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@xxxxxxx>
Signed-off-by: Andrew Boyer <andrew.boyer@xxxxxxx>
Co-developed-by: Allen Hubbe <allen.hubbe@xxxxxxx>
Signed-off-by: Allen Hubbe <allen.hubbe@xxxxxxx>
Signed-off-by: Abhijit Gangurde <abhijit.gangurde@xxxxxxx>
---
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.
