Setup RDMA admin queues using device command exposed over
auxiliary device and manage these queues using ida.
Co-developed-by: Andrew Boyer <andrew.boyer@amd.com>
Signed-off-by: Andrew Boyer <andrew.boyer@amd.com>
Co-developed-by: Allen Hubbe <allen.hubbe@amd.com>
Signed-off-by: Allen Hubbe <allen.hubbe@amd.com>
Signed-off-by: Abhijit Gangurde <abhijit.gangurde@amd.com>
---
v2->v3
- Fixed lockdep warning
- Used IDA for resource id allocation
- Removed rw locks around xarrays
drivers/infiniband/hw/ionic/ionic_admin.c | 1169 +++++++++++++++++
.../infiniband/hw/ionic/ionic_controlpath.c | 184 +++
drivers/infiniband/hw/ionic/ionic_fw.h | 164 +++
drivers/infiniband/hw/ionic/ionic_ibdev.c | 56 +
drivers/infiniband/hw/ionic/ionic_ibdev.h | 225 ++++
drivers/infiniband/hw/ionic/ionic_pgtbl.c | 113 ++
drivers/infiniband/hw/ionic/ionic_queue.c | 52 +
drivers/infiniband/hw/ionic/ionic_queue.h | 234 ++++
drivers/infiniband/hw/ionic/ionic_res.h | 154 +++
9 files changed, 2351 insertions(+)
create mode 100644 drivers/infiniband/hw/ionic/ionic_admin.c
create mode 100644 drivers/infiniband/hw/ionic/ionic_controlpath.c
create mode 100644 drivers/infiniband/hw/ionic/ionic_fw.h
create mode 100644 drivers/infiniband/hw/ionic/ionic_pgtbl.c
create mode 100644 drivers/infiniband/hw/ionic/ionic_queue.c
create mode 100644 drivers/infiniband/hw/ionic/ionic_queue.h
create mode 100644 drivers/infiniband/hw/ionic/ionic_res.h
diff --git a/drivers/infiniband/hw/ionic/ionic_admin.c b/drivers/infiniband/hw/ionic/ionic_admin.c
new file mode 100644
index 000000000000..bf2ab58e3f13
--- /dev/null
+++ b/drivers/infiniband/hw/ionic/ionic_admin.c
@@ -0,0 +1,1169 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (C) 2018-2025, Advanced Micro Devices, Inc. */
+
+#include <linux/interrupt.h>
+#include <linux/module.h>
+#include <linux/printk.h>
+
+#include "ionic_fw.h"
+#include "ionic_ibdev.h"
+
+#define IONIC_EQ_COUNT_MIN 4
+#define IONIC_AQ_COUNT_MIN 1
+
+/* not a valid queue position or negative error status */
+#define IONIC_ADMIN_POSTED 0x10000
+
+/* cpu can be held with irq disabled for COUNT * MS (for create/destroy_ah) */
+#define IONIC_ADMIN_BUSY_RETRY_COUNT 2000
+#define IONIC_ADMIN_BUSY_RETRY_MS 1
+
+/* admin queue will be considered failed if a command takes longer */
+#define IONIC_ADMIN_TIMEOUT (HZ * 2)
+#define IONIC_ADMIN_WARN (HZ / 8)
+
+/* will poll for admin cq to tolerate and report from missed event */
+#define IONIC_ADMIN_DELAY (HZ / 8)
+
+/* work queue for polling the event queue and admin cq */
+struct workqueue_struct *ionic_evt_workq;
+
+static void ionic_admin_timedout(struct ionic_aq *aq)
+{
+ struct ionic_cq *cq = &aq->vcq->cq[0];
+ struct ionic_ibdev *dev = aq->dev;
+ unsigned long irqflags;
+ u16 pos;
+
+ spin_lock_irqsave(&aq->lock, irqflags);
+ if (ionic_queue_empty(&aq->q))
+ goto out;
+
+ /* Reset ALL adminq if any one times out */
+ if (aq->admin_state < IONIC_ADMIN_KILLED)
+ queue_work(ionic_evt_workq, &dev->reset_work);
+
+ ibdev_err(&dev->ibdev, "admin command timed out, aq %d\n", aq->aqid);
+
+ ibdev_warn(&dev->ibdev, "admin timeout was set for %ums\n",
+ (u32)jiffies_to_msecs(IONIC_ADMIN_TIMEOUT));
+ ibdev_warn(&dev->ibdev, "admin inactivity for %ums\n",
+ (u32)jiffies_to_msecs(jiffies - aq->stamp));
+
+ ibdev_warn(&dev->ibdev, "admin commands outstanding %u\n",
+ ionic_queue_length(&aq->q));
+ ibdev_warn(&dev->ibdev, "%s more commands pending\n",
+ list_empty(&aq->wr_post) ? "no" : "some");
+
+ pos = cq->q.prod;
+
+ ibdev_warn(&dev->ibdev, "admin cq pos %u (next to complete)\n", pos);
+ print_hex_dump(KERN_WARNING, "cqe ", DUMP_PREFIX_OFFSET, 16, 1,
+ ionic_queue_at(&cq->q, pos),
+ BIT(cq->q.stride_log2), true);
+
+ pos = (pos - 1) & cq->q.mask;
+
+ ibdev_warn(&dev->ibdev, "admin cq pos %u (last completed)\n", pos);
+ print_hex_dump(KERN_WARNING, "cqe ", DUMP_PREFIX_OFFSET, 16, 1,
+ ionic_queue_at(&cq->q, pos),
+ BIT(cq->q.stride_log2), true);
+
+ pos = aq->q.cons;
+
+ ibdev_warn(&dev->ibdev, "admin pos %u (next to complete)\n", pos);
+ print_hex_dump(KERN_WARNING, "cmd ", DUMP_PREFIX_OFFSET, 16, 1,
+ ionic_queue_at(&aq->q, pos),
+ BIT(aq->q.stride_log2), true);
+
+ pos = (aq->q.prod - 1) & aq->q.mask;
+ if (pos == aq->q.cons)
+ goto out;
+
+ ibdev_warn(&dev->ibdev, "admin pos %u (last posted)\n", pos);
+ print_hex_dump(KERN_WARNING, "cmd ", DUMP_PREFIX_OFFSET, 16, 1,
+ ionic_queue_at(&aq->q, pos),
+ BIT(aq->q.stride_log2), true);
+
+out:
+ spin_unlock_irqrestore(&aq->lock, irqflags);
+}
+
+static void ionic_admin_reset_dwork(struct ionic_ibdev *dev)
+{
+ if (atomic_read(&dev->admin_state) >= IONIC_ADMIN_KILLED)
+ return;
+
+ queue_delayed_work(ionic_evt_workq, &dev->admin_dwork,
+ IONIC_ADMIN_DELAY);
+}
+
+static void ionic_admin_reset_wdog(struct ionic_aq *aq)
+{
+ if (aq->admin_state >= IONIC_ADMIN_KILLED)
+ return;
+
+ aq->stamp = jiffies;
+ ionic_admin_reset_dwork(aq->dev);
+}
+
+static bool ionic_admin_next_cqe(struct ionic_ibdev *dev, struct ionic_cq *cq,
+ struct ionic_v1_cqe **cqe)
+{
+ struct ionic_v1_cqe *qcqe = ionic_queue_at_prod(&cq->q);
+
+ if (unlikely(cq->color != ionic_v1_cqe_color(qcqe)))
+ return false;
+
+ /* Prevent out-of-order reads of the CQE */
+ dma_rmb();
+
+ ibdev_dbg(&dev->ibdev, "poll admin cq %u prod %u\n",
+ cq->cqid, cq->q.prod);
+ print_hex_dump_debug("cqe ", DUMP_PREFIX_OFFSET, 16, 1,
+ qcqe, BIT(cq->q.stride_log2), true);
+ *cqe = qcqe;
+
+ return true;
+}
+
+static void ionic_admin_poll_locked(struct ionic_aq *aq)
+{
+ struct ionic_cq *cq = &aq->vcq->cq[0];
+ struct ionic_admin_wr *wr, *wr_next;
+ struct ionic_ibdev *dev = aq->dev;
+ u32 wr_strides, avlbl_strides;
+ struct ionic_v1_cqe *cqe;
+ u32 qtf, qid;
+ u16 old_prod;
+ u8 type;
+
+ lockdep_assert_held(&aq->lock);
+
+ if (aq->admin_state >= IONIC_ADMIN_KILLED) {
+ list_for_each_entry_safe(wr, wr_next, &aq->wr_prod, aq_ent) {
+ INIT_LIST_HEAD(&wr->aq_ent);
+ aq->q_wr[wr->status].wr = NULL;
+ wr->status = aq->admin_state;
+ complete_all(&wr->work);
+ }
+ INIT_LIST_HEAD(&aq->wr_prod);
+
+ list_for_each_entry_safe(wr, wr_next, &aq->wr_post, aq_ent) {
+ INIT_LIST_HEAD(&wr->aq_ent);
+ wr->status = aq->admin_state;
+ complete_all(&wr->work);
+ }
+ INIT_LIST_HEAD(&aq->wr_post);
+
+ return;
+ }
+
+ old_prod = cq->q.prod;
+
+ while (ionic_admin_next_cqe(dev, cq, &cqe)) {
+ qtf = ionic_v1_cqe_qtf(cqe);
+ qid = ionic_v1_cqe_qtf_qid(qtf);
+ type = ionic_v1_cqe_qtf_type(qtf);
+
+ if (unlikely(type != IONIC_V1_CQE_TYPE_ADMIN)) {
+ ibdev_warn_ratelimited(&dev->ibdev,
+ "bad cqe type %u\n", type);
+ goto cq_next;
+ }
+
+ if (unlikely(qid != aq->aqid)) {
+ ibdev_warn_ratelimited(&dev->ibdev,
+ "bad cqe qid %u\n", qid);
+ goto cq_next;
+ }
+
+ if (unlikely(be16_to_cpu(cqe->admin.cmd_idx) != aq->q.cons)) {
+ ibdev_warn_ratelimited(&dev->ibdev,
+ "bad idx %u cons %u qid %u\n",
+ be16_to_cpu(cqe->admin.cmd_idx),
+ aq->q.cons, qid);
+ goto cq_next;
+ }
+
+ if (unlikely(ionic_queue_empty(&aq->q))) {
+ ibdev_warn_ratelimited(&dev->ibdev,
+ "bad cqe for empty adminq\n");
+ goto cq_next;
+ }
+
+ wr = aq->q_wr[aq->q.cons].wr;
+ if (wr) {
+ aq->q_wr[aq->q.cons].wr = NULL;
+ list_del_init(&wr->aq_ent);
+
+ wr->cqe = *cqe;
+ wr->status = aq->admin_state;
+ complete_all(&wr->work);
+ }
+
+ ionic_queue_consume_entries(&aq->q,
+ aq->q_wr[aq->q.cons].wqe_strides);
+
+cq_next:
+ ionic_queue_produce(&cq->q);
+ cq->color = ionic_color_wrap(cq->q.prod, cq->color);
+ }
+
+ if (old_prod != cq->q.prod) {
+ ionic_admin_reset_wdog(aq);
+ cq->q.cons = cq->q.prod;
+ ionic_dbell_ring(dev->lif_cfg.dbpage, dev->lif_cfg.cq_qtype,
+ ionic_queue_dbell_val(&cq->q));
+ queue_work(ionic_evt_workq, &aq->work);
+ } else if (!aq->armed) {
+ aq->armed = true;
+ cq->arm_any_prod = ionic_queue_next(&cq->q, cq->arm_any_prod);
+ ionic_dbell_ring(dev->lif_cfg.dbpage, dev->lif_cfg.cq_qtype,
+ cq->q.dbell | IONIC_CQ_RING_ARM |
+ cq->arm_any_prod);
+ queue_work(ionic_evt_workq, &aq->work);
+ }
+
+ if (aq->admin_state != IONIC_ADMIN_ACTIVE)
+ return;
+
+ old_prod = aq->q.prod;
+
+ if (ionic_queue_empty(&aq->q) && !list_empty(&aq->wr_post))
+ ionic_admin_reset_wdog(aq);
+
+ if (list_empty(&aq->wr_post))
+ return;
+
+ do {
+ u8 *src;
+ int i, src_len;
+ size_t stride_len;
+
+ wr = list_first_entry(&aq->wr_post, struct ionic_admin_wr,
+ aq_ent);
+ wr_strides = (le16_to_cpu(wr->wqe.len) + ADMIN_WQE_HDR_LEN +
+ (ADMIN_WQE_STRIDE - 1)) >> aq->q.stride_log2;
+ avlbl_strides = ionic_queue_length_remaining(&aq->q);
+
+ if (wr_strides > avlbl_strides)
+ break;
+
+ list_move(&wr->aq_ent, &aq->wr_prod);
+ wr->status = aq->q.prod;
+ aq->q_wr[aq->q.prod].wr = wr;
+ aq->q_wr[aq->q.prod].wqe_strides = wr_strides;
+
+ src_len = le16_to_cpu(wr->wqe.len);
+ src = (uint8_t *)&wr->wqe.cmd;
+
+ /* First stride */
+ memcpy(ionic_queue_at_prod(&aq->q), &wr->wqe,
+ ADMIN_WQE_HDR_LEN);
+ stride_len = ADMIN_WQE_STRIDE - ADMIN_WQE_HDR_LEN;
+ if (stride_len > src_len)
+ stride_len = src_len;
+ memcpy(ionic_queue_at_prod(&aq->q) + ADMIN_WQE_HDR_LEN,
+ src, stride_len);
+ ibdev_dbg(&dev->ibdev, "post admin prod %u (%u strides)\n",
+ aq->q.prod, wr_strides);
+ print_hex_dump_debug("wqe ", DUMP_PREFIX_OFFSET, 16, 1,
+ ionic_queue_at_prod(&aq->q),
+ BIT(aq->q.stride_log2), true);
+ ionic_queue_produce(&aq->q);
+
+ /* Remaining strides */
+ for (i = stride_len; i < src_len; i += stride_len) {
+ stride_len = ADMIN_WQE_STRIDE;
+
+ if (i + stride_len > src_len)
+ stride_len = src_len - i;
+
+ memcpy(ionic_queue_at_prod(&aq->q), src + i,
+ stride_len);
+ print_hex_dump_debug("wqe ", DUMP_PREFIX_OFFSET, 16, 1,
+ ionic_queue_at_prod(&aq->q),
+ BIT(aq->q.stride_log2), true);
+ ionic_queue_produce(&aq->q);
+ }
+ } while (!list_empty(&aq->wr_post));
+
+ if (old_prod != aq->q.prod)
+ ionic_dbell_ring(dev->lif_cfg.dbpage, dev->lif_cfg.aq_qtype,
+ ionic_queue_dbell_val(&aq->q));
+}
+
+static void ionic_admin_dwork(struct work_struct *ws)
+{
+ struct ionic_ibdev *dev =
+ container_of(ws, struct ionic_ibdev, admin_dwork.work);
+ struct ionic_aq *aq, *bad_aq = NULL;
+ bool do_reschedule = false;
+ unsigned long irqflags;
+ bool do_reset = false;
+ u16 pos;
+ int i;
+
+ for (i = 0; i < dev->lif_cfg.aq_count; i++) {
+ aq = dev->aq_vec[i];
+
+ spin_lock_irqsave(&aq->lock, irqflags);
+
+ if (ionic_queue_empty(&aq->q))
+ goto next_aq;
+
+ /* Reschedule if any queue has outstanding work */
+ do_reschedule = true;
+
+ if (time_is_after_eq_jiffies(aq->stamp + IONIC_ADMIN_WARN))
+ /* Warning threshold not met, nothing to do */
+ goto next_aq;
+
+ /* See if polling now makes some progress */
+ pos = aq->q.cons;
+ ionic_admin_poll_locked(aq);
+ if (pos != aq->q.cons) {
+ ibdev_dbg(&dev->ibdev,
+ "missed event for acq %d\n", aq->cqid);
+ goto next_aq;
+ }
+
+ if (time_is_after_eq_jiffies(aq->stamp +
+ IONIC_ADMIN_TIMEOUT)) {
+ /* Timeout threshold not met */
+ ibdev_dbg(&dev->ibdev, "no progress after %ums\n",
+ (u32)jiffies_to_msecs(jiffies - aq->stamp));
+ goto next_aq;
+ }
+
+ /* Queue timed out */
+ bad_aq = aq;
+ do_reset = true;
+next_aq:
+ spin_unlock_irqrestore(&aq->lock, irqflags);
+ }
+
+ if (do_reset)
+ /* Reset device on a timeout */
+ ionic_admin_timedout(bad_aq);
+ else if (do_reschedule)
+ /* Try to poll again later */
+ ionic_admin_reset_dwork(dev);
+}
+
+static void ionic_admin_work(struct work_struct *ws)
+{
+ struct ionic_aq *aq = container_of(ws, struct ionic_aq, work);
+ unsigned long irqflags;
+
+ spin_lock_irqsave(&aq->lock, irqflags);
+ ionic_admin_poll_locked(aq);
+ spin_unlock_irqrestore(&aq->lock, irqflags);
+}
+
+static void ionic_admin_post_aq(struct ionic_aq *aq, struct ionic_admin_wr *wr)
+{
+ unsigned long irqflags;
+ bool poll;
+
+ wr->status = IONIC_ADMIN_POSTED;
+ wr->aq = aq;
+
+ spin_lock_irqsave(&aq->lock, irqflags);
+ poll = list_empty(&aq->wr_post);
+ list_add(&wr->aq_ent, &aq->wr_post);
+ if (poll)
+ ionic_admin_poll_locked(aq);
+ spin_unlock_irqrestore(&aq->lock, irqflags);
+}
+
+void ionic_admin_post(struct ionic_ibdev *dev, struct ionic_admin_wr *wr)
+{
+ int aq_idx;
+
+ /* Use cpu id for the adminq selection */
+ aq_idx = raw_smp_processor_id() % dev->lif_cfg.aq_count;
+ ionic_admin_post_aq(dev->aq_vec[aq_idx], wr);
+}
+
+static void ionic_admin_cancel(struct ionic_admin_wr *wr)
+{
+ struct ionic_aq *aq = wr->aq;
+ unsigned long irqflags;
+
+ spin_lock_irqsave(&aq->lock, irqflags);
+
+ if (!list_empty(&wr->aq_ent)) {
+ list_del(&wr->aq_ent);
+ if (wr->status != IONIC_ADMIN_POSTED)
+ aq->q_wr[wr->status].wr = NULL;
+ }
+
+ spin_unlock_irqrestore(&aq->lock, irqflags);
+}
+
+static int ionic_admin_busy_wait(struct ionic_admin_wr *wr)
+{
+ struct ionic_aq *aq = wr->aq;
+ unsigned long irqflags;
+ int try_i;
+
+ for (try_i = 0; try_i < IONIC_ADMIN_BUSY_RETRY_COUNT; ++try_i) {
+ if (completion_done(&wr->work))
+ return 0;
+
+ mdelay(IONIC_ADMIN_BUSY_RETRY_MS);
+
+ spin_lock_irqsave(&aq->lock, irqflags);
+ ionic_admin_poll_locked(aq);
+ spin_unlock_irqrestore(&aq->lock, irqflags);
+ }
+
+ /*
+ * we timed out. Initiate RDMA LIF reset and indicate
+ * error to caller.
+ */
+ ionic_admin_timedout(aq);
+ return -ETIMEDOUT;
+}
+
+int ionic_admin_wait(struct ionic_ibdev *dev, struct ionic_admin_wr *wr,
+ enum ionic_admin_flags flags)
+{
+ int rc, timo;
+
+ if (flags & IONIC_ADMIN_F_BUSYWAIT) {
+ /* Spin */
+ rc = ionic_admin_busy_wait(wr);
+ } else if (flags & IONIC_ADMIN_F_INTERRUPT) {
+ /*
+ * Interruptible sleep, 1s timeout
+ * This is used for commands which are safe for the caller
+ * to clean up without killing and resetting the adminq.
+ */
+ timo = wait_for_completion_interruptible_timeout(&wr->work,
+ HZ);
+ if (timo > 0)
+ rc = 0;
+ else if (timo == 0)
+ rc = -ETIMEDOUT;
+ else
+ rc = timo;
+ } else {
+ /*
+ * Uninterruptible sleep
+ * This is used for commands which are NOT safe for the
+ * caller to clean up. Cleanup must be handled by the
+ * adminq kill and reset process so that host memory is
+ * not corrupted by the device.
+ */
+ wait_for_completion(&wr->work);
+ rc = 0;
+ }
+
+ if (rc) {
+ ibdev_warn(&dev->ibdev, "wait status %d\n", rc);
+ ionic_admin_cancel(wr);
+ } else if (wr->status == IONIC_ADMIN_KILLED) {
+ ibdev_dbg(&dev->ibdev, "admin killed\n");
+
+ /* No error if admin already killed during teardown */
+ rc = (flags & IONIC_ADMIN_F_TEARDOWN) ? 0 : -ENODEV;
+ } else if (ionic_v1_cqe_error(&wr->cqe)) {
+ ibdev_warn(&dev->ibdev, "opcode %u error %u\n",
+ wr->wqe.op,
+ be32_to_cpu(wr->cqe.status_length));
+ rc = -EINVAL;
+ }
+ return rc;
+}
+
+static int ionic_rdma_devcmd(struct ionic_ibdev *dev,
+ struct ionic_admin_ctx *admin)
+{
+ int rc;
+
+ rc = ionic_adminq_post_wait(dev->lif_cfg.lif, admin);
+ if (rc)
+ return rc;
+
+ return ionic_error_to_errno(admin->comp.comp.status);
+}
+
+int ionic_rdma_reset_devcmd(struct ionic_ibdev *dev)
+{
+ struct ionic_admin_ctx admin = {
+ .work = COMPLETION_INITIALIZER_ONSTACK(admin.work),
+ .cmd.rdma_reset = {
+ .opcode = IONIC_CMD_RDMA_RESET_LIF,
+ .lif_index = cpu_to_le16(dev->lif_cfg.lif_index),
+ },
+ };
+
+ return ionic_rdma_devcmd(dev, &admin);
+}
+
+static int ionic_rdma_queue_devcmd(struct ionic_ibdev *dev,
+ struct ionic_queue *q,
+ u32 qid, u32 cid, u16 opcode)
+{
+ struct ionic_admin_ctx admin = {
+ .work = COMPLETION_INITIALIZER_ONSTACK(admin.work),
+ .cmd.rdma_queue = {
+ .opcode = opcode,
+ .lif_index = cpu_to_le16(dev->lif_cfg.lif_index),
+ .qid_ver = cpu_to_le32(qid),
+ .cid = cpu_to_le32(cid),
+ .dbid = cpu_to_le16(dev->lif_cfg.dbid),
+ .depth_log2 = q->depth_log2,
+ .stride_log2 = q->stride_log2,
+ .dma_addr = cpu_to_le64(q->dma),
+ },
+ };
+
+ return ionic_rdma_devcmd(dev, &admin);
+}
+
+static void ionic_rdma_admincq_comp(struct ib_cq *ibcq, void *cq_context)
+{
+ struct ionic_aq *aq = cq_context;
+ unsigned long irqflags;
+
+ spin_lock_irqsave(&aq->lock, irqflags);
+ aq->armed = false;
+ if (aq->admin_state < IONIC_ADMIN_KILLED)
+ queue_work(ionic_evt_workq, &aq->work);
+ spin_unlock_irqrestore(&aq->lock, irqflags);
+}
+
+static void ionic_rdma_admincq_event(struct ib_event *event, void *cq_context)
+{
+ struct ionic_aq *aq = cq_context;
+
+ ibdev_err(&aq->dev->ibdev, "admincq event %d\n", event->event);
+}
+
+static struct ionic_vcq *ionic_create_rdma_admincq(struct ionic_ibdev *dev,
+ int comp_vector)
+{
+ struct ib_cq_init_attr attr = {
+ .cqe = IONIC_AQ_DEPTH,
+ .comp_vector = comp_vector,
+ };
+ struct ionic_tbl_buf buf = {};
+ struct ionic_vcq *vcq;
+ struct ionic_cq *cq;
+ int rc;
+
+ vcq = kzalloc(sizeof(*vcq), GFP_KERNEL);
+ if (!vcq) {
+ rc = -ENOMEM;
+ goto err_alloc;
+ }
+
+ vcq->ibcq.device = &dev->ibdev;
+ vcq->ibcq.uobject = NULL;
+ vcq->ibcq.comp_handler = ionic_rdma_admincq_comp;
+ vcq->ibcq.event_handler = ionic_rdma_admincq_event;
+ vcq->ibcq.cq_context = NULL;
+ atomic_set(&vcq->ibcq.usecnt, 0);
+
+ vcq->udma_mask = 1;
+ cq = &vcq->cq[0];
+
+ rc = ionic_create_cq_common(vcq, &buf, &attr, NULL, NULL,
+ NULL, NULL, 0);
+ if (rc)
+ goto err_init;
+
+ rc = ionic_rdma_queue_devcmd(dev, &cq->q, cq->cqid, cq->eqid,
+ IONIC_CMD_RDMA_CREATE_CQ);
+ if (rc)
+ goto err_cmd;
+
+ return vcq;
+
+err_cmd:
+ ionic_destroy_cq_common(dev, cq);
+err_init:
+ kfree(vcq);
+err_alloc:
+ return ERR_PTR(rc);
+}
+
+static struct ionic_aq *__ionic_create_rdma_adminq(struct ionic_ibdev *dev,
+ u32 aqid, u32 cqid)
+{
+ struct ionic_aq *aq;
+ int rc;
+
+ aq = kzalloc(sizeof(*aq), GFP_KERNEL);
+ if (!aq) {
+ rc = -ENOMEM;
+ goto err_aq;
+ }
+
+ aq->admin_state = IONIC_ADMIN_KILLED;
+ aq->dev = dev;
+ aq->aqid = aqid;
+ aq->cqid = cqid;
+ spin_lock_init(&aq->lock);
+
+ rc = ionic_queue_init(&aq->q, dev->lif_cfg.hwdev, IONIC_EQ_DEPTH,
+ ADMIN_WQE_STRIDE);
+ if (rc)
+ goto err_q;
+
+ ionic_queue_dbell_init(&aq->q, aq->aqid);
+
+ aq->q_wr = kcalloc((u32)aq->q.mask + 1, sizeof(*aq->q_wr), GFP_KERNEL);
+ if (!aq->q_wr) {
+ rc = -ENOMEM;
+ goto err_wr;
+ }
+
+ INIT_LIST_HEAD(&aq->wr_prod);
+ INIT_LIST_HEAD(&aq->wr_post);
+
+ INIT_WORK(&aq->work, ionic_admin_work);
+ aq->armed = false;
+
+ return aq;
+
+err_wr:
+ ionic_queue_destroy(&aq->q, dev->lif_cfg.hwdev);
+err_q:
+ kfree(aq);
+err_aq:
+ return ERR_PTR(rc);
+}
+
+static void __ionic_destroy_rdma_adminq(struct ionic_ibdev *dev,
+ struct ionic_aq *aq)
+{
+ ionic_queue_destroy(&aq->q, dev->lif_cfg.hwdev);
+ kfree(aq);
+}
+
+static struct ionic_aq *ionic_create_rdma_adminq(struct ionic_ibdev *dev,
+ u32 aqid, u32 cqid)
+{
+ struct ionic_aq *aq;
+ int rc;
+
+ aq = __ionic_create_rdma_adminq(dev, aqid, cqid);
+ if (IS_ERR(aq)) {
+ rc = PTR_ERR(aq);
+ goto err_aq;
+ }
+
+ rc = ionic_rdma_queue_devcmd(dev, &aq->q, aq->aqid, aq->cqid,
+ IONIC_CMD_RDMA_CREATE_ADMINQ);
+ if (rc)
+ goto err_cmd;
+
+ return aq;
+
+err_cmd:
+ __ionic_destroy_rdma_adminq(dev, aq);
+err_aq:
+ return ERR_PTR(rc);
+}
+
+static void ionic_kill_ibdev(struct ionic_ibdev *dev, bool fatal_path)
+{
+ unsigned long irqflags;
+ bool do_flush = false;
+ int i;
+
+ /* Mark AQs for drain and flush the QPs while irq is disabled */
+ local_irq_save(irqflags);
+
+ /* Mark the admin queue, flushing at most once */
+ for (i = 0; i < dev->lif_cfg.aq_count; i++) {
+ struct ionic_aq *aq = dev->aq_vec[i];
+
+ spin_lock(&aq->lock);
+ if (aq->admin_state != IONIC_ADMIN_KILLED) {
+ aq->admin_state = IONIC_ADMIN_KILLED;
+ /* Flush incomplete admin commands */
+ ionic_admin_poll_locked(aq);
+ do_flush = true;
+ }
+ spin_unlock(&aq->lock);
+ }
+
+ local_irq_restore(irqflags);
+
+ /* Post a fatal event if requested */
+ if (fatal_path) {
+ struct ib_event ev;
+
+ ev.device = &dev->ibdev;
+ ev.element.port_num = 1;
+ ev.event = IB_EVENT_DEVICE_FATAL;
+
+ ib_dispatch_event(&ev);
+ }
+
+ atomic_set(&dev->admin_state, IONIC_ADMIN_KILLED);
+}
+
+void ionic_kill_rdma_admin(struct ionic_ibdev *dev, bool fatal_path)
+{
+ enum ionic_admin_state old_state;
+ unsigned long irqflags = 0;
+ int i, rc;
+
+ if (!dev->aq_vec)
+ return;
+
+ /*
+ * Admin queues are transitioned from active to paused to killed state.
+ * When in paused state, no new commands are issued to the device,
+ * nor are any completed locally. After resetting the device, it will be
+ * safe to resume the rdma admin queues in the killed state. Commands
+ * will not be issued to the device, but will complete locally with status
+ * IONIC_ADMIN_KILLED. Handling completion will ensure that creating or
+ * modifying resources fails, but destroying resources succeeds.
+ * If there was a failure resetting the device using this strategy,
+ * then the state of the device is unknown.
+ */
+ old_state = atomic_cmpxchg(&dev->admin_state, IONIC_ADMIN_ACTIVE,
+ IONIC_ADMIN_PAUSED);
+ if (old_state != IONIC_ADMIN_ACTIVE)
+ return;
+
+ /* Pause all the AQs */
+ local_irq_save(irqflags);
+ for (i = 0; i < dev->lif_cfg.aq_count; i++) {
+ struct ionic_aq *aq = dev->aq_vec[i];
+
+ spin_lock(&aq->lock);
+ /* pause rdma admin queues to reset device */
+ if (aq->admin_state == IONIC_ADMIN_ACTIVE)
+ aq->admin_state = IONIC_ADMIN_PAUSED;
+ spin_unlock(&aq->lock);
+ }
+ local_irq_restore(irqflags);
+
+ rc = ionic_rdma_reset_devcmd(dev);
+ if (unlikely(rc)) {
+ ibdev_err(&dev->ibdev, "failed to reset rdma %d\n", rc);
+ ionic_request_rdma_reset(dev->lif_cfg.lif);
+ }
+
+ ionic_kill_ibdev(dev, fatal_path);
+}
+
+static void ionic_reset_work(struct work_struct *ws)
+{
+ struct ionic_ibdev *dev =
+ container_of(ws, struct ionic_ibdev, reset_work);
+
+ ionic_kill_rdma_admin(dev, true);
+}
+
+static bool ionic_next_eqe(struct ionic_eq *eq, struct ionic_v1_eqe *eqe)
+{
+ struct ionic_v1_eqe *qeqe;
+ bool color;
+
+ qeqe = ionic_queue_at_prod(&eq->q);
+ color = ionic_v1_eqe_color(qeqe);
+
+ /* cons is color for eq */
+ if (eq->q.cons != color)
+ return false;
+
+ /* Prevent out-of-order reads of the EQE */
+ dma_rmb();
+
+ ibdev_dbg(&eq->dev->ibdev, "poll eq prod %u\n", eq->q.prod);
+ print_hex_dump_debug("eqe ", DUMP_PREFIX_OFFSET, 16, 1,
+ qeqe, BIT(eq->q.stride_log2), true);
+ *eqe = *qeqe;
+
+ return true;
+}
+
+static void ionic_cq_event(struct ionic_ibdev *dev, u32 cqid, u8 code)
+{
+ struct ib_event ibev;
+ struct ionic_cq *cq;
+
+ rcu_read_lock();
+ cq = xa_load(&dev->cq_tbl, cqid);
+ if (cq)
+ kref_get(&cq->cq_kref);
+ rcu_read_unlock();
+
+ if (!cq) {
+ ibdev_dbg(&dev->ibdev,
+ "missing cqid %#x code %u\n", cqid, code);
+ return;
+ }
+
+ switch (code) {
+ case IONIC_V1_EQE_CQ_NOTIFY:
+ if (cq->vcq->ibcq.comp_handler)
+ cq->vcq->ibcq.comp_handler(&cq->vcq->ibcq,
+ cq->vcq->ibcq.cq_context);
+ break;
+
+ case IONIC_V1_EQE_CQ_ERR:
+ if (cq->vcq->ibcq.event_handler) {
+ ibev.event = IB_EVENT_CQ_ERR;
+ ibev.device = &dev->ibdev;
+ ibev.element.cq = &cq->vcq->ibcq;
+
+ cq->vcq->ibcq.event_handler(&ibev,
+ cq->vcq->ibcq.cq_context);
+ }
+ break;
+
+ default:
+ ibdev_dbg(&dev->ibdev,
+ "unrecognized cqid %#x code %u\n", cqid, code);
+ break;
+ }
+
+ kref_put(&cq->cq_kref, ionic_cq_complete);
+}
+
+static u16 ionic_poll_eq(struct ionic_eq *eq, u16 budget)
+{
+ struct ionic_ibdev *dev = eq->dev;
+ struct ionic_v1_eqe eqe;
+ u16 npolled = 0;
+ u8 type, code;
+ u32 evt, qid;
+
+ while (npolled < budget) {
+ if (!ionic_next_eqe(eq, &eqe))
+ break;
+
+ ionic_queue_produce(&eq->q);
+
+ /* cons is color for eq */
+ eq->q.cons = ionic_color_wrap(eq->q.prod, eq->q.cons);
+
+ ++npolled;
+
+ evt = ionic_v1_eqe_evt(&eqe);
+ type = ionic_v1_eqe_evt_type(evt);
+ code = ionic_v1_eqe_evt_code(evt);
+ qid = ionic_v1_eqe_evt_qid(evt);
+
+ switch (type) {
+ case IONIC_V1_EQE_TYPE_CQ:
+ ionic_cq_event(dev, qid, code);
+ break;
+
+ default:
+ ibdev_dbg(&dev->ibdev,
+ "unknown event %#x type %u\n", evt, type);
+ }
+ }
+
+ return npolled;
+}
+
+static void ionic_poll_eq_work(struct work_struct *work)
+{
+ struct ionic_eq *eq = container_of(work, struct ionic_eq, work);
+ u32 npolled;
+
+ if (unlikely(!eq->enable) || WARN_ON(eq->armed))
+ return;
+
+ npolled = ionic_poll_eq(eq, IONIC_EQ_WORK_BUDGET);
+ if (npolled == IONIC_EQ_WORK_BUDGET) {
+ ionic_intr_credits(eq->dev->lif_cfg.intr_ctrl, eq->intr,
+ npolled, 0);
+ queue_work(ionic_evt_workq, &eq->work);
+ } else {
+ xchg(&eq->armed, true);
+ ionic_intr_credits(eq->dev->lif_cfg.intr_ctrl, eq->intr,
+ 0, IONIC_INTR_CRED_UNMASK);
+ }
+}
+
+static irqreturn_t ionic_poll_eq_isr(int irq, void *eqptr)
+{
+ struct ionic_eq *eq = eqptr;
+ bool was_armed;
+ u32 npolled;
+
+ was_armed = xchg(&eq->armed, false);
+
+ if (unlikely(!eq->enable) || !was_armed)
+ return IRQ_HANDLED;
+
+ npolled = ionic_poll_eq(eq, IONIC_EQ_ISR_BUDGET);
+ if (npolled == IONIC_EQ_ISR_BUDGET) {
+ ionic_intr_credits(eq->dev->lif_cfg.intr_ctrl, eq->intr,
+ npolled, 0);
+ queue_work(ionic_evt_workq, &eq->work);
+ } else {
+ xchg(&eq->armed, true);
+ ionic_intr_credits(eq->dev->lif_cfg.intr_ctrl, eq->intr,
+ 0, IONIC_INTR_CRED_UNMASK);
+ }
+
+ return IRQ_HANDLED;
+}
+
+static struct ionic_eq *ionic_create_eq(struct ionic_ibdev *dev, int eqid)
+{
+ struct ionic_intr_info intr_obj = { };
+ struct ionic_eq *eq;
+ int rc;
+
+ eq = kzalloc(sizeof(*eq), GFP_KERNEL);
+ if (!eq) {
+ rc = -ENOMEM;
+ goto err_eq;
+ }
+
+ eq->dev = dev;
+
+ rc = ionic_queue_init(&eq->q, dev->lif_cfg.hwdev, IONIC_EQ_DEPTH,
+ sizeof(struct ionic_v1_eqe));
+ if (rc)
+ goto err_q;
+
+ eq->eqid = eqid;
+
+ eq->armed = true;
+ eq->enable = false;
+ INIT_WORK(&eq->work, ionic_poll_eq_work);
+
+ rc = ionic_intr_alloc(dev->lif_cfg.lif, &intr_obj);
+ if (rc < 0)
+ goto err_intr;
+
+ eq->irq = intr_obj.vector;
+ eq->intr = intr_obj.index;
+
+ ionic_queue_dbell_init(&eq->q, eq->eqid);
+
+ /* cons is color for eq */
+ eq->q.cons = true;
+
+ snprintf(eq->name, sizeof(eq->name), "%s-%d-%d-eq",
+ DRIVER_SHORTNAME, dev->lif_cfg.lif_index, eq->eqid);
+
+ ionic_intr_mask(dev->lif_cfg.intr_ctrl, eq->intr, IONIC_INTR_MASK_SET);
+ ionic_intr_mask_assert(dev->lif_cfg.intr_ctrl, eq->intr, IONIC_INTR_MASK_SET);
+ ionic_intr_coal_init(dev->lif_cfg.intr_ctrl, eq->intr, 0);
+ ionic_intr_clean(dev->lif_cfg.intr_ctrl, eq->intr);
+
+ eq->enable = true;
+
+ rc = request_irq(eq->irq, ionic_poll_eq_isr, 0, eq->name, eq);
+ if (rc)
+ goto err_irq;
+
+ rc = ionic_rdma_queue_devcmd(dev, &eq->q, eq->eqid, eq->intr,
+ IONIC_CMD_RDMA_CREATE_EQ);
+ if (rc)
+ goto err_cmd;
+
+ ionic_intr_mask(dev->lif_cfg.intr_ctrl, eq->intr, IONIC_INTR_MASK_CLEAR);
+
+ return eq;
+
+err_cmd:
+ eq->enable = false;
+ free_irq(eq->irq, eq);
+ flush_work(&eq->work);
+err_irq:
+ ionic_intr_free(dev->lif_cfg.lif, eq->intr);
+err_intr:
+ ionic_queue_destroy(&eq->q, dev->lif_cfg.hwdev);
+err_q:
+ kfree(eq);
+err_eq:
+ return ERR_PTR(rc);
+}
+
+static void ionic_destroy_eq(struct ionic_eq *eq)
+{
+ struct ionic_ibdev *dev = eq->dev;
+
+ eq->enable = false;
+ free_irq(eq->irq, eq);
+ flush_work(&eq->work);
+
+ ionic_intr_free(dev->lif_cfg.lif, eq->intr);
+ ionic_queue_destroy(&eq->q, dev->lif_cfg.hwdev);
+ kfree(eq);
+}
+
+int ionic_create_rdma_admin(struct ionic_ibdev *dev)
+{
+ int eq_i = 0, aq_i = 0, rc = 0;
+ struct ionic_vcq *vcq;
+ struct ionic_aq *aq;
+ struct ionic_eq *eq;
+
+ dev->eq_vec = NULL;
+ dev->aq_vec = NULL;
+
+ INIT_WORK(&dev->reset_work, ionic_reset_work);
+ INIT_DELAYED_WORK(&dev->admin_dwork, ionic_admin_dwork);
+ atomic_set(&dev->admin_state, IONIC_ADMIN_KILLED);
+
+ if (dev->lif_cfg.aq_count > IONIC_AQ_COUNT) {
+ ibdev_dbg(&dev->ibdev, "limiting adminq count to %d\n",
+ IONIC_AQ_COUNT);
+ dev->lif_cfg.aq_count = IONIC_AQ_COUNT;
+ }
+
+ if (dev->lif_cfg.eq_count > IONIC_EQ_COUNT) {
+ dev_dbg(&dev->ibdev.dev, "limiting eventq count to %d\n",
+ IONIC_EQ_COUNT);
+ dev->lif_cfg.eq_count = IONIC_EQ_COUNT;
+ }
+
+ /* need at least two eq and one aq */
+ if (dev->lif_cfg.eq_count < IONIC_EQ_COUNT_MIN ||
+ dev->lif_cfg.aq_count < IONIC_AQ_COUNT_MIN) {
+ rc = -EINVAL;
+ goto out;
+ }
+
+ dev->eq_vec = kmalloc_array(dev->lif_cfg.eq_count, sizeof(*dev->eq_vec),
+ GFP_KERNEL);
+ if (!dev->eq_vec) {
+ rc = -ENOMEM;
+ goto out;
+ }
+
+ for (eq_i = 0; eq_i < dev->lif_cfg.eq_count; ++eq_i) {
+ eq = ionic_create_eq(dev, eq_i + dev->lif_cfg.eq_base);
+ if (IS_ERR(eq)) {
+ rc = PTR_ERR(eq);
+
+ if (eq_i < IONIC_EQ_COUNT_MIN) {
+ ibdev_err(&dev->ibdev,
+ "fail create eq %d\n", rc);
+ goto out;
+ }
+
+ /* ok, just fewer eq than device supports */
+ ibdev_dbg(&dev->ibdev, "eq count %d want %d rc %d\n",
+ eq_i, dev->lif_cfg.eq_count, rc);
+
+ rc = 0;
+ break;
+ }
+
+ dev->eq_vec[eq_i] = eq;
+ }
+
+ dev->lif_cfg.eq_count = eq_i;
+
+ dev->aq_vec = kmalloc_array(dev->lif_cfg.aq_count, sizeof(*dev->aq_vec),
+ GFP_KERNEL);
+ if (!dev->aq_vec) {
+ rc = -ENOMEM;
+ goto out;
+ }
+
+ /* Create one CQ per AQ */
+ for (aq_i = 0; aq_i < dev->lif_cfg.aq_count; ++aq_i) {
+ vcq = ionic_create_rdma_admincq(dev, aq_i % eq_i);
+ if (IS_ERR(vcq)) {
+ rc = PTR_ERR(vcq);
+
+ if (!aq_i) {
+ ibdev_err(&dev->ibdev,
+ "failed to create acq %d\n", rc);
+ goto out;
+ }
+
+ /* ok, just fewer adminq than device supports */
+ ibdev_dbg(&dev->ibdev, "acq count %d want %d rc %d\n",
+ aq_i, dev->lif_cfg.aq_count, rc);
+ break;
+ }
+
+ aq = ionic_create_rdma_adminq(dev, aq_i + dev->lif_cfg.aq_base,
+ vcq->cq[0].cqid);
+ if (IS_ERR(aq)) {
+ /* Clean up the dangling CQ */
+ ionic_destroy_cq_common(dev, &vcq->cq[0]);
+ kfree(vcq);
+
+ rc = PTR_ERR(aq);
+
+ if (!aq_i) {
+ ibdev_err(&dev->ibdev,
+ "failed to create aq %d\n", rc);
+ goto out;
+ }
+
+ /* ok, just fewer adminq than device supports */
+ ibdev_dbg(&dev->ibdev, "aq count %d want %d rc %d\n",
+ aq_i, dev->lif_cfg.aq_count, rc);
+ break;
+ }
+
+ vcq->ibcq.cq_context = aq;
+ aq->vcq = vcq;
+
+ aq->admin_state = IONIC_ADMIN_ACTIVE;
+ dev->aq_vec[aq_i] = aq;
+ }
+
+ atomic_set(&dev->admin_state, IONIC_ADMIN_ACTIVE);
+out:
+ dev->lif_cfg.eq_count = eq_i;
+ dev->lif_cfg.aq_count = aq_i;
+
+ return rc;
+}
+
+void ionic_destroy_rdma_admin(struct ionic_ibdev *dev)
+{
+ struct ionic_vcq *vcq;
+ struct ionic_aq *aq;
+ struct ionic_eq *eq;
+
+ /*
+ * Killing the admin before destroy makes sure all admin and
+ * completions are flushed. admin_state = IONIC_ADMIN_KILLED
+ * stops queueing up further works.
+ */
+ cancel_delayed_work_sync(&dev->admin_dwork);
+ cancel_work_sync(&dev->reset_work);
+
+ if (dev->aq_vec) {
+ while (dev->lif_cfg.aq_count > 0) {
+ aq = dev->aq_vec[--dev->lif_cfg.aq_count];
+ vcq = aq->vcq;
+
+ cancel_work_sync(&aq->work);
+
+ __ionic_destroy_rdma_adminq(dev, aq);
+ if (vcq) {
+ ionic_destroy_cq_common(dev, &vcq->cq[0]);
+ kfree(vcq);
+ }
+ }
+
+ kfree(dev->aq_vec);
+ }
+
+ if (dev->eq_vec) {
+ while (dev->lif_cfg.eq_count > 0) {
+ eq = dev->eq_vec[--dev->lif_cfg.eq_count];
+ ionic_destroy_eq(eq);
+ }
+
+ kfree(dev->eq_vec);
+ }
+}
diff --git a/drivers/infiniband/hw/ionic/ionic_controlpath.c b/drivers/infiniband/hw/ionic/ionic_controlpath.c
new file mode 100644
index 000000000000..84a39e9fed40
--- /dev/null
+++ b/drivers/infiniband/hw/ionic/ionic_controlpath.c
@@ -0,0 +1,184 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (C) 2018-2025, Advanced Micro Devices, Inc. */
+
+#include "ionic_ibdev.h"
+
+static int ionic_validate_qdesc(struct ionic_qdesc *q)
+{
+ if (!q->addr || !q->size || !q->mask ||
+ !q->depth_log2 || !q->stride_log2)
+ return -EINVAL;
+
+ if (q->addr & (PAGE_SIZE - 1))
+ return -EINVAL;
+
+ if (q->mask != BIT(q->depth_log2) - 1)
+ return -EINVAL;
+
+ if (q->size < BIT_ULL(q->depth_log2 + q->stride_log2))
+ return -EINVAL;
+
+ return 0;
+}
+
+static u32 ionic_get_eqid(struct ionic_ibdev *dev, u32 comp_vector, u8 udma_idx)
+{
+ /* EQ per vector per udma, and the first eqs reserved for async events.
+ * The rest of the vectors can be requested for completions.
+ */
+ u32 comp_vec_count = dev->lif_cfg.eq_count / dev->lif_cfg.udma_count - 1;
+
+ return (comp_vector % comp_vec_count + 1) * dev->lif_cfg.udma_count + udma_idx;
+}
+
+static int ionic_get_cqid(struct ionic_ibdev *dev, u32 *cqid, u8 udma_idx)
+{
+ unsigned int size, base, bound;
+ int rc;
+
+ size = dev->lif_cfg.cq_count / dev->lif_cfg.udma_count;
+ base = size * udma_idx;
+ bound = base + size;
+
+ rc = ionic_resid_get_shared(&dev->inuse_cqid, base, bound);
+ if (rc >= 0) {
+ /* cq_base is zero or a multiple of two queue groups */
+ *cqid = dev->lif_cfg.cq_base +
+ ionic_bitid_to_qid(rc, dev->lif_cfg.udma_qgrp_shift,
+ dev->half_cqid_udma_shift);
+
+ rc = 0;
+ }
+
+ return rc;
+}
+
+static void ionic_put_cqid(struct ionic_ibdev *dev, u32 cqid)
+{
+ u32 bitid = ionic_qid_to_bitid(cqid - dev->lif_cfg.cq_base,
+ dev->lif_cfg.udma_qgrp_shift,
+ dev->half_cqid_udma_shift);
+
+ ionic_resid_put(&dev->inuse_cqid, bitid);
+}
+
+int ionic_create_cq_common(struct ionic_vcq *vcq,
+ struct ionic_tbl_buf *buf,
+ const struct ib_cq_init_attr *attr,
+ struct ionic_ctx *ctx,
+ struct ib_udata *udata,
+ struct ionic_qdesc *req_cq,
+ __u32 *resp_cqid,
+ int udma_idx)
+{
+ struct ionic_ibdev *dev = to_ionic_ibdev(vcq->ibcq.device);
+ struct ionic_cq *cq = &vcq->cq[udma_idx];
+ void *entry;
+ int rc;
+
+ cq->vcq = vcq;
+
+ if (attr->cqe < 1 || attr->cqe + IONIC_CQ_GRACE > 0xffff) {
+ rc = -EINVAL;
+ goto err_args;
+ }
+
+ rc = ionic_get_cqid(dev, &cq->cqid, udma_idx);
+ if (rc)
+ goto err_cqid;
+
+ cq->eqid = ionic_get_eqid(dev, attr->comp_vector, udma_idx);
+
+ spin_lock_init(&cq->lock);
+ INIT_LIST_HEAD(&cq->poll_sq);
+ INIT_LIST_HEAD(&cq->flush_sq);
+ INIT_LIST_HEAD(&cq->flush_rq);
+
+ if (udata) {
+ rc = ionic_validate_qdesc(req_cq);
+ if (rc)
+ goto err_qdesc;
+
+ cq->umem = ib_umem_get(&dev->ibdev, req_cq->addr, req_cq->size,
+ IB_ACCESS_LOCAL_WRITE);
+ if (IS_ERR(cq->umem)) {
+ rc = PTR_ERR(cq->umem);
+ goto err_umem;
+ }
+
+ cq->q.ptr = NULL;
+ cq->q.size = req_cq->size;
+ cq->q.mask = req_cq->mask;
+ cq->q.depth_log2 = req_cq->depth_log2;
+ cq->q.stride_log2 = req_cq->stride_log2;
+
+ *resp_cqid = cq->cqid;
+ } else {
+ rc = ionic_queue_init(&cq->q, dev->lif_cfg.hwdev,
+ attr->cqe + IONIC_CQ_GRACE,
+ sizeof(struct ionic_v1_cqe));
+ if (rc)
+ goto err_q_init;
+
+ ionic_queue_dbell_init(&cq->q, cq->cqid);
+ cq->color = true;
+ cq->credit = cq->q.mask;
+ }
+
+ rc = ionic_pgtbl_init(dev, buf, cq->umem, cq->q.dma, 1, PAGE_SIZE);
+ if (rc)
+ goto err_pgtbl_init;
+
+ init_completion(&cq->cq_rel_comp);
+ kref_init(&cq->cq_kref);
+
+ entry = xa_store_irq(&dev->cq_tbl, cq->cqid, cq, GFP_KERNEL);
+ if (entry) {
+ if (!xa_is_err(entry))
+ rc = -EINVAL;
+ else
+ rc = xa_err(entry);
+
+ goto err_xa;
+ }
+
+ return 0;
+
+err_xa:
+ ionic_pgtbl_unbuf(dev, buf);
+err_pgtbl_init:
+ if (!udata)
+ ionic_queue_destroy(&cq->q, dev->lif_cfg.hwdev);
+err_q_init:
+ if (cq->umem)
+ ib_umem_release(cq->umem);
+err_umem:
+err_qdesc:
+ ionic_put_cqid(dev, cq->cqid);
+err_cqid:
+err_args:
+ cq->vcq = NULL;
+
+ return rc;
+}
+
+void ionic_destroy_cq_common(struct ionic_ibdev *dev, struct ionic_cq *cq)
+{
+ if (!cq->vcq)
+ return;
+
+ xa_erase_irq(&dev->cq_tbl, cq->cqid);
+ synchronize_rcu();
+
+ kref_put(&cq->cq_kref, ionic_cq_complete);
+ wait_for_completion(&cq->cq_rel_comp);
+
+ if (cq->umem)
+ ib_umem_release(cq->umem);
+ else
+ ionic_queue_destroy(&cq->q, dev->lif_cfg.hwdev);
+
+ ionic_put_cqid(dev, cq->cqid);
+
+ cq->vcq = NULL;
+}
diff --git a/drivers/infiniband/hw/ionic/ionic_fw.h b/drivers/infiniband/hw/ionic/ionic_fw.h
new file mode 100644
index 000000000000..44ec69487519
--- /dev/null
+++ b/drivers/infiniband/hw/ionic/ionic_fw.h
@@ -0,0 +1,164 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright (C) 2018-2025, Advanced Micro Devices, Inc. */
+
+#ifndef _IONIC_FW_H_
+#define _IONIC_FW_H_
+
+#include <linux/kernel.h>
+
+/* completion queue v1 cqe */
+struct ionic_v1_cqe {
+ union {
+ struct {
+ __be16 cmd_idx;
+ __u8 cmd_op;
+ __u8 rsvd[17];
+ __le16 old_sq_cindex;
+ __le16 old_rq_cq_cindex;
+ } admin;
+ struct {
+ __u64 wqe_id;
+ __be32 src_qpn_op;
+ __u8 src_mac[6];
+ __be16 vlan_tag;
+ __be32 imm_data_rkey;
+ } recv;
+ struct {
+ __u8 rsvd[4];
+ __be32 msg_msn;
+ __u8 rsvd2[8];
+ __u64 npg_wqe_id;
+ } send;
+ };
+ __be32 status_length;
+ __be32 qid_type_flags;
+};
+
+/* bits for cqe qid_type_flags */
+enum ionic_v1_cqe_qtf_bits {
+ IONIC_V1_CQE_COLOR = BIT(0),
+ IONIC_V1_CQE_ERROR = BIT(1),
+ IONIC_V1_CQE_TYPE_SHIFT = 5,
+ IONIC_V1_CQE_TYPE_MASK = 0x7,
+ IONIC_V1_CQE_QID_SHIFT = 8,
+
+ IONIC_V1_CQE_TYPE_ADMIN = 0,
+ IONIC_V1_CQE_TYPE_RECV = 1,
+ IONIC_V1_CQE_TYPE_SEND_MSN = 2,
+ IONIC_V1_CQE_TYPE_SEND_NPG = 3,
+};
+
+static inline bool ionic_v1_cqe_color(struct ionic_v1_cqe *cqe)
+{
+ return cqe->qid_type_flags & cpu_to_be32(IONIC_V1_CQE_COLOR);
+}
+
+static inline bool ionic_v1_cqe_error(struct ionic_v1_cqe *cqe)
+{
+ return cqe->qid_type_flags & cpu_to_be32(IONIC_V1_CQE_ERROR);
+}
+
+static inline void ionic_v1_cqe_clean(struct ionic_v1_cqe *cqe)
+{
+ cqe->qid_type_flags |= cpu_to_be32(~0u << IONIC_V1_CQE_QID_SHIFT);
+}
+
+static inline u32 ionic_v1_cqe_qtf(struct ionic_v1_cqe *cqe)
+{
+ return be32_to_cpu(cqe->qid_type_flags);
+}
+
+static inline u8 ionic_v1_cqe_qtf_type(u32 qtf)
+{
+ return (qtf >> IONIC_V1_CQE_TYPE_SHIFT) & IONIC_V1_CQE_TYPE_MASK;
+}
+
+static inline u32 ionic_v1_cqe_qtf_qid(u32 qtf)
+{
+ return qtf >> IONIC_V1_CQE_QID_SHIFT;
+}
+
+#define ADMIN_WQE_STRIDE 64
+#define ADMIN_WQE_HDR_LEN 4
+
+/* admin queue v1 wqe */
+struct ionic_v1_admin_wqe {
+ __u8 op;
+ __u8 rsvd;
+ __le16 len;
+
+ union {
+ } cmd;
+};
+
+/* admin queue v1 cqe status */
+enum ionic_v1_admin_status {
+ IONIC_V1_ASTS_OK,
+ IONIC_V1_ASTS_BAD_CMD,
+ IONIC_V1_ASTS_BAD_INDEX,
+ IONIC_V1_ASTS_BAD_STATE,
+ IONIC_V1_ASTS_BAD_TYPE,
+ IONIC_V1_ASTS_BAD_ATTR,
+ IONIC_V1_ASTS_MSG_TOO_BIG,
+};
+
+/* event queue v1 eqe */
+struct ionic_v1_eqe {
+ __be32 evt;
+};
+
+/* bits for cqe queue_type_flags */
+enum ionic_v1_eqe_evt_bits {
+ IONIC_V1_EQE_COLOR = BIT(0),
+ IONIC_V1_EQE_TYPE_SHIFT = 1,
+ IONIC_V1_EQE_TYPE_MASK = 0x7,
+ IONIC_V1_EQE_CODE_SHIFT = 4,
+ IONIC_V1_EQE_CODE_MASK = 0xf,
+ IONIC_V1_EQE_QID_SHIFT = 8,
+
+ /* cq events */
+ IONIC_V1_EQE_TYPE_CQ = 0,
+ /* cq normal events */
+ IONIC_V1_EQE_CQ_NOTIFY = 0,
+ /* cq error events */
+ IONIC_V1_EQE_CQ_ERR = 8,
+
+ /* qp and srq events */
+ IONIC_V1_EQE_TYPE_QP = 1,
+ /* qp normal events */
+ IONIC_V1_EQE_SRQ_LEVEL = 0,
+ IONIC_V1_EQE_SQ_DRAIN = 1,
+ IONIC_V1_EQE_QP_COMM_EST = 2,
+ IONIC_V1_EQE_QP_LAST_WQE = 3,
+ /* qp error events */
+ IONIC_V1_EQE_QP_ERR = 8,
+ IONIC_V1_EQE_QP_ERR_REQUEST = 9,
+ IONIC_V1_EQE_QP_ERR_ACCESS = 10,
+};
+
+static inline bool ionic_v1_eqe_color(struct ionic_v1_eqe *eqe)
+{
+ return eqe->evt & cpu_to_be32(IONIC_V1_EQE_COLOR);
+}
+
+static inline u32 ionic_v1_eqe_evt(struct ionic_v1_eqe *eqe)
+{
+ return be32_to_cpu(eqe->evt);
+}
+
+static inline u8 ionic_v1_eqe_evt_type(u32 evt)
+{
+ return (evt >> IONIC_V1_EQE_TYPE_SHIFT) & IONIC_V1_EQE_TYPE_MASK;
+}
+
+static inline u8 ionic_v1_eqe_evt_code(u32 evt)
+{
+ return (evt >> IONIC_V1_EQE_CODE_SHIFT) & IONIC_V1_EQE_CODE_MASK;
+}
+
+static inline u32 ionic_v1_eqe_evt_qid(u32 evt)
+{
+ return evt >> IONIC_V1_EQE_QID_SHIFT;
+}
+
+#endif /* _IONIC_FW_H_ */
diff --git a/drivers/infiniband/hw/ionic/ionic_ibdev.c b/drivers/infiniband/hw/ionic/ionic_ibdev.c
index 6bd0175cd7b1..34febbeb08a5 100644
--- a/drivers/infiniband/hw/ionic/ionic_ibdev.c
+++ b/drivers/infiniband/hw/ionic/ionic_ibdev.c
@@ -15,9 +15,41 @@ MODULE_DESCRIPTION(DRIVER_DESCRIPTION);
MODULE_LICENSE("GPL");
MODULE_IMPORT_NS("NET_IONIC");
+static void ionic_init_resids(struct ionic_ibdev *dev)
+{
+ ionic_resid_init(&dev->inuse_cqid, dev->lif_cfg.cq_count);
+ dev->half_cqid_udma_shift =
+ order_base_2(dev->lif_cfg.cq_count / dev->lif_cfg.udma_count);
+ ionic_resid_init(&dev->inuse_pdid, IONIC_MAX_PD);
+ ionic_resid_init(&dev->inuse_ahid, dev->lif_cfg.nahs_per_lif);
+ ionic_resid_init(&dev->inuse_mrid, dev->lif_cfg.nmrs_per_lif);
+ /* skip reserved lkey */
+ dev->next_mrkey = 1;
+ ionic_resid_init(&dev->inuse_qpid, dev->lif_cfg.qp_count);
+ /* skip reserved SMI and GSI qpids */
+ dev->half_qpid_udma_shift =
+ order_base_2(dev->lif_cfg.qp_count / dev->lif_cfg.udma_count);
+ ionic_resid_init(&dev->inuse_dbid, dev->lif_cfg.dbid_count);
+}
+
+static void ionic_destroy_resids(struct ionic_ibdev *dev)
+{
+ ionic_resid_destroy(&dev->inuse_cqid);
+ ionic_resid_destroy(&dev->inuse_pdid);
+ ionic_resid_destroy(&dev->inuse_ahid);
+ ionic_resid_destroy(&dev->inuse_mrid);
+ ionic_resid_destroy(&dev->inuse_qpid);
+ ionic_resid_destroy(&dev->inuse_dbid);
+}
+
static void ionic_destroy_ibdev(struct ionic_ibdev *dev)
{
+ ionic_kill_rdma_admin(dev, false);
ib_unregister_device(&dev->ibdev);
+ ionic_destroy_rdma_admin(dev);
+ ionic_destroy_resids(dev);
+ WARN_ON(!xa_empty(&dev->cq_tbl));
+ xa_destroy(&dev->cq_tbl);
ib_dealloc_device(&dev->ibdev);
}
@@ -37,6 +69,18 @@ static struct ionic_ibdev *ionic_create_ibdev(struct ionic_aux_dev *ionic_adev)
ionic_fill_lif_cfg(ionic_adev->lif, &dev->lif_cfg);
+ xa_init_flags(&dev->cq_tbl, GFP_ATOMIC);
+
+ ionic_init_resids(dev);
+
+ rc = ionic_rdma_reset_devcmd(dev);
+ if (rc)
+ goto err_reset;
+
+ rc = ionic_create_rdma_admin(dev);
+ if (rc)
+ goto err_admin;
+
ibdev = &dev->ibdev;
ibdev->dev.parent = dev->lif_cfg.hwdev;
@@ -64,6 +108,11 @@ static struct ionic_ibdev *ionic_create_ibdev(struct ionic_aux_dev *ionic_adev)
err_register:
err_admin:
+ ionic_kill_rdma_admin(dev, false);
+ ionic_destroy_rdma_admin(dev);
+err_reset:
+ ionic_destroy_resids(dev);
+ xa_destroy(&dev->cq_tbl);
ib_dealloc_device(&dev->ibdev);
return ERR_PTR(rc);
@@ -114,6 +163,10 @@ static int __init ionic_mod_init(void)
{
int rc;
+ ionic_evt_workq = create_workqueue(DRIVER_NAME "-evt");
+ if (!ionic_evt_workq)
+ return -ENOMEM;
+
rc = auxiliary_driver_register(&ionic_aux_r_driver);
if (rc)
goto err_aux;
@@ -121,12 +174,15 @@ static int __init ionic_mod_init(void)
return 0;
err_aux:
+ destroy_workqueue(ionic_evt_workq);
+
return rc;
}
static void __exit ionic_mod_exit(void)
{
auxiliary_driver_unregister(&ionic_aux_r_driver);
+ destroy_workqueue(ionic_evt_workq);
}
module_init(ionic_mod_init);
diff --git a/drivers/infiniband/hw/ionic/ionic_ibdev.h b/drivers/infiniband/hw/ionic/ionic_ibdev.h
index e13adff390d7..e7563c0429fc 100644
--- a/drivers/infiniband/hw/ionic/ionic_ibdev.h
+++ b/drivers/infiniband/hw/ionic/ionic_ibdev.h
@@ -4,18 +4,243 @@
#ifndef _IONIC_IBDEV_H_
#define _IONIC_IBDEV_H_
+#include <rdma/ib_umem.h>
#include <rdma/ib_verbs.h>
+
#include <ionic_api.h>
+#include <ionic_regs.h>
+
+#include "ionic_fw.h"
+#include "ionic_queue.h"
+#include "ionic_res.h"
#include "ionic_lif_cfg.h"
+#define DRIVER_NAME "ionic_rdma"
+#define DRIVER_SHORTNAME "ionr"
+
#define IONIC_MIN_RDMA_VERSION 0
#define IONIC_MAX_RDMA_VERSION 2
+/* Config knobs */
+#define IONIC_EQ_DEPTH 511
+#define IONIC_EQ_COUNT 32
+#define IONIC_AQ_DEPTH 63
+#define IONIC_AQ_COUNT 4
+#define IONIC_EQ_ISR_BUDGET 10
+#define IONIC_EQ_WORK_BUDGET 1000
+#define IONIC_MAX_PD 1024
+
+#define IONIC_CQ_GRACE 100
+
+struct ionic_aq;
+struct ionic_cq;
+struct ionic_eq;
+struct ionic_vcq;
+
+enum ionic_admin_state {
+ IONIC_ADMIN_ACTIVE, /* submitting admin commands to queue */
+ IONIC_ADMIN_PAUSED, /* not submitting, but may complete normally */
+ IONIC_ADMIN_KILLED, /* not submitting, locally completed */
+};
+
+enum ionic_admin_flags {
+ IONIC_ADMIN_F_BUSYWAIT = BIT(0), /* Don't sleep */
+ IONIC_ADMIN_F_TEARDOWN = BIT(1), /* In destroy path */
+ IONIC_ADMIN_F_INTERRUPT = BIT(2), /* Interruptible w/timeout */
+};
+
+struct ionic_qdesc {
+ __aligned_u64 addr;
+ __u32 size;
+ __u16 mask;
+ __u8 depth_log2;
+ __u8 stride_log2;
+};
+
+enum ionic_mmap_flag {
+ IONIC_MMAP_WC = BIT(0),
+};
+
+struct ionic_mmap_entry {
+ struct rdma_user_mmap_entry rdma_entry;
+ unsigned long size;
+ unsigned long pfn;
+ u8 mmap_flags;
+};
+
struct ionic_ibdev {
struct ib_device ibdev;
struct ionic_lif_cfg lif_cfg;
+
+ struct xarray qp_tbl;
+ struct xarray cq_tbl;
+
+ struct ionic_resid_bits inuse_dbid;
+ struct ionic_resid_bits inuse_pdid;
+ struct ionic_resid_bits inuse_ahid;
+ struct ionic_resid_bits inuse_mrid;
+ struct ionic_resid_bits inuse_qpid;
+ struct ionic_resid_bits inuse_cqid;
+
+ u8 half_cqid_udma_shift;
+ u8 half_qpid_udma_shift;
+ u8 next_qpid_udma_idx;
+ u8 next_mrkey;
+
+ struct work_struct reset_work;
+ bool reset_posted;
+ u32 reset_cnt;
+
+ struct delayed_work admin_dwork;
+ struct ionic_aq **aq_vec;
+ atomic_t admin_state;
+
+ struct ionic_eq **eq_vec;
};
+struct ionic_eq {
+ struct ionic_ibdev *dev;
+
+ u32 eqid;
+ u32 intr;
+
+ struct ionic_queue q;
+
+ bool armed;
+ bool enable;
+
+ struct work_struct work;
+
+ int irq;
+ char name[32];
+};
+
+struct ionic_admin_wr {
+ struct completion work;
+ struct list_head aq_ent;
+ struct ionic_v1_admin_wqe wqe;
+ struct ionic_v1_cqe cqe;
+ struct ionic_aq *aq;
+ int status;
+};
+
+struct ionic_admin_wr_q {
+ struct ionic_admin_wr *wr;
+ int wqe_strides;
+};
+
+struct ionic_aq {
+ struct ionic_ibdev *dev;
+ struct ionic_vcq *vcq;
+
+ struct work_struct work;
+
+ enum ionic_admin_state admin_state;
+ unsigned long stamp;
+ bool armed;
+
+ u32 aqid;
+ u32 cqid;
+
+ spinlock_t lock; /* for posting */
+ struct ionic_queue q;
+ struct ionic_admin_wr_q *q_wr;
+ struct list_head wr_prod;
+ struct list_head wr_post;
+};
+
+struct ionic_ctx {
+ struct ib_ucontext ibctx;
+ u32 dbid;
+ struct rdma_user_mmap_entry *mmap_dbell;
+};
+
+struct ionic_tbl_buf {
+ u32 tbl_limit;
+ u32 tbl_pages;
+ size_t tbl_size;
+ __le64 *tbl_buf;
+ dma_addr_t tbl_dma;
+ u8 page_size_log2;
+};
+
+struct ionic_cq {
+ struct ionic_vcq *vcq;
+
+ u32 cqid;
+ u32 eqid;
+
+ spinlock_t lock; /* for polling */
+ struct list_head poll_sq;
+ bool flush;
+ struct list_head flush_sq;
+ struct list_head flush_rq;
+ struct list_head ibkill_flush_ent;
+
+ struct ionic_queue q;
+ bool color;
+ int credit;
+ u16 arm_any_prod;
+ u16 arm_sol_prod;
+
+ struct kref cq_kref;
+ struct completion cq_rel_comp;
+
+ /* infrequently accessed, keep at end */
+ struct ib_umem *umem;
+};
+
+struct ionic_vcq {
+ struct ib_cq ibcq;
+ struct ionic_cq cq[2];
+ u8 udma_mask;
+ u8 poll_idx;
+};
+
+static inline struct ionic_ibdev *to_ionic_ibdev(struct ib_device *ibdev)
+{
+ return container_of(ibdev, struct ionic_ibdev, ibdev);
+}
+
+static inline void ionic_cq_complete(struct kref *kref)
+{
+ struct ionic_cq *cq = container_of(kref, struct ionic_cq, cq_kref);
+
+ complete(&cq->cq_rel_comp);
+}
+
+/* ionic_admin.c */
+extern struct workqueue_struct *ionic_evt_workq;
+void ionic_admin_post(struct ionic_ibdev *dev, struct ionic_admin_wr *wr);
+int ionic_admin_wait(struct ionic_ibdev *dev, struct ionic_admin_wr *wr,
+ enum ionic_admin_flags);
+
+int ionic_rdma_reset_devcmd(struct ionic_ibdev *dev);
+
+int ionic_create_rdma_admin(struct ionic_ibdev *dev);
+void ionic_destroy_rdma_admin(struct ionic_ibdev *dev);
+void ionic_kill_rdma_admin(struct ionic_ibdev *dev, bool fatal_path);
+
+/* ionic_controlpath.c */
+int ionic_create_cq_common(struct ionic_vcq *vcq,
+ struct ionic_tbl_buf *buf,
+ const struct ib_cq_init_attr *attr,
+ struct ionic_ctx *ctx,
+ struct ib_udata *udata,
+ struct ionic_qdesc *req_cq,
+ __u32 *resp_cqid,
+ int udma_idx);
+void ionic_destroy_cq_common(struct ionic_ibdev *dev, struct ionic_cq *cq);
+
+/* ionic_pgtbl.c */
+int ionic_pgtbl_page(struct ionic_tbl_buf *buf, u64 dma);
+int ionic_pgtbl_init(struct ionic_ibdev *dev,
+ struct ionic_tbl_buf *buf,
+ struct ib_umem *umem,
+ dma_addr_t dma,
+ int limit,
+ u64 page_size);
+void ionic_pgtbl_unbuf(struct ionic_ibdev *dev, struct ionic_tbl_buf *buf);
#endif /* _IONIC_IBDEV_H_ */
diff --git a/drivers/infiniband/hw/ionic/ionic_pgtbl.c b/drivers/infiniband/hw/ionic/ionic_pgtbl.c
new file mode 100644
index 000000000000..11461f7642bc
--- /dev/null
+++ b/drivers/infiniband/hw/ionic/ionic_pgtbl.c
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (C) 2018-2025, Advanced Micro Devices, Inc. */
+
+#include <linux/mman.h>
+#include <linux/dma-mapping.h>
+
+#include "ionic_fw.h"
+#include "ionic_ibdev.h"
+
+int ionic_pgtbl_page(struct ionic_tbl_buf *buf, u64 dma)
+{
+ if (unlikely(buf->tbl_pages == buf->tbl_limit))
+ return -ENOMEM;
+
+ if (buf->tbl_buf)
+ buf->tbl_buf[buf->tbl_pages] = cpu_to_le64(dma);
+ else
+ buf->tbl_dma = dma;
+
+ ++buf->tbl_pages;
+
+ return 0;
+}
+
+static int ionic_tbl_buf_alloc(struct ionic_ibdev *dev,
+ struct ionic_tbl_buf *buf)
+{
+ int rc;
+
+ buf->tbl_size = buf->tbl_limit * sizeof(*buf->tbl_buf);
+ buf->tbl_buf = kmalloc(buf->tbl_size, GFP_KERNEL);
+ if (!buf->tbl_buf)
+ return -ENOMEM;
+
+ buf->tbl_dma = dma_map_single(dev->lif_cfg.hwdev, buf->tbl_buf,
+ buf->tbl_size, DMA_TO_DEVICE);
+ rc = dma_mapping_error(dev->lif_cfg.hwdev, buf->tbl_dma);
+ if (rc) {
+ kfree(buf->tbl_buf);
+ return rc;
+ }
+
+ return 0;
+}
+
+static int ionic_pgtbl_umem(struct ionic_tbl_buf *buf, struct ib_umem *umem)
+{
+ struct ib_block_iter biter;
+ u64 page_dma;
+ int rc;
+
+ rdma_umem_for_each_dma_block(umem, &biter, BIT_ULL(buf->page_size_log2)) {
+ page_dma = rdma_block_iter_dma_address(&biter);
+ rc = ionic_pgtbl_page(buf, page_dma);
+ if (rc)
+ return rc;
+ }
+
+ return 0;
+}
+
+void ionic_pgtbl_unbuf(struct ionic_ibdev *dev, struct ionic_tbl_buf *buf)
+{
+ if (buf->tbl_buf)
+ dma_unmap_single(dev->lif_cfg.hwdev, buf->tbl_dma,
+ buf->tbl_size, DMA_TO_DEVICE);
+
+ kfree(buf->tbl_buf);
+ memset(buf, 0, sizeof(*buf));
+}
+
+int ionic_pgtbl_init(struct ionic_ibdev *dev,
+ struct ionic_tbl_buf *buf,
+ struct ib_umem *umem,
+ dma_addr_t dma,
+ int limit,
+ u64 page_size)
+{
+ int rc;
+
+ memset(buf, 0, sizeof(*buf));
+
+ if (umem) {
+ limit = ib_umem_num_dma_blocks(umem, page_size);
+ buf->page_size_log2 = order_base_2(page_size);
+ }
+
+ if (limit < 1)
+ return -EINVAL;
+
+ buf->tbl_limit = limit;
+
+ /* skip pgtbl if contiguous / direct translation */
+ if (limit > 1) {
+ rc = ionic_tbl_buf_alloc(dev, buf);
+ if (rc)
+ return rc;
+ }
+
+ if (umem)
+ rc = ionic_pgtbl_umem(buf, umem);
+ else
+ rc = ionic_pgtbl_page(buf, dma);
+
+ if (rc)
+ goto err_unbuf;
+
+ return 0;
+
+err_unbuf:
+ ionic_pgtbl_unbuf(dev, buf);
+ return rc;
+}
diff --git a/drivers/infiniband/hw/ionic/ionic_queue.c b/drivers/infiniband/hw/ionic/ionic_queue.c
new file mode 100644
index 000000000000..aa897ed2a412
--- /dev/null
+++ b/drivers/infiniband/hw/ionic/ionic_queue.c
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (C) 2018-2025, Advanced Micro Devices, Inc. */
+
+#include <linux/dma-mapping.h>
+
+#include "ionic_queue.h"
+
+int ionic_queue_init(struct ionic_queue *q, struct device *dma_dev,
+ int depth, size_t stride)
+{
+ if (depth < 0 || depth > 0xffff)
+ return -EINVAL;
+
+ if (stride == 0 || stride > 0x10000)
+ return -EINVAL;
+
+ if (depth == 0)
+ depth = 1;
+
+ q->depth_log2 = order_base_2(depth + 1);
+ q->stride_log2 = order_base_2(stride);
+
+ if (q->depth_log2 + q->stride_log2 < PAGE_SHIFT)
+ q->depth_log2 = PAGE_SHIFT - q->stride_log2;
+
+ if (q->depth_log2 > 16 || q->stride_log2 > 16)
+ return -EINVAL;
+
+ q->size = BIT_ULL(q->depth_log2 + q->stride_log2);
+ q->mask = BIT(q->depth_log2) - 1;
+
+ q->ptr = dma_alloc_coherent(dma_dev, q->size, &q->dma, GFP_KERNEL);
+ if (!q->ptr)
+ return -ENOMEM;
+
+ /* it will always be page aligned, but just to be sure... */
+ if (!PAGE_ALIGNED(q->ptr)) {
+ dma_free_coherent(dma_dev, q->size, q->ptr, q->dma);
+ return -ENOMEM;
+ }
+
+ q->prod = 0;
+ q->cons = 0;
+ q->dbell = 0;
+
+ return 0;
+}
+
+void ionic_queue_destroy(struct ionic_queue *q, struct device *dma_dev)
+{
+ dma_free_coherent(dma_dev, q->size, q->ptr, q->dma);
+}
diff --git a/drivers/infiniband/hw/ionic/ionic_queue.h b/drivers/infiniband/hw/ionic/ionic_queue.h
new file mode 100644
index 000000000000..d18020d4cad5
--- /dev/null
+++ b/drivers/infiniband/hw/ionic/ionic_queue.h
@@ -0,0 +1,234 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright (C) 2018-2025, Advanced Micro Devices, Inc. */
+
+#ifndef _IONIC_QUEUE_H_
+#define _IONIC_QUEUE_H_
+
+#include <linux/io.h>
+#include <ionic_regs.h>
+
+#define IONIC_MAX_DEPTH 0xffff
+#define IONIC_MAX_CQ_DEPTH 0xffff
+#define IONIC_CQ_RING_ARM IONIC_DBELL_RING_1
+#define IONIC_CQ_RING_SOL IONIC_DBELL_RING_2
+
+/**
+ * struct ionic_queue - Ring buffer used between device and driver
+ * @size: Size of the buffer, in bytes
+ * @dma: Dma address of the buffer
+ * @ptr: Buffer virtual address
+ * @prod: Driver position in the queue
+ * @cons: Device position in the queue
+ * @mask: Capacity of the queue, subtracting the hole
+ * This value is equal to ((1 << depth_log2) - 1)
+ * @depth_log2: Log base two size depth of the queue
+ * @stride_log2: Log base two size of an element in the queue
+ * @dbell: Doorbell identifying bits
+ */
+struct ionic_queue {
+ size_t size;
+ dma_addr_t dma;
+ void *ptr;
+ u16 prod;
+ u16 cons;
+ u16 mask;
+ u8 depth_log2;
+ u8 stride_log2;
+ u64 dbell;
+};
+
+/**
+ * ionic_queue_init() - Initialize user space queue
+ * @q: Uninitialized queue structure
+ * @dma_dev: DMA device for mapping
+ * @depth: Depth of the queue
+ * @stride: Size of each element of the queue
+ *
+ * Return: status code
+ */
+int ionic_queue_init(struct ionic_queue *q, struct device *dma_dev,
+ int depth, size_t stride);
+
+/**
+ * ionic_queue_destroy() - Destroy user space queue
+ * @q: Queue structure
+ * @dma_dev: DMA device for mapping
+ *
+ * Return: status code
+ */
+void ionic_queue_destroy(struct ionic_queue *q, struct device *dma_dev);
+
+/**
+ * ionic_queue_empty() - Test if queue is empty
+ * @q: Queue structure
+ *
+ * This is only valid for to-device queues.
+ *
+ * Return: is empty
+ */
+static inline bool ionic_queue_empty(struct ionic_queue *q)
+{
+ return q->prod == q->cons;
+}
+
+/**
+ * ionic_queue_length() - Get the current length of the queue
+ * @q: Queue structure
+ *
+ * This is only valid for to-device queues.
+ *
+ * Return: length
+ */
+static inline u16 ionic_queue_length(struct ionic_queue *q)
+{
+ return (q->prod - q->cons) & q->mask;
+}
+
+/**
+ * ionic_queue_length_remaining() - Get the remaining length of the queue
+ * @q: Queue structure
+ *
+ * This is only valid for to-device queues.
+ *
+ * Return: length remaining
+ */
+static inline u16 ionic_queue_length_remaining(struct ionic_queue *q)
+{
+ return q->mask - ionic_queue_length(q);
+}
+
+/**
+ * ionic_queue_full() - Test if queue is full
+ * @q: Queue structure
+ *
+ * This is only valid for to-device queues.
+ *
+ * Return: is full
+ */
+static inline bool ionic_queue_full(struct ionic_queue *q)
+{
+ return q->mask == ionic_queue_length(q);
+}
+
+/**
+ * ionic_color_wrap() - Flip the color if prod is wrapped
+ * @prod: Queue index just after advancing
+ * @color: Queue color just prior to advancing the index
+ *
+ * Return: color after advancing the index
+ */
+static inline bool ionic_color_wrap(u16 prod, bool color)
+{
+ /* logical xor color with (prod == 0) */
+ return color != (prod == 0);
+}
+
+/**
+ * ionic_queue_at() - Get the element at the given index
+ * @q: Queue structure
+ * @idx: Index in the queue
+ *
+ * The index must be within the bounds of the queue. It is not checked here.
+ *
+ * Return: pointer to element at index
+ */
+static inline void *ionic_queue_at(struct ionic_queue *q, u16 idx)
+{
+ return q->ptr + ((unsigned long)idx << q->stride_log2);
+}
+
+/**
+ * ionic_queue_at_prod() - Get the element at the producer index
+ * @q: Queue structure
+ *
+ * Return: pointer to element at producer index
+ */
+static inline void *ionic_queue_at_prod(struct ionic_queue *q)
+{
+ return ionic_queue_at(q, q->prod);
+}
+
+/**
+ * ionic_queue_at_cons() - Get the element at the consumer index
+ * @q: Queue structure
+ *
+ * Return: pointer to element at consumer index
+ */
+static inline void *ionic_queue_at_cons(struct ionic_queue *q)
+{
+ return ionic_queue_at(q, q->cons);
+}
+
+/**
+ * ionic_queue_next() - Compute the next index
+ * @q: Queue structure
+ * @idx: Index
+ *
+ * Return: next index after idx
+ */
+static inline u16 ionic_queue_next(struct ionic_queue *q, u16 idx)
+{
+ return (idx + 1) & q->mask;
+}
+
+/**
+ * ionic_queue_produce() - Increase the producer index
+ * @q: Queue structure
+ *
+ * Caller must ensure that the queue is not full. It is not checked here.
+ */
+static inline void ionic_queue_produce(struct ionic_queue *q)
+{
+ q->prod = ionic_queue_next(q, q->prod);
+}
+
+/**
+ * ionic_queue_consume() - Increase the consumer index
+ * @q: Queue structure
+ *
+ * Caller must ensure that the queue is not empty. It is not checked here.
+ *
+ * This is only valid for to-device queues.
+ */
+static inline void ionic_queue_consume(struct ionic_queue *q)
+{
+ q->cons = ionic_queue_next(q, q->cons);
+}
+
+/**
+ * ionic_queue_consume_entries() - Increase the consumer index by entries
+ * @q: Queue structure
+ * @entries: Number of entries to increment
+ *
+ * Caller must ensure that the queue is not empty. It is not checked here.
+ *
+ * This is only valid for to-device queues.
+ */
+static inline void ionic_queue_consume_entries(struct ionic_queue *q,
+ u16 entries)
+{
+ q->cons = (q->cons + entries) & q->mask;
+}
+
+/**
+ * ionic_queue_dbell_init() - Initialize doorbell bits for queue id
+ * @q: Queue structure
+ * @qid: Queue identifying number
+ */
+static inline void ionic_queue_dbell_init(struct ionic_queue *q, u32 qid)
+{
+ q->dbell = IONIC_DBELL_QID(qid);
+}
+
+/**
+ * ionic_queue_dbell_val() - Get current doorbell update value
+ * @q: Queue structure
+ *
+ * Return: current doorbell update value
+ */
+static inline u64 ionic_queue_dbell_val(struct ionic_queue *q)
+{
+ return q->dbell | q->prod;
+}
+
+#endif /* _IONIC_QUEUE_H_ */
diff --git a/drivers/infiniband/hw/ionic/ionic_res.h b/drivers/infiniband/hw/ionic/ionic_res.h
new file mode 100644
index 000000000000..46c8c584bd9a
--- /dev/null
+++ b/drivers/infiniband/hw/ionic/ionic_res.h
@@ -0,0 +1,154 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright (C) 2018-2025, Advanced Micro Devices, Inc. */
+
+#ifndef _IONIC_RES_H_
+#define _IONIC_RES_H_
+
+#include <linux/kernel.h>
+#include <linux/idr.h>
+
+/**
+ * struct ionic_resid_bits - Number allocator based on IDA
+ *
+ * @inuse: IDA handle
+ * @inuse_size: Highest ID limit for IDA
+ */
+struct ionic_resid_bits {
+ struct ida inuse;
+ unsigned int inuse_size;
+};
+
+/**
+ * ionic_resid_init() - Initialize a resid allocator
+ * @resid: Uninitialized resid allocator
+ * @size: Capacity of the allocator
+ *
+ * Return: Zero on success, or negative error number
+ */
+static inline void ionic_resid_init(struct ionic_resid_bits *resid,
+ unsigned int size)
+{
+ resid->inuse_size = size;
+ ida_init(&resid->inuse);
+}
+
+/**
+ * ionic_resid_destroy() - Destroy a resid allocator
+ * @resid: Resid allocator
+ */
+static inline void ionic_resid_destroy(struct ionic_resid_bits *resid)
+{
+ ida_destroy(&resid->inuse);
+}
+
+/**
+ * ionic_resid_get_shared() - Allocate an available shared resource id
+ * @resid: Resid allocator
+ * @min: Smallest valid resource id
+ * @size: One after largest valid resource id
+ *
+ * Return: Resource id, or negative error number
+ */
+static inline int ionic_resid_get_shared(struct ionic_resid_bits *resid,
+ unsigned int min,
+ unsigned int size)
+{
+ return ida_alloc_range(&resid->inuse, min, size - 1, GFP_KERNEL);
+}
+
+/**
+ * ionic_resid_get() - Allocate an available resource id
+ * @resid: Resid allocator
+ *
+ * Return: Resource id, or negative error number
+ */
+static inline int ionic_resid_get(struct ionic_resid_bits *resid)
+{
+ return ionic_resid_get_shared(resid, 0, resid->inuse_size);
+}
+
+/**
+ * ionic_resid_put() - Free a resource id
+ * @resid: Resid allocator
+ * @id: Resource id
+ */
+static inline void ionic_resid_put(struct ionic_resid_bits *resid, int id)
+{
+ ida_free(&resid->inuse, id);
+}
+
+/**
+ * ionic_bitid_to_qid() - Transform a resource bit index into a queue id
+ * @bitid: Bit index
+ * @qgrp_shift: Log2 number of queues per queue group
+ * @half_qid_shift: Log2 of half the total number of queues
+ *
+ * Return: Queue id
+ *
+ * Udma-constrained queues (QPs and CQs) are associated with their udma by
+ * queue group. Even queue groups are associated with udma0, and odd queue
+ * groups with udma1.
+ *
+ * For allocating queue ids, we want to arrange the bits into two halves,
+ * with the even queue groups of udma0 in the lower half of the bitset,
+ * and the odd queue groups of udma1 in the upper half of the bitset.
+ * Then, one or two calls of find_next_zero_bit can examine all the bits
+ * for queues of an entire udma.
+ *
+ * For example, assuming eight queue groups with qgrp qids per group:
+ *
+ * bitid 0*qgrp..1*qgrp-1 : qid 0*qgrp..1*qgrp-1
+ * bitid 1*qgrp..2*qgrp-1 : qid 2*qgrp..3*qgrp-1
+ * bitid 2*qgrp..3*qgrp-1 : qid 4*qgrp..5*qgrp-1
+ * bitid 3*qgrp..4*qgrp-1 : qid 6*qgrp..7*qgrp-1
+ * bitid 4*qgrp..5*qgrp-1 : qid 1*qgrp..2*qgrp-1
+ * bitid 5*qgrp..6*qgrp-1 : qid 3*qgrp..4*qgrp-1
+ * bitid 6*qgrp..7*qgrp-1 : qid 5*qgrp..6*qgrp-1
+ * bitid 7*qgrp..8*qgrp-1 : qid 7*qgrp..8*qgrp-1
+ *
+ * There are three important ranges of bits in the qid. There is the udma
+ * bit "U" at qgrp_shift, which is the least significant bit of the group
+ * index, and determines which udma a queue is associated with.
+ * The bits of lesser significance we can call the idx bits "I", which are
+ * the index of the queue within the group. The bits of greater significance
+ * we can call the grp bits "G", which are other bits of the group index that
+ * do not determine the udma. Those bits are just rearranged in the bit index
+ * in the bitset. A bitid has the udma bit in the most significant place,
+ * then the grp bits, then the idx bits.
+ *
+ * bitid: 00000000000000 U GGG IIIIII
+ * qid: 00000000000000 GGG U IIIIII
+ *
+ * Transforming from bit index to qid, or from qid to bit index, can be
+ * accomplished by rearranging the bits by masking and shifting.
+ */
+static inline u32 ionic_bitid_to_qid(u32 bitid, u8 qgrp_shift,
+ u8 half_qid_shift)
+{
+ u32 udma_bit =
+ (bitid & BIT(half_qid_shift)) >> (half_qid_shift - qgrp_shift);
+ u32 grp_bits = (bitid & GENMASK(half_qid_shift - 1, qgrp_shift)) << 1;
+ u32 idx_bits = bitid & (BIT(qgrp_shift) - 1);
+
+ return grp_bits | udma_bit | idx_bits;
+}
+
+/**
+ * ionic_qid_to_bitid() - Transform a queue id into a resource bit index
+ * @qid: queue index
+ * @qgrp_shift: Log2 number of queues per queue group
+ * @half_qid_shift: Log2 of half the total number of queues
+ *
+ * Return: Resource bit index
+ *
+ * This is the inverse of ionic_bitid_to_qid().
+ */
+static inline u32 ionic_qid_to_bitid(u32 qid, u8 qgrp_shift, u8 half_qid_shift)
+{
+ u32 udma_bit = (qid & BIT(qgrp_shift)) << (half_qid_shift - qgrp_shift);
+ u32 grp_bits = (qid & GENMASK(half_qid_shift, qgrp_shift + 1)) >> 1;
+ u32 idx_bits = qid & (BIT(qgrp_shift) - 1);
+
+ return udma_bit | grp_bits | idx_bits;
+}
+#endif /* _IONIC_RES_H_ */
--
2.43.0
On Tue, Jun 24, 2025 at 05:43:10PM +0530, Abhijit Gangurde wrote: > Setup RDMA admin queues using device command exposed over > auxiliary device and manage these queues using ida. > > Co-developed-by: Andrew Boyer <andrew.boyer@amd.com> > Signed-off-by: Andrew Boyer <andrew.boyer@amd.com> > Co-developed-by: Allen Hubbe <allen.hubbe@amd.com> > Signed-off-by: Allen Hubbe <allen.hubbe@amd.com> > Signed-off-by: Abhijit Gangurde <abhijit.gangurde@amd.com> > --- > v2->v3 > - Fixed lockdep warning > - Used IDA for resource id allocation > - Removed rw locks around xarrays > > drivers/infiniband/hw/ionic/ionic_admin.c | 1169 +++++++++++++++++ > .../infiniband/hw/ionic/ionic_controlpath.c | 184 +++ > drivers/infiniband/hw/ionic/ionic_fw.h | 164 +++ > drivers/infiniband/hw/ionic/ionic_ibdev.c | 56 + > drivers/infiniband/hw/ionic/ionic_ibdev.h | 225 ++++ > drivers/infiniband/hw/ionic/ionic_pgtbl.c | 113 ++ > drivers/infiniband/hw/ionic/ionic_queue.c | 52 + > drivers/infiniband/hw/ionic/ionic_queue.h | 234 ++++ > drivers/infiniband/hw/ionic/ionic_res.h | 154 +++ > 9 files changed, 2351 insertions(+) > create mode 100644 drivers/infiniband/hw/ionic/ionic_admin.c > create mode 100644 drivers/infiniband/hw/ionic/ionic_controlpath.c > create mode 100644 drivers/infiniband/hw/ionic/ionic_fw.h > create mode 100644 drivers/infiniband/hw/ionic/ionic_pgtbl.c > create mode 100644 drivers/infiniband/hw/ionic/ionic_queue.c > create mode 100644 drivers/infiniband/hw/ionic/ionic_queue.h > create mode 100644 drivers/infiniband/hw/ionic/ionic_res.h <...> > +static void ionic_admin_timedout(struct ionic_aq *aq) > +{ > + struct ionic_cq *cq = &aq->vcq->cq[0]; > + struct ionic_ibdev *dev = aq->dev; > + unsigned long irqflags; > + u16 pos; > + > + spin_lock_irqsave(&aq->lock, irqflags); > + if (ionic_queue_empty(&aq->q)) > + goto out; > + > + /* Reset ALL adminq if any one times out */ > + if (aq->admin_state < IONIC_ADMIN_KILLED) > + queue_work(ionic_evt_workq, &dev->reset_work); > + > + ibdev_err(&dev->ibdev, "admin command timed out, aq %d\n", aq->aqid); > + > + ibdev_warn(&dev->ibdev, "admin timeout was set for %ums\n", > + (u32)jiffies_to_msecs(IONIC_ADMIN_TIMEOUT)); > + ibdev_warn(&dev->ibdev, "admin inactivity for %ums\n", > + (u32)jiffies_to_msecs(jiffies - aq->stamp)); > + > + ibdev_warn(&dev->ibdev, "admin commands outstanding %u\n", > + ionic_queue_length(&aq->q)); > + ibdev_warn(&dev->ibdev, "%s more commands pending\n", > + list_empty(&aq->wr_post) ? "no" : "some"); > + > + pos = cq->q.prod; > + > + ibdev_warn(&dev->ibdev, "admin cq pos %u (next to complete)\n", pos); > + print_hex_dump(KERN_WARNING, "cqe ", DUMP_PREFIX_OFFSET, 16, 1, > + ionic_queue_at(&cq->q, pos), > + BIT(cq->q.stride_log2), true); > + > + pos = (pos - 1) & cq->q.mask; > + > + ibdev_warn(&dev->ibdev, "admin cq pos %u (last completed)\n", pos); > + print_hex_dump(KERN_WARNING, "cqe ", DUMP_PREFIX_OFFSET, 16, 1, > + ionic_queue_at(&cq->q, pos), > + BIT(cq->q.stride_log2), true); > + > + pos = aq->q.cons; > + > + ibdev_warn(&dev->ibdev, "admin pos %u (next to complete)\n", pos); > + print_hex_dump(KERN_WARNING, "cmd ", DUMP_PREFIX_OFFSET, 16, 1, > + ionic_queue_at(&aq->q, pos), > + BIT(aq->q.stride_log2), true); > + > + pos = (aq->q.prod - 1) & aq->q.mask; > + if (pos == aq->q.cons) > + goto out; > + > + ibdev_warn(&dev->ibdev, "admin pos %u (last posted)\n", pos); > + print_hex_dump(KERN_WARNING, "cmd ", DUMP_PREFIX_OFFSET, 16, 1, > + ionic_queue_at(&aq->q, pos), > + BIT(aq->q.stride_log2), true); > + > +out: > + spin_unlock_irqrestore(&aq->lock, irqflags); > +} Please reduce number of debug prints. You are supposed to send driver that works and not for the debug session. > + > +static void ionic_admin_reset_dwork(struct ionic_ibdev *dev) > +{ > + if (atomic_read(&dev->admin_state) >= IONIC_ADMIN_KILLED) > + return; <...> > + if (aq->admin_state >= IONIC_ADMIN_KILLED) > + return; <...> > + ibdev_dbg(&dev->ibdev, "poll admin cq %u prod %u\n", > + cq->cqid, cq->q.prod); > + print_hex_dump_debug("cqe ", DUMP_PREFIX_OFFSET, 16, 1, > + qcqe, BIT(cq->q.stride_log2), true); We have restrack to print CQE and other objects, please use it. > + *cqe = qcqe; > + > + return true; > +} > + > +static void ionic_admin_poll_locked(struct ionic_aq *aq) > +{ > + struct ionic_cq *cq = &aq->vcq->cq[0]; > + struct ionic_admin_wr *wr, *wr_next; > + struct ionic_ibdev *dev = aq->dev; > + u32 wr_strides, avlbl_strides; > + struct ionic_v1_cqe *cqe; > + u32 qtf, qid; > + u16 old_prod; > + u8 type; > + > + lockdep_assert_held(&aq->lock); > + > + if (aq->admin_state >= IONIC_ADMIN_KILLED) { IONIC_ADMIN_KILLED is the last, there is no ">" option. > + list_for_each_entry_safe(wr, wr_next, &aq->wr_prod, aq_ent) { > + INIT_LIST_HEAD(&wr->aq_ent); > + aq->q_wr[wr->status].wr = NULL; > + wr->status = aq->admin_state; > + complete_all(&wr->work); > + } > + INIT_LIST_HEAD(&aq->wr_prod); <...> > + if (do_reset) > + /* Reset device on a timeout */ > + ionic_admin_timedout(bad_aq); I wonder why RDMA driver resets device and not the one who owns PCI. > + else if (do_reschedule) > + /* Try to poll again later */ > + ionic_admin_reset_dwork(dev); > +} <...> > + vcq = kzalloc(sizeof(*vcq), GFP_KERNEL); > + if (!vcq) { > + rc = -ENOMEM; > + goto err_alloc; > + } > + > + vcq->ibcq.device = &dev->ibdev; > + vcq->ibcq.uobject = NULL; 1. There is no need in explicit NULL here, vcq was allocated with kzalloc() 2. Maybe rdma_zalloc_drv_obj() should be used here. > + vcq->ibcq.comp_handler = ionic_rdma_admincq_comp; > + vcq->ibcq.event_handler = ionic_rdma_admincq_event; > + vcq->ibcq.cq_context = NULL; > + atomic_set(&vcq->ibcq.usecnt, 0); <...> > + aq->admin_state = IONIC_ADMIN_KILLED; <...> > + old_state = atomic_cmpxchg(&dev->admin_state, IONIC_ADMIN_ACTIVE, > + IONIC_ADMIN_PAUSED); > + if (old_state != IONIC_ADMIN_ACTIVE) In all these places you are mixing enum_admin_state and atomic_t for same values, but different variable. Please chose or atomic_t or enum. > + return; > + > + /* Pause all the AQs */ > + local_irq_save(irqflags); > + for (i = 0; i < dev->lif_cfg.aq_count; i++) { > + struct ionic_aq *aq = dev->aq_vec[i]; > + > + spin_lock(&aq->lock); > + /* pause rdma admin queues to reset device */ > + if (aq->admin_state == IONIC_ADMIN_ACTIVE) > + aq->admin_state = IONIC_ADMIN_PAUSED; > + spin_unlock(&aq->lock); > + } > + local_irq_restore(irqflags); > + > + rc = ionic_rdma_reset_devcmd(dev); > + if (unlikely(rc)) { > + ibdev_err(&dev->ibdev, "failed to reset rdma %d\n", rc); > + ionic_request_rdma_reset(dev->lif_cfg.lif); > + } > + > + ionic_kill_ibdev(dev, fatal_path); > +} <...> > +static void ionic_cq_event(struct ionic_ibdev *dev, u32 cqid, u8 code) > +{ > + struct ib_event ibev; > + struct ionic_cq *cq; > + > + rcu_read_lock(); > + cq = xa_load(&dev->cq_tbl, cqid); > + if (cq) > + kref_get(&cq->cq_kref); > + rcu_read_unlock(); What and how does this RCU protect? > + > + if (!cq) { Is it possible? > + ibdev_dbg(&dev->ibdev, > + "missing cqid %#x code %u\n", cqid, code); > + return; > + } <...> > module_init(ionic_mod_init); > diff --git a/drivers/infiniband/hw/ionic/ionic_ibdev.h b/drivers/infiniband/hw/ionic/ionic_ibdev.h > index e13adff390d7..e7563c0429fc 100644 > --- a/drivers/infiniband/hw/ionic/ionic_ibdev.h > +++ b/drivers/infiniband/hw/ionic/ionic_ibdev.h > @@ -4,18 +4,243 @@ > #ifndef _IONIC_IBDEV_H_ > #define _IONIC_IBDEV_H_ > > +#include <rdma/ib_umem.h> > #include <rdma/ib_verbs.h> > + > #include <ionic_api.h> > +#include <ionic_regs.h> > + > +#include "ionic_fw.h" > +#include "ionic_queue.h" > +#include "ionic_res.h" > > #include "ionic_lif_cfg.h" > > +#define DRIVER_NAME "ionic_rdma" It is KBUILD_MODNAME, please use it. > +#define DRIVER_SHORTNAME "ionr" > + > #define IONIC_MIN_RDMA_VERSION 0 > #define IONIC_MAX_RDMA_VERSION 2 Nothing from the above is applicable to upstream code. Thanks
On 7/1/25 15:54, Leon Romanovsky wrote: > On Tue, Jun 24, 2025 at 05:43:10PM +0530, Abhijit Gangurde wrote: >> Setup RDMA admin queues using device command exposed over >> auxiliary device and manage these queues using ida. >> >> Co-developed-by: Andrew Boyer <andrew.boyer@amd.com> >> Signed-off-by: Andrew Boyer <andrew.boyer@amd.com> >> Co-developed-by: Allen Hubbe <allen.hubbe@amd.com> >> Signed-off-by: Allen Hubbe <allen.hubbe@amd.com> >> Signed-off-by: Abhijit Gangurde <abhijit.gangurde@amd.com> >> --- >> v2->v3 >> - Fixed lockdep warning >> - Used IDA for resource id allocation >> - Removed rw locks around xarrays >> >> drivers/infiniband/hw/ionic/ionic_admin.c | 1169 +++++++++++++++++ >> .../infiniband/hw/ionic/ionic_controlpath.c | 184 +++ >> drivers/infiniband/hw/ionic/ionic_fw.h | 164 +++ >> drivers/infiniband/hw/ionic/ionic_ibdev.c | 56 + >> drivers/infiniband/hw/ionic/ionic_ibdev.h | 225 ++++ >> drivers/infiniband/hw/ionic/ionic_pgtbl.c | 113 ++ >> drivers/infiniband/hw/ionic/ionic_queue.c | 52 + >> drivers/infiniband/hw/ionic/ionic_queue.h | 234 ++++ >> drivers/infiniband/hw/ionic/ionic_res.h | 154 +++ >> 9 files changed, 2351 insertions(+) >> create mode 100644 drivers/infiniband/hw/ionic/ionic_admin.c >> create mode 100644 drivers/infiniband/hw/ionic/ionic_controlpath.c >> create mode 100644 drivers/infiniband/hw/ionic/ionic_fw.h >> create mode 100644 drivers/infiniband/hw/ionic/ionic_pgtbl.c >> create mode 100644 drivers/infiniband/hw/ionic/ionic_queue.c >> create mode 100644 drivers/infiniband/hw/ionic/ionic_queue.h >> create mode 100644 drivers/infiniband/hw/ionic/ionic_res.h > <...> > >> +static void ionic_admin_timedout(struct ionic_aq *aq) >> +{ >> + struct ionic_cq *cq = &aq->vcq->cq[0]; >> + struct ionic_ibdev *dev = aq->dev; >> + unsigned long irqflags; >> + u16 pos; >> + >> + spin_lock_irqsave(&aq->lock, irqflags); >> + if (ionic_queue_empty(&aq->q)) >> + goto out; >> + >> + /* Reset ALL adminq if any one times out */ >> + if (aq->admin_state < IONIC_ADMIN_KILLED) >> + queue_work(ionic_evt_workq, &dev->reset_work); >> + >> + ibdev_err(&dev->ibdev, "admin command timed out, aq %d\n", aq->aqid); >> + >> + ibdev_warn(&dev->ibdev, "admin timeout was set for %ums\n", >> + (u32)jiffies_to_msecs(IONIC_ADMIN_TIMEOUT)); >> + ibdev_warn(&dev->ibdev, "admin inactivity for %ums\n", >> + (u32)jiffies_to_msecs(jiffies - aq->stamp)); >> + >> + ibdev_warn(&dev->ibdev, "admin commands outstanding %u\n", >> + ionic_queue_length(&aq->q)); >> + ibdev_warn(&dev->ibdev, "%s more commands pending\n", >> + list_empty(&aq->wr_post) ? "no" : "some"); >> + >> + pos = cq->q.prod; >> + >> + ibdev_warn(&dev->ibdev, "admin cq pos %u (next to complete)\n", pos); >> + print_hex_dump(KERN_WARNING, "cqe ", DUMP_PREFIX_OFFSET, 16, 1, >> + ionic_queue_at(&cq->q, pos), >> + BIT(cq->q.stride_log2), true); >> + >> + pos = (pos - 1) & cq->q.mask; >> + >> + ibdev_warn(&dev->ibdev, "admin cq pos %u (last completed)\n", pos); >> + print_hex_dump(KERN_WARNING, "cqe ", DUMP_PREFIX_OFFSET, 16, 1, >> + ionic_queue_at(&cq->q, pos), >> + BIT(cq->q.stride_log2), true); >> + >> + pos = aq->q.cons; >> + >> + ibdev_warn(&dev->ibdev, "admin pos %u (next to complete)\n", pos); >> + print_hex_dump(KERN_WARNING, "cmd ", DUMP_PREFIX_OFFSET, 16, 1, >> + ionic_queue_at(&aq->q, pos), >> + BIT(aq->q.stride_log2), true); >> + >> + pos = (aq->q.prod - 1) & aq->q.mask; >> + if (pos == aq->q.cons) >> + goto out; >> + >> + ibdev_warn(&dev->ibdev, "admin pos %u (last posted)\n", pos); >> + print_hex_dump(KERN_WARNING, "cmd ", DUMP_PREFIX_OFFSET, 16, 1, >> + ionic_queue_at(&aq->q, pos), >> + BIT(aq->q.stride_log2), true); >> + >> +out: >> + spin_unlock_irqrestore(&aq->lock, irqflags); >> +} > Please reduce number of debug prints. You are supposed to send driver > that works and not for the debug session. I will keep it minimal. > >> + >> +static void ionic_admin_reset_dwork(struct ionic_ibdev *dev) >> +{ >> + if (atomic_read(&dev->admin_state) >= IONIC_ADMIN_KILLED) >> + return; > <...> > >> + if (aq->admin_state >= IONIC_ADMIN_KILLED) >> + return; > <...> > >> + ibdev_dbg(&dev->ibdev, "poll admin cq %u prod %u\n", >> + cq->cqid, cq->q.prod); >> + print_hex_dump_debug("cqe ", DUMP_PREFIX_OFFSET, 16, 1, >> + qcqe, BIT(cq->q.stride_log2), true); > We have restrack to print CQE and other objects, please use it. sure > >> + *cqe = qcqe; >> + >> + return true; >> +} >> + >> +static void ionic_admin_poll_locked(struct ionic_aq *aq) >> +{ >> + struct ionic_cq *cq = &aq->vcq->cq[0]; >> + struct ionic_admin_wr *wr, *wr_next; >> + struct ionic_ibdev *dev = aq->dev; >> + u32 wr_strides, avlbl_strides; >> + struct ionic_v1_cqe *cqe; >> + u32 qtf, qid; >> + u16 old_prod; >> + u8 type; >> + >> + lockdep_assert_held(&aq->lock); >> + >> + if (aq->admin_state >= IONIC_ADMIN_KILLED) { > IONIC_ADMIN_KILLED is the last, there is no ">" option. Will correct this. > >> + list_for_each_entry_safe(wr, wr_next, &aq->wr_prod, aq_ent) { >> + INIT_LIST_HEAD(&wr->aq_ent); >> + aq->q_wr[wr->status].wr = NULL; >> + wr->status = aq->admin_state; >> + complete_all(&wr->work); >> + } >> + INIT_LIST_HEAD(&aq->wr_prod); > <...> > >> + if (do_reset) >> + /* Reset device on a timeout */ >> + ionic_admin_timedout(bad_aq); > I wonder why RDMA driver resets device and not the one who owns PCI. RDMA driver is requesting the reset via eth driver which holds the privilege. > >> + else if (do_reschedule) >> + /* Try to poll again later */ >> + ionic_admin_reset_dwork(dev); >> +} > <...> > >> + vcq = kzalloc(sizeof(*vcq), GFP_KERNEL); >> + if (!vcq) { >> + rc = -ENOMEM; >> + goto err_alloc; >> + } >> + >> + vcq->ibcq.device = &dev->ibdev; >> + vcq->ibcq.uobject = NULL; > 1. There is no need in explicit NULL here, vcq was allocated with kzalloc() > 2. Maybe rdma_zalloc_drv_obj() should be used here. Will remove the explicit NULLs. > >> + vcq->ibcq.comp_handler = ionic_rdma_admincq_comp; >> + vcq->ibcq.event_handler = ionic_rdma_admincq_event; >> + vcq->ibcq.cq_context = NULL; >> + atomic_set(&vcq->ibcq.usecnt, 0); > <...> > >> + aq->admin_state = IONIC_ADMIN_KILLED; > <...> > >> + old_state = atomic_cmpxchg(&dev->admin_state, IONIC_ADMIN_ACTIVE, >> + IONIC_ADMIN_PAUSED); >> + if (old_state != IONIC_ADMIN_ACTIVE) > In all these places you are mixing enum_admin_state and atomic_t for > same values, but different variable. Please chose or atomic_t or enum. admin_state within the admin queues is protected by the spinlock, hence it is used as enum_admin_state. However device's admin_state is used as as atomic to avoid reset race of reset. > >> + return; >> + >> + /* Pause all the AQs */ >> + local_irq_save(irqflags); >> + for (i = 0; i < dev->lif_cfg.aq_count; i++) { >> + struct ionic_aq *aq = dev->aq_vec[i]; >> + >> + spin_lock(&aq->lock); >> + /* pause rdma admin queues to reset device */ >> + if (aq->admin_state == IONIC_ADMIN_ACTIVE) >> + aq->admin_state = IONIC_ADMIN_PAUSED; >> + spin_unlock(&aq->lock); >> + } >> + local_irq_restore(irqflags); >> + >> + rc = ionic_rdma_reset_devcmd(dev); >> + if (unlikely(rc)) { >> + ibdev_err(&dev->ibdev, "failed to reset rdma %d\n", rc); >> + ionic_request_rdma_reset(dev->lif_cfg.lif); >> + } >> + >> + ionic_kill_ibdev(dev, fatal_path); >> +} > <...> > >> +static void ionic_cq_event(struct ionic_ibdev *dev, u32 cqid, u8 code) >> +{ >> + struct ib_event ibev; >> + struct ionic_cq *cq; >> + >> + rcu_read_lock(); >> + cq = xa_load(&dev->cq_tbl, cqid); >> + if (cq) >> + kref_get(&cq->cq_kref); >> + rcu_read_unlock(); > What and how does this RCU protect? This is to protect kref_get against free of cq in destroy_cq. > >> + >> + if (!cq) { > Is it possible? Possible when HCA goes bad. > >> + ibdev_dbg(&dev->ibdev, >> + "missing cqid %#x code %u\n", cqid, code); >> + return; >> + } > <...> > >> module_init(ionic_mod_init); >> diff --git a/drivers/infiniband/hw/ionic/ionic_ibdev.h b/drivers/infiniband/hw/ionic/ionic_ibdev.h >> index e13adff390d7..e7563c0429fc 100644 >> --- a/drivers/infiniband/hw/ionic/ionic_ibdev.h >> +++ b/drivers/infiniband/hw/ionic/ionic_ibdev.h >> @@ -4,18 +4,243 @@ >> #ifndef _IONIC_IBDEV_H_ >> #define _IONIC_IBDEV_H_ >> >> +#include <rdma/ib_umem.h> >> #include <rdma/ib_verbs.h> >> + >> #include <ionic_api.h> >> +#include <ionic_regs.h> >> + >> +#include "ionic_fw.h" >> +#include "ionic_queue.h" >> +#include "ionic_res.h" >> >> #include "ionic_lif_cfg.h" >> >> +#define DRIVER_NAME "ionic_rdma" > It is KBUILD_MODNAME, please use it. sure. > >> +#define DRIVER_SHORTNAME "ionr" >> + >> #define IONIC_MIN_RDMA_VERSION 0 >> #define IONIC_MAX_RDMA_VERSION 2 > Nothing from the above is applicable to upstream code. > > Thanks I will remove this. Thanks, Abhijit.
On Thu, Jul 03, 2025 at 12:29:45PM +0530, Abhijit Gangurde wrote: > > On 7/1/25 15:54, Leon Romanovsky wrote: > > On Tue, Jun 24, 2025 at 05:43:10PM +0530, Abhijit Gangurde wrote: > > > Setup RDMA admin queues using device command exposed over > > > auxiliary device and manage these queues using ida. > > > > > > Co-developed-by: Andrew Boyer <andrew.boyer@amd.com> > > > Signed-off-by: Andrew Boyer <andrew.boyer@amd.com> > > > Co-developed-by: Allen Hubbe <allen.hubbe@amd.com> > > > Signed-off-by: Allen Hubbe <allen.hubbe@amd.com> > > > Signed-off-by: Abhijit Gangurde <abhijit.gangurde@amd.com> > > > --- > > > v2->v3 > > > - Fixed lockdep warning > > > - Used IDA for resource id allocation > > > - Removed rw locks around xarrays <...> > > > > > + list_for_each_entry_safe(wr, wr_next, &aq->wr_prod, aq_ent) { > > > + INIT_LIST_HEAD(&wr->aq_ent); > > > + aq->q_wr[wr->status].wr = NULL; > > > + wr->status = aq->admin_state; > > > + complete_all(&wr->work); > > > + } > > > + INIT_LIST_HEAD(&aq->wr_prod); > > <...> > > > > > + if (do_reset) > > > + /* Reset device on a timeout */ > > > + ionic_admin_timedout(bad_aq); > > I wonder why RDMA driver resets device and not the one who owns PCI. > > RDMA driver is requesting the reset via eth driver which holds the > privilege. I wonder if the one who owns CMD interface should decide and reset device and not the clients. > <...> > > > + old_state = atomic_cmpxchg(&dev->admin_state, IONIC_ADMIN_ACTIVE, > > > + IONIC_ADMIN_PAUSED); > > > + if (old_state != IONIC_ADMIN_ACTIVE) > > In all these places you are mixing enum_admin_state and atomic_t for > > same values, but different variable. Please chose or atomic_t or enum. > > admin_state within the admin queues is protected by the spinlock, > hence it is used as enum_admin_state. However device's admin_state > is used as as atomic to avoid reset race of reset. The issue is in mixing types. > <...> > > > + > > > + if (!cq) { > > Is it possible? > > Possible when HCA goes bad. Do you have errata for that? Generally speaking, kernel is not written to be protected from broken HW. The overall assumption is that HW works correctly. Thanks
On 7/3/25 14:11, Leon Romanovsky wrote: > On Thu, Jul 03, 2025 at 12:29:45PM +0530, Abhijit Gangurde wrote: >> On 7/1/25 15:54, Leon Romanovsky wrote: >>> On Tue, Jun 24, 2025 at 05:43:10PM +0530, Abhijit Gangurde wrote: >>>> Setup RDMA admin queues using device command exposed over >>>> auxiliary device and manage these queues using ida. >>>> >>>> Co-developed-by: Andrew Boyer <andrew.boyer@amd.com> >>>> Signed-off-by: Andrew Boyer <andrew.boyer@amd.com> >>>> Co-developed-by: Allen Hubbe <allen.hubbe@amd.com> >>>> Signed-off-by: Allen Hubbe <allen.hubbe@amd.com> >>>> Signed-off-by: Abhijit Gangurde <abhijit.gangurde@amd.com> >>>> --- >>>> v2->v3 >>>> - Fixed lockdep warning >>>> - Used IDA for resource id allocation >>>> - Removed rw locks around xarrays > <...> > >>>> + list_for_each_entry_safe(wr, wr_next, &aq->wr_prod, aq_ent) { >>>> + INIT_LIST_HEAD(&wr->aq_ent); >>>> + aq->q_wr[wr->status].wr = NULL; >>>> + wr->status = aq->admin_state; >>>> + complete_all(&wr->work); >>>> + } >>>> + INIT_LIST_HEAD(&aq->wr_prod); >>> <...> >>> >>>> + if (do_reset) >>>> + /* Reset device on a timeout */ >>>> + ionic_admin_timedout(bad_aq); >>> I wonder why RDMA driver resets device and not the one who owns PCI. >> RDMA driver is requesting the reset via eth driver which holds the >> privilege. > I wonder if the one who owns CMD interface should decide and reset device > and not the clients. To be precise, this operation resets the RDMA logical interface built on top of the base device, and does not affect the PCI device or the Ethernet driver's interface. Apologies for the lack of clarity in the previous comment. I will update the comment to reflect this accurately in the next version. > <...> > >>>> + old_state = atomic_cmpxchg(&dev->admin_state, IONIC_ADMIN_ACTIVE, >>>> + IONIC_ADMIN_PAUSED); >>>> + if (old_state != IONIC_ADMIN_ACTIVE) >>> In all these places you are mixing enum_admin_state and atomic_t for >>> same values, but different variable. Please chose or atomic_t or enum. >> admin_state within the admin queues is protected by the spinlock, >> hence it is used as enum_admin_state. However device's admin_state >> is used as as atomic to avoid reset race of reset. > The issue is in mixing types. I will correct this. > > <...> > >>>> + >>>> + if (!cq) { >>> Is it possible? >> Possible when HCA goes bad. > Do you have errata for that? Generally speaking, kernel is not written > to be protected from broken HW. The overall assumption is that HW works > correctly. > > Thanks There is no known hw issue around this. This was added just as a precautionary check so that wrong cqid does not result in kernel panic. I can remove this check if this is unwarranted. Thanks, Abhijit
© 2016 - 2025 Red Hat, Inc.