On Thu, Aug 25, 2016 at 05:41:41 -0700, Yuval Shaia wrote:
> On Wed, Aug 03, 2016 at 04:27:37PM -0700, Adit Ranadive wrote:
> > This patch adds the support for creating and destroying completion
> > queues on the paravirtual RDMA device.
> >
> > Changes v2->v3:
> > - Removed boolean from pvrdma_cmd_post.
> > - Return -EAGAIN if qp retrieved from CQE is bogus.
> > - Check for invalid index of ring.
> >
> > Reviewed-by: Jorgen Hansen <jhansen@xxxxxxxxxx>
> > Reviewed-by: George Zhang <georgezhang@xxxxxxxxxx>
> > Reviewed-by: Aditya Sarwade <asarwade@xxxxxxxxxx>
> > Reviewed-by: Bryan Tan <bryantan@xxxxxxxxxx>
> > Signed-off-by: Adit Ranadive <aditr@xxxxxxxxxx>
> > ---
> > drivers/infiniband/hw/pvrdma/pvrdma_cq.c | 437
> > +++++++++++++++++++++++++++++++
> > 1 file changed, 437 insertions(+)
> > create mode 100644 drivers/infiniband/hw/pvrdma/pvrdma_cq.c
...
> > +int pvrdma_req_notify_cq(struct ib_cq *ibcq,
> > + enum ib_cq_notify_flags notify_flags) {
> > + struct pvrdma_dev *dev = to_vdev(ibcq->device);
> > + struct pvrdma_cq *cq = to_vcq(ibcq);
> > + u32 val = cq->cq_handle;
> > +
> > + val |= (notify_flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED
> ?
> > + PVRDMA_UAR_CQ_ARM_SOL : PVRDMA_UAR_CQ_ARM;
> > +
> > + writel(cpu_to_le32(val), dev->driver_uar.map +
> > +PVRDMA_UAR_CQ_OFFSET);
>
> With such a nice wrappers to dev->regs i wonder why you choose not to do
> the same for driver_uar.map.
I added those in v4.
> > +
> > + return 0;
> > +}
> > +
> > +/**
> > + * pvrdma_create_cq - create completion queue
> > + * @ibdev: the device
> > + * @attr: completion queue attributes
> > + * @context: user context
> > + * @udata: user data
> > + *
> > + * @return: ib_cq completion queue pointer on success,
> > + * otherwise returns negative errno.
> > + */
> > +struct ib_cq *pvrdma_create_cq(struct ib_device *ibdev,
> > + const struct ib_cq_init_attr *attr,
> > + struct ib_ucontext *context,
> > + struct ib_udata *udata)
> > +{
> > + int entries = attr->cqe;
> > + struct pvrdma_dev *dev = to_vdev(ibdev);
> > + struct pvrdma_cq *cq;
> > + int ret;
> > + int npages;
> > + unsigned long flags;
> > + union pvrdma_cmd_req req;
> > + union pvrdma_cmd_resp rsp;
> > + struct pvrdma_cmd_create_cq *cmd = &req.create_cq;
> > + struct pvrdma_cmd_create_cq_resp *resp = &rsp.create_cq_resp;
> > + struct pvrdma_create_cq ucmd;
> > +
> > + BUILD_BUG_ON(sizeof(struct pvrdma_cqe) != 64);
> > +
> > + entries = roundup_pow_of_two(entries);
> > + if (entries < 1 || entries > dev->dsr->caps.max_cqe)
> > + return ERR_PTR(-EINVAL);
> > +
> > + if (!atomic_add_unless(&dev->num_cqs, 1, dev->dsr-
> >caps.max_cq))
> > + return ERR_PTR(-EINVAL);
> > +
> > + cq = kzalloc(sizeof(*cq), GFP_KERNEL);
> > + if (!cq) {
> > + atomic_dec(&dev->num_cqs);
> > + return ERR_PTR(-ENOMEM);
> > + }
> > +
> > + cq->ibcq.cqe = entries;
> > +
> > + if (context) {
> > + if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
> > + ret = -EFAULT;
> > + goto err_cq;
> > + }
> > +
> > + cq->umem = ib_umem_get(context, ucmd.buf_addr,
> ucmd.buf_size,
> > + IB_ACCESS_LOCAL_WRITE, 1);
> > + if (IS_ERR(cq->umem)) {
> > + ret = PTR_ERR(cq->umem);
> > + goto err_cq;
> > + }
> > +
> > + npages = ib_umem_page_count(cq->umem);
> > + } else {
> > + cq->is_kernel = true;
> > +
> > + /* One extra page for shared ring state */
> > + npages = 1 + (entries * sizeof(struct pvrdma_cqe) +
> > + PAGE_SIZE - 1) / PAGE_SIZE;
> > +
> > + /* Skip header page. */
> > + cq->offset = PAGE_SIZE;
> > + }
> > +
> > + if (npages < 0 || npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
> > + dev_warn(&dev->pdev->dev,
> > + "overflow pages in completion queue\n");
> > + ret = -EINVAL;
> > + goto err_umem;
> > + }
> > +
> > + ret = pvrdma_page_dir_init(dev, &cq->pdir, npages, cq->is_kernel);
> > + if (ret) {
> > + dev_warn(&dev->pdev->dev,
> > + "could not allocate page directory\n");
> > + goto err_umem;
> > + }
> > +
> > + if (cq->is_kernel) {
> > + /* Ring state is always the first page. */
> > + cq->ring_state = cq->pdir.pages[0];
> > + } else {
> > + pvrdma_page_dir_insert_umem(&cq->pdir, cq->umem, 0);
> > + }
> > +
> > + atomic_set(&cq->refcnt, 1);
> > + init_waitqueue_head(&cq->wait);
> > + spin_lock_init(&cq->cq_lock);
> > +
> > + memset(cmd, 0, sizeof(*cmd));
> > + cmd->hdr.cmd = PVRDMA_CMD_CREATE_CQ;
> > + cmd->nchunks = npages;
> > + cmd->ctx_handle = (context) ?
> > + (u64)to_vucontext(context)->ctx_handle : 0;
> > + cmd->cqe = entries;
> > + cmd->pdir_dma = cq->pdir.dir_dma;
> > + ret = pvrdma_cmd_post(dev, &req, &rsp);
> > +
> > + if (ret < 0 || rsp.hdr.ack != PVRDMA_CMD_CREATE_CQ_RESP) {
> > + dev_warn(&dev->pdev->dev,
> > + "could not create completion queue\n");
> > + goto err_page_dir;
> > + }
> > +
> > + cq->ibcq.cqe = resp->cqe;
> > + cq->cq_handle = resp->cq_handle;
> > + spin_lock_irqsave(&dev->cq_tbl_lock, flags);
> > + dev->cq_tbl[cq->cq_handle % dev->dsr->caps.max_cq] = cq;
> > + spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
> > +
> > + if (context) {
> > + cq->uar = &(to_vucontext(context)->uar);
> > +
> > + /* Copy udata back. */
> > + if (ib_copy_to_udata(udata, &cq->cq_handle,
> sizeof(__u32))) {
> > + dev_warn(&dev->pdev->dev,
> > + "failed to copy back udata\n");
> > + ret = -EINVAL;
>
> Don't we need to destroy the cq here?
You are right. Added.
>
> > + goto err_page_dir;
> > + }
> > + }
> > +
> > + return &cq->ibcq;
> > +
> > +err_page_dir:
> > + pvrdma_page_dir_cleanup(dev, &cq->pdir);
> > +err_umem:
> > + if (context)
> > + ib_umem_release(cq->umem);
> > +err_cq:
> > + atomic_dec(&dev->num_cqs);
> > + kfree(cq);
> > +
> > + return ERR_PTR(ret);
> > +}
> > +
> > +static void pvrdma_free_cq(struct pvrdma_dev *dev, struct pvrdma_cq
> > +*cq) {
> > + atomic_dec(&cq->refcnt);
> > + wait_event(cq->wait, !atomic_read(&cq->refcnt));
> > +
> > + if (!cq->is_kernel)
> > + ib_umem_release(cq->umem);
> > +
> > + pvrdma_page_dir_cleanup(dev, &cq->pdir);
> > + kfree(cq);
> > +}
> > +
> > +/**
> > + * pvrdma_destroy_cq - destroy completion queue
> > + * @cq: the completion queue to destroy.
> > + *
> > + * @return: 0 for success.
> > + */
> > +int pvrdma_destroy_cq(struct ib_cq *cq) {
> > + struct pvrdma_cq *vcq = to_vcq(cq);
> > + union pvrdma_cmd_req req;
> > + struct pvrdma_cmd_destroy_cq *cmd = &req.destroy_cq;
> > + struct pvrdma_dev *dev = to_vdev(cq->device);
> > + unsigned long flags;
> > + int ret;
> > +
> > + memset(cmd, 0, sizeof(*cmd));
> > + cmd->hdr.cmd = PVRDMA_CMD_DESTROY_CQ;
> > + cmd->cq_handle = vcq->cq_handle;
> > +
> > + ret = pvrdma_cmd_post(dev, &req, NULL);
> > + if (ret < 0)
> > + dev_warn(&dev->pdev->dev,
> > + "could not destroy completion queue\n");
> > +
> > + /* free cq's resources */
> > + spin_lock_irqsave(&dev->cq_tbl_lock, flags);
> > + dev->cq_tbl[vcq->cq_handle] = NULL;
> > + spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
> > +
> > + pvrdma_free_cq(dev, vcq);
> > + atomic_dec(&dev->num_cqs);
> > +
> > + return ret;
> > +}
> > +
> > +/**
> > + * pvrdma_modify_cq - modify the CQ moderation parameters
> > + * @ibcq: the CQ to modify
> > + * @cq_count: number of CQEs that will trigger an event
> > + * @cq_period: max period of time in usec before triggering an event
> > + *
> > + * @return: -EOPNOTSUPP as CQ resize is not supported.
> > + */
> > +int pvrdma_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period) {
> > + return -EOPNOTSUPP;
> > +}
> > +
> > +static inline struct pvrdma_cqe *get_cqe(struct pvrdma_cq *cq, int i)
> > +{
> > + return (struct pvrdma_cqe *)pvrdma_page_dir_get_ptr(
> > + &cq->pdir,
> > + cq->offset +
> > + sizeof(struct pvrdma_cqe) * i);
> > +}
> > +
> > +void pvrdma_flush_cqe(struct pvrdma_qp *qp, struct pvrdma_cq *cq) {
> > + int head;
> > + int has_data;
> > +
> > + if (!cq->is_kernel)
> > + return;
> > +
> > + /* Lock held */
> > + has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
> > + cq->ibcq.cqe, &head);
> > + if (unlikely(has_data > 0)) {
> > + int items;
> > + int curr;
> > + int tail = pvrdma_idx(&cq->ring_state->rx.prod_tail,
> > + cq->ibcq.cqe);
> > + struct pvrdma_cqe *cqe;
> > + struct pvrdma_cqe *curr_cqe;
> > +
> > + items = (tail > head) ? (tail - head) :
> > + (cq->ibcq.cqe - head + tail);
> > + curr = --tail;
> > + while (items-- > 0) {
> > + if (curr < 0)
> > + curr = cq->ibcq.cqe - 1;
> > + if (tail < 0)
> > + tail = cq->ibcq.cqe - 1;
> > + curr_cqe = get_cqe(cq, curr);
> > + if ((curr_cqe->qp & 0xFFFF) != qp->qp_handle) {
> > + if (curr != tail) {
> > + cqe = get_cqe(cq, tail);
> > + *cqe = *curr_cqe;
> > + }
> > + tail--;
> > + } else {
> > + pvrdma_idx_ring_inc(
> > + &cq->ring_state->rx.cons_head,
> > + cq->ibcq.cqe);
> > + }
> > + curr--;
> > + }
> > + }
> > +}
> > +
> > +static int pvrdma_poll_one(struct pvrdma_cq *cq, struct pvrdma_qp
> **cur_qp,
> > + struct ib_wc *wc)
> > +{
> > + struct pvrdma_dev *dev = to_vdev(cq->ibcq.device);
> > + int has_data;
> > + unsigned int head;
> > + bool tried = false;
> > + struct pvrdma_cqe *cqe;
> > +
> > +retry:
> > + has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
> > + cq->ibcq.cqe, &head);
> > + if (has_data == 0) {
> > + u32 val;
> > +
> > + if (tried)
> > + return -EAGAIN;
> > +
> > + /* Pass down POLL to give physical HCA a chance to poll. */
> > + val = cq->cq_handle | PVRDMA_UAR_CQ_POLL;
> > + writel(cpu_to_le32(val),
> > + dev->driver_uar.map + PVRDMA_UAR_CQ_OFFSET);
> > +
> > + tried = true;
> > + goto retry;
> > + } else if (has_data == PVRDMA_INVALID_IDX) {
> > + /* Ring state could be corrupted. Treat as unrecoverable. */
> > + return -EINVAL;
>
> Please try your best not to return anything but 0 or -EAGAIN from this
> function.
> Returning -EINVAL will cause caller (pvrdma_poll_cq) to drop all successive
> CQE in tn batch.
This reason I want to return -EINVAL here is that its possible that the ring state is
corrupted so any successive calls to poll_cq will likely not give any valid data.
If that means successive CQEs are dropped then that should be okay since we have
no idea now which CQEs were consumed.
Thanks,
Adit
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
