq.cons_index;
+ while (get_sw_cqe(cq, i & cq->ibcq.cqe))
+ ++i;
+
+ return i - cq->mcq.cons_index;
+}
+
+static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
+{
+ struct mlx4_cqe *cqe;
+ int i;
+
+ i = cq->mcq.cons_index;
+ cqe = get_cqe(cq, i & cq->ibcq.cqe);
+ while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
+ memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
+ (i + 1) & cq->resize_buf->cqe),
+ get_cqe(cq, i & cq->ibcq.cqe), sizeof(struct mlx4_cqe));
+ cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
+ }
+ ++cq->mcq.cons_index;
+}
+
+int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
+{
+ struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
+ struct mlx4_ib_cq *cq = to_mcq(ibcq);
+ int outst_cqe;
+ int err;
+
+ mutex_lock(&cq->resize_mutex);
+
+ if (entries < 1 || entries > dev->dev->caps.max_cqes) {
+ err = -EINVAL;
+ goto out;
+ }
+
+ entries = roundup_pow_of_two(entries + 1);
+ if (entries == ibcq->cqe + 1) {
+ err = 0;
+ goto out;
+ }
+
+ if (ibcq->uobject) {
+ err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
+ if (err)
+ goto out;
+ } else {
+ /* Can't be smaller then the number of outstanding CQEs */
+ outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
+ if (entries < outst_cqe + 1) {
+ err = 0;
+ goto out;
+ }
+
+ err = mlx4_alloc_resize_buf(dev, cq, entries);
+ if (err)
+ goto out;
+ }
+
+ err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
+ if (err)
+ goto err_buf;
+
+ if (ibcq->uobject) {
+ cq->buf = cq->resize_buf->buf;
+ cq->ibcq.cqe = cq->resize_buf->cqe;
+ ib_umem_release(cq->umem);
+ cq->umem = cq->resize_umem;
+
+ kfree(cq->resize_buf);
+ cq->resize_buf = NULL;
+ cq->resize_umem = NULL;
+ } else {
+ spin_lock_irq(&cq->lock);
+ if (cq->resize_buf) {
+ mlx4_ib_cq_resize_copy_cqes(cq);
+ mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
+ cq->buf = cq->resize_buf->buf;
+ cq->ibcq.cqe = cq->resize_buf->cqe;
+
+ kfree(cq->resize_buf);
+ cq->resize_buf = NULL;
+ }
+ spin_unlock_irq(&cq->lock);
+ }
+
+ goto out;
+
+err_buf:
+ if (!ibcq->uobject)
+ mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
+ cq->resize_buf->cqe);
+
+ kfree(cq->resize_buf);
+ cq->resize_buf = NULL;
+
+ if (cq->resize_umem) {
+ ib_umem_release(cq->resize_umem);
+ cq->resize_umem = NULL;
+ }
+
+out:
+ mutex_unlock(&cq->resize_mutex);
+ return err;
+}
+
int mlx4_ib_destroy_cq(struct ib_cq *cq)
{
struct mlx4_ib_dev *dev = to_mdev(cq->device);
@@ -224,8 +434,7 @@ int mlx4_ib_destroy_cq(struct ib_cq *cq)
mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
ib_umem_release(mcq->umem);
} else {
- mlx4_buf_free(dev->dev, (cq->cqe + 1) * sizeof (struct mlx4_cqe),
- &mcq->buf.buf);
+ mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe + 1);
mlx4_ib_db_free(dev, &mcq->db);
}
@@ -332,6 +541,7 @@ static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
u32 g_mlpath_rqpn;
u16 wqe_ctr;
+repoll:
cqe = next_cqe_sw(cq);
if (!cqe)
return -EAGAIN;
@@ -354,6 +564,22 @@ static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
return -EINVAL;
}
+ /* Resize CQ in progress */
+ if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
+ if (cq->resize_buf) {
+ struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
+
+ mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
+ cq->buf = cq->resize_buf->buf;
+ cq->ibcq.cqe = cq->resize_buf->cqe;
+
+ kfree(cq->resize_buf);
+ cq->resize_buf = NULL;
+ }
+
+ goto repoll;
+ }
+
if (!*cur_qp ||
(be32_to_cpu(cqe->my_qpn) & 0xffffff) != (*cur_qp)->mqp.qpn) {
/*
diff --git a/drivers/infiniband/hw/mlx4/main.c b/drivers/infiniband/hw/mlx4/main.c
index 76dd45c..57885cd 100644
--- a/drivers/infiniband/hw/mlx4/main.c
+++ b/drivers/infiniband/hw/mlx4/main.c
@@ -571,6 +571,7 @@ static void *mlx4_ib_add(struct mlx4_dev *dev)
(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) |
@@ -610,6 +611,7 @@ static void *mlx4_ib_add(struct mlx4_dev *dev)
ibdev->ib_dev.post_recv = mlx4_ib_post_recv;
ibdev->ib_dev.create_cq = mlx4_ib_create_cq;
ibdev->ib_dev.modify_cq = mlx4_ib_modify_cq;
+ ibdev->ib_dev.resize_cq = mlx4_ib_resize_cq;
ibdev->ib_dev.destroy_cq = mlx4_ib_destroy_cq;
ibdev->ib_dev.poll_cq = mlx4_ib_poll_cq;
ibdev->ib_dev.req_notify_cq = mlx4_ib_arm_cq;
diff --git a/drivers/infiniband/hw/mlx4/mlx4_ib.h b/drivers/infiniband/hw/mlx4/mlx4_ib.h
index ef8ad96..9e63732 100644
--- a/drivers/infiniband/hw/mlx4/mlx4_ib.h
+++ b/drivers/infiniband/hw/mlx4/mlx4_ib.h
@@ -78,13 +78,21 @@ struct mlx4_ib_cq_buf {
struct mlx4_mtt mtt;
};
+struct mlx4_ib_cq_resize {
+ struct mlx4_ib_cq_buf buf;
+ int cqe;
+};
+
struct mlx4_ib_cq {
struct ib_cq ibcq;
struct mlx4_cq mcq;
struct mlx4_ib_cq_buf buf;
+ struct mlx4_ib_cq_resize *resize_buf;
struct mlx4_ib_db db;
spinlock_t lock;
+ struct mutex resize_mutex;
struct ib_umem *umem;
+ struct ib_umem *resize_umem;
};
struct mlx4_ib_mr {
@@ -255,6 +263,7 @@ struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
int mlx4_ib_dereg_mr(struct ib_mr *mr);
int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period);
+int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata);
struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector,
struct ib_ucontext *context,
struct ib_udata *udata);
diff --git a/drivers/net/mlx4/cq.c b/drivers/net/mlx4/cq.c
index 8c31434..caa5bcf 100644
--- a/drivers/net/mlx4/cq.c
+++ b/drivers/net/mlx4/cq.c
@@ -159,6 +159,34 @@ int mlx4_cq_modify(struct mlx4_dev *dev, struct mlx4_cq *cq,
}
EXPORT_SYMBOL_GPL(mlx4_cq_modify);
+int mlx4_cq_resize(struct mlx4_dev *dev, struct mlx4_cq *cq,
+ int entries, struct mlx4_mtt *mtt)
+{
+ struct mlx4_cmd_mailbox *mailbox;
+ struct mlx4_cq_context *cq_context;
+ u64 mtt_addr;
+ int err;
+
+ mailbox = mlx4_alloc_cmd_mailbox(dev);
+ if (IS_ERR(mailbox))
+ return PTR_ERR(mailbox);
+
+ cq_context = mailbox->buf;
+ memset(cq_context, 0, sizeof *cq_context);
+
+ cq_context->logsize_usrpage = cpu_to_be32(ilog2(entries) << 24);
+ cq_context->log_page_size = mtt->page_shift - 12;
+ mtt_addr = mlx4_mtt_addr(dev, mtt);
+ cq_context->mtt_base_addr_h = mtt_addr >> 32;
+ cq_context->mtt_base_addr_l = cpu_to_be32(mtt_addr & 0xffffffff);
+
+ err = mlx4_MODIFY_CQ(dev, mailbox, cq->cqn, 1);
+
+ mlx4_free_cmd_mailbox(dev, mailbox);
+ return err;
+}
+EXPORT_SYMBOL_GPL(mlx4_cq_resize);
+
int mlx4_cq_alloc(struct mlx4_dev *dev, int nent, struct mlx4_mtt *mtt,
struct mlx4_uar *uar, u64 db_rec, struct mlx4_cq *cq)
{
diff --git a/include/linux/mlx4/cq.h b/include/linux/mlx4/cq.h
index f7c3511..071cf96 100644
--- a/include/linux/mlx4/cq.h
+++ b/include/linux/mlx4/cq.h
@@ -132,5 +132,7 @@ enum {
int mlx4_cq_modify(struct mlx4_dev *dev, struct mlx4_cq *cq,
u16 count, u16 period);
+int mlx4_cq_resize(struct mlx4_dev *dev, struct mlx4_cq *cq,
+ int entries, struct mlx4_mtt *mtt);
#endif /* MLX4_CQ_H */
--
1.5.4.5
More information about the general
mailing list