When one side of UALink nodes reboots, its ring wptr resets to 0 while
the peer keeps a stale value. Read the remote wptr and adopt it as the
local wptr and rptr to resync, before the first message is sent. Either
side can notice first, so do this both when setting up the connection
and when receiving the HELLO from the peer.
On the sdma path, read the wptr in a fence callback because the IB buffer
holding the result is recycled once the fence signals, so reading it
after dma_fence_wait_timeout() is too late. The lsdma path copies
synchronously, so it reads the wptr directly.
Take peer->lock in the callers (amdgpu_ualink_remote_interrupt and
amdgpu_ualink_remote_shootdown) instead of in amdgpu_ualink_send_command,
so the ring->ready check and the send are atomic against a concurrent
ring reset.
Move the rptr refresh and ring-full check out of amdgpu_ualink_send_command()
into amdgpu_ualink_ring_check_space(), so it runs under the caller's
peer->lock before the job is allocated.
On a failed NPA-REQ send, reset the connection like the NPA-RSP timeout path
already does, so the next attempt resyncs the ring wptr.
On connection reset, amdgpu_ualink_reset_peer_rings() only clears
ring->ready, so the writeback address is re-registered on the next send.
v10:
- Fix the resync missed on the receiving side: after a reboot the peer
may send HELLO before we set up the connection, so resync the ring
wptr in amdgpu_ualink_process_hello_msg() too.
v9:
- Fix the unnecessary connection reset after an app exits: the peer is
idle here but still busy handling NPA revokes on the remote, so the
HELLO reset a healthy connection. Only send it if the wptr changed.
- Resync both rings together under peer->lock, read both remote wptr
before updating either so a failed read changes nothing.
v8:
- Fix regression with concurrent import ioctls: do the idle check and
ring reset in amdgpu_ualink_setup_connection() under conn_state->lock,
and return -EAGAIN for IN_PROGRESS/PENDING first.
- Reset the peer rings only, the imp/exp xarray cleanup is not needed to
resync wptr.
v7:
- With v6 reboot one node of vPOD test passed on minirack.
- Use the per-remote imported and exported handle lists for the idle
check instead of walking the xarrays.
- Rename amdgpu_ualink_resync_peer_rings() to
amdgpu_ualink_reset_peer_rings(), it only clears ring->ready.
v6:
- Check idleness per remote accelerator instead of globally, and reset
the stale connection at the import ioctl so the peer re-runs the
HELLO handshake after a reboot.
v5:
- Resync the ring pointers at the import ioctl when the connection is
completely idle (no exported or imported BOs), so the first command
from a new application after a remote reboot/reset succeeds.
- Put the sdma wptr read fence callback on the stack instead of
allocating it; it is cancelled on timeout so it never outlives the
call.
v4:
- Cancel the sdma fence callback on the wait timeout/failure path so it
cannot write into the caller's stack after the function returns, and
free the callback state in one place.
v3:
- Resync both rings at connection setup instead of on the send path.
- Set the writeback rptr too, the stale value disagrees with the
resynced wptr until the remote FW writes it back.
v2:
- Match the FW change: resync wptr only on the first send or after a
timeout, instead of checking the remote periodically.
Suggested-by: Felix Kuehling<[email protected]>
Signed-off-by: Philip Yang<[email protected]>
Reviewed-by: Felix Kuehling<[email protected]>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c | 339 +++++++++++++++++----
1 file changed, 287 insertions(+), 52 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
index b2c2650b0149..c0327f6854a6 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
@@ -56,6 +56,10 @@ static int amdgpu_ualink_remote_shootdown(struct
amdgpu_device *adev,
u32 size_in_pages, u32 flush_type);
static void __amdgpu_ualink_activate_vpod_locked(struct amdgpu_device *adev);
static bool amdgpu_ualink_vpod_membership_changed(struct amdgpu_device *adev);
+static void amdgpu_ualink_reset_peer_rings(struct amdgpu_device *adev,
+ u32 remote_accel_id);
+static int amdgpu_ualink_ring_resync_ptrs(struct amdgpu_device *adev,
+ u32 remote_accel_id);
#define STRIP_NPA(addr) \
(((u64)(addr) & ~AMDGPU_UALINK_NPA_ADDR_GPUID_MASK))
@@ -2282,11 +2286,23 @@ static void amdgpu_ualink_process_hello_msg(struct
amdgpu_device *adev,
generation_count);
}
+ /* Peer may have rebooted. */
+ r = amdgpu_ualink_ring_resync_ptrs(adev, sender_acc_id);
+ if (r < 0) {
+ dev_err(adev->dev,
+ "HELLO: resync sender AccId %u ptrs failed %d\n",
+ sender_acc_id, r);
+ goto out_set_state;
+ }
+ if (r)
+ amdgpu_ualink_reset_peer_rings(adev, sender_acc_id);
+
r = amdgpu_ualink_send_hello_ack_msg(adev, sender_acc_id);
if (r)
dev_err(adev->dev, "HELLO-ACK: send failed to remote
AccId:%u\n",
sender_acc_id);
+out_set_state:
mutex_lock(&conn_state->lock);
if (r) {
conn_state->state = AMDGPU_UALINK_CONN_NOT_READY;
@@ -2303,6 +2319,14 @@ static void amdgpu_ualink_process_hello_msg(struct
amdgpu_device *adev,
mutex_unlock(&conn_state->lock);
}
+/* True if nothing is exported to, or imported from, this remote accelerator. */
+static bool amdgpu_ualink_peer_idle(struct amdgpu_device *adev, u32
remote_accel_id)
+{
+ /* Unlocked: a racing import at worst costs one redundant ring resync.
*/
+ return list_empty(&adev->ualink.imp_handles_list[remote_accel_id]) &&
+ list_empty(&adev->ualink.exp_handles_list[remote_accel_id]);
+}
+
static int amdgpu_ualink_setup_connection(struct amdgpu_device *adev,
u32 remote_acc_id)
{
@@ -2324,16 +2348,31 @@ static int amdgpu_ualink_setup_connection(struct
amdgpu_device *adev,
* already done by another thread.
*/
mutex_lock(&conn_state->lock);
- if (conn_state->state == AMDGPU_UALINK_CONN_ESTABLISHED) {
- r = 0;
- goto out;
- } else if (conn_state->state == AMDGPU_UALINK_CONN_IN_PROGRESS ||
- conn_state->state == AMDGPU_UALINK_CONN_PENDING) {
+ if (conn_state->state == AMDGPU_UALINK_CONN_IN_PROGRESS ||
+ conn_state->state == AMDGPU_UALINK_CONN_PENDING) {
r = -EAGAIN;
goto out;
}
+ if (conn_state->state == AMDGPU_UALINK_CONN_ESTABLISHED &&
+ !amdgpu_ualink_peer_idle(adev, remote_acc_id)) {
+ r = 0;
+ goto out;
+ }
+
+ /* Peer is idle or not connected, it may have rebooted. Skip the
+ * handshake if the rings are still in sync, it would reset the remote.
+ */
+ r = amdgpu_ualink_ring_resync_ptrs(adev, remote_acc_id);
+ if (r < 0)
+ goto out;
+ if (!r && conn_state->state == AMDGPU_UALINK_CONN_ESTABLISHED)
+ goto out;
+
+ amdgpu_ualink_reset_peer_rings(adev, remote_acc_id);
+
conn_state->state = AMDGPU_UALINK_CONN_IN_PROGRESS;
+
mutex_unlock(&conn_state->lock);
/* Send HELLO message */
@@ -3670,7 +3709,7 @@ static int amdgpu_ualink_do_import_handle(struct
amdgpu_device *adev,
dev_warn(adev->dev,
"IMPORT: NPA-REQ send failed to remote AccId:%u\n",
remote_acc_id);
- return r;
+ goto reset_conn;
}
/* Wait for the NPA_RSP to come back */
@@ -5511,6 +5550,34 @@ static void amdgpu_ualink_emit_update_wb_addr(u32
**cpu_addr_p, u32 wb_data,
typedef void (*ualink_emit_packet)(u32 **cpu_addr_p, u32 wb, u32 dw0,
u32 dw1, u32 dw2, u32 dw3);
+/*
+ * Refresh rptr from the writeback and check the ring has room for one more
+ * command. The caller must hold the peer lock.
+ *
+ * Return 0 on success or a negative error code.
+ */
+static int amdgpu_ualink_ring_check_space(struct amdgpu_device *adev,
+ u32 remote_accel_id,
+ struct amdgpu_ualink_ring *ring,
+ struct amdgpu_ualink_wb *wb_cpu)
+{
+ ring->rptr = READ_ONCE(wb_cpu->rptr);
+
+ if (WARN_ON_ONCE(ring->rptr > ring->wptr)) {
+ dev_err(adev->dev, "accel_id %u ring overflow wptr 0x%llx rptr
0x%llx\n",
+ remote_accel_id, ring->wptr, ring->rptr);
+ return -EFAULT;
+ }
+
+ if ((ring->wptr + 1 - ring->rptr) >= ring->rb_size) {
+ dev_err(adev->dev, "accel_id %u command ring full wptr 0x%llx rptr
0x%llx\n",
+ remote_accel_id, ring->wptr, ring->rptr);
+ return -ENOSPC;
+ }
+
+ return 0;
+}
+
static int amdgpu_ualink_send_command(struct amdgpu_device *adev,
u32 remote_accel_id,
struct amdgpu_ualink_ring *ring,
@@ -5536,6 +5603,10 @@ static int amdgpu_ualink_send_command(struct
amdgpu_device *adev,
peer = &remote->peer[remote_accel_id];
+ r = amdgpu_ualink_ring_check_space(adev, remote_accel_id, ring, wb_cpu);
+ if (r)
+ goto out;
+
/*
* 3 sdma copy commands: write data to ring buffer, update wptr, ring
doorbell
*
@@ -5552,25 +5623,7 @@ static int amdgpu_ualink_send_command(struct
amdgpu_device *adev,
ndw * 4, AMDGPU_IB_POOL_IMMEDIATE,
AMDGPU_KERNEL_JOB_ID_TTM_COPY_BUFFER,
&job);
if (r)
- return r;
-
- mutex_lock(&peer->lock);
-
- ring->rptr = READ_ONCE(wb_cpu->rptr);
-
- if (WARN_ON_ONCE(ring->rptr > ring->wptr)) {
- dev_err(adev->dev, "accel_id %u ring overflow wptr 0x%llx rptr
0x%llx\n",
- remote_accel_id, ring->wptr, ring->rptr);
- r = -EFAULT;
- goto unlock_free;
- }
-
- if ((ring->wptr + 1 - ring->rptr) >= ring->rb_size) {
- dev_err(adev->dev, "accel_id %u command ring full wptr 0x%llx rptr
0x%llx\n",
- remote_accel_id, ring->wptr, ring->rptr);
- r = -ENOSPC;
- goto unlock_free;
- }
+ goto out;
ib = &job->ibs[0];
src = ib->gpu_addr + ndw_copy_cmd * 4;
@@ -5628,12 +5681,11 @@ static int amdgpu_ualink_send_command(struct
amdgpu_device *adev,
dev_dbg(adev->dev,
"remote %u lsdma copy failed (r %d), skip completion
wait\n",
remote_accel_id, r);
- mutex_unlock(&peer->lock);
return r;
}
goto out_wait_complete;
- }
+ }
sdma_ring = &adev->sdma.instance[0].ring;
@@ -5664,12 +5716,9 @@ static int amdgpu_ualink_send_command(struct amdgpu_device *adev,
if (r <= 0) {
dev_dbg(adev->dev, "remote %u sdma fence wait return r %d\n",
remote_accel_id, r);
-
if (r == 0)
r = -ETIME;
-
- mutex_unlock(&peer->lock);
- return r;
+ goto out;
}
out_wait_complete:
@@ -5678,21 +5727,15 @@ static int amdgpu_ualink_send_command(struct
amdgpu_device *adev,
*/
r = amdgpu_ualink_remote_wait_timeout(adev, remote_accel_id, wb_cpu,
seq);
- /* increase local copy ring wptr, only if FW not timeout */
- if (r != -ETIME)
- ring->wptr++;
-
/*
- * Release ring lock after the remote FW handle command completes to
- * prevent race conditions.
+ * Advance wptr unless the remote timed out, even if the FW returned an
+ * error status. On timeout the caller resets the connection, which
+ * resyncs wptr from the remote in amdgpu_ualink_setup_connection.
*/
- mutex_unlock(&peer->lock);
- return r;
-
+ if (r != -ETIME)
+ ring->wptr++;
-unlock_free:
- mutex_unlock(&peer->lock);
- amdgpu_job_free(job);
+out:
dev_dbg(adev->dev, "ret r = %d\n", r);
return r;
}
@@ -5732,6 +5775,192 @@ static void amdgpu_ualink_get_wb_addr(struct
amdgpu_device *adev,
remote_accel_id, npa + offset);
}
+struct amdgpu_ualink_wptr_fence_cb {
+ struct dma_fence_cb base;
+ u64 *wptr_cpu;
+ u64 result;
+};
+
+static void amdgpu_ualink_read_wptr_fence_cb(struct dma_fence *fence,
+ struct dma_fence_cb *cb)
+{
+ struct amdgpu_ualink_wptr_fence_cb *wptr_cb;
+
+ wptr_cb = container_of(cb, typeof(*wptr_cb), base);
+ wptr_cb->result = *wptr_cb->wptr_cpu;
+}
+
+static int amdgpu_ualink_read_remote_wptr(struct amdgpu_device *adev,
+ u32 remote_accel_id,
+ struct amdgpu_ualink_ring *ring,
+ u64 *wptr)
+{
+ struct amdgpu_ualink_remote *remote = to_remote(adev);
+ struct amdgpu_ualink_peer *peer = &remote->peer[remote_accel_id];
+ struct amdgpu_ualink_wptr_fence_cb wptr_cb = { };
+ struct amdgpu_ring *sdma_ring;
+ struct dma_fence *fence;
+ struct amdgpu_job *job;
+ struct amdgpu_ib *ib;
+ u32 ndw, ndw_copy_cmd;
+ u64 wptr_gpu, *wptr_cpu;
+ int r = 0;
+
