Skip to content

Always request a cqe on xfer submits #10901

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 2 commits into
base: device-api
Choose a base branch
from
Draft

Always request a cqe on xfer submits #10901

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7704a47
UCT/GDAKI: Fix WQE write ordering with threadfence
michal-shalev Sep 18, 2025
1d06b2f
Always request cqes
itayalroy Sep 18, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 4 additions & 9 deletions src/uct/cuda/gdaki/gdaki.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -126,8 +126,7 @@ __device__ static inline ucs_status_t uct_gdaki_progress_thread(uct_gdaki_dev_ep
}

cuda::atomic_ref<uint64_t, cuda::thread_scope_device> ref(qp->sq_wqe_pi);
uint64_t sq_wqe_pi = qp->sq_wqe_pi;
ref.fetch_max(((wqe_cnt - sq_wqe_pi) & 0xffff) + sq_wqe_pi + 1);
ref.fetch_add(1);

doca_gpu_dev_verbs_fence_release<DOCA_GPUNETIO_VERBS_SYNC_SCOPE_GPU>();
}
Expand Down Expand Up @@ -254,6 +253,7 @@ uct_gdaki_db(struct doca_gpu_dev_verbs_qp *qp, uint64_t wqe_base, unsigned count
{
cuda::atomic_ref<uint64_t, cuda::thread_scope_device> ref(qp->sq_ready_index);
uint64_t wqe_base_orig = wqe_base;
__threadfence();
while (!ref.compare_exchange_strong(wqe_base, wqe_base + count,
cuda::std::memory_order_relaxed)) {
wqe_base = wqe_base_orig;
Expand Down Expand Up @@ -553,7 +553,6 @@ uct_gdaki_put_batch_part_impl(uct_gdaki_batch_t *batch, uint64_t flags,
uint64_t dst;
uint32_t rkey;
int opcode;
uint32_t fc;
unsigned lane_id;
unsigned num_lanes;
#if ENABLE_PARAMS_CHECK
Expand All @@ -580,7 +579,6 @@ uct_gdaki_put_batch_part_impl(uct_gdaki_batch_t *batch, uint64_t flags,
return UCS_ERR_NO_RESOURCE;
}

fc = doca_gpu_dev_verbs_wqe_idx_inc_mask(qp->sq_wqe_pi, qp->sq_wqe_num / 2);
wqe_idx = doca_gpu_dev_verbs_wqe_idx_inc_mask(wqe_base, lane_id);
for (uint32_t i = lane_id; i < count; i += num_lanes) {
uint32_t idx;
Expand All @@ -601,11 +599,8 @@ uct_gdaki_put_batch_part_impl(uct_gdaki_batch_t *batch, uint64_t flags,
continue;
}

if (((flags & UCT_DEV_BATCH_FLAG_COMP) && (i == count - 1)) ||
(!(flags & UCT_DEV_BATCH_FLAG_COMP) && (wqe_idx == fc))) {
cflag = DOCA_GPUNETIO_MLX5_WQE_CTRL_CQ_UPDATE;
ep->ops[wqe_idx & qp->sq_wqe_mask].comp = comp;
}
cflag = DOCA_GPUNETIO_MLX5_WQE_CTRL_CQ_UPDATE;
ep->ops[wqe_idx & qp->sq_wqe_mask].comp = comp;

wqe_ptr = doca_gpu_dev_verbs_get_wqe_ptr(qp, wqe_idx);
src = batch->list[idx].src + src_off;
Expand Down