github-mirror/linux
1
0
Fork 0
mirror of https://github.com/torvalds/linux.git synced 2026-09-25 17:42:03 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity Graph

RDMA/hns: Support drain SQ and RQ

Browse Source 
Some ULPs, e.g. rpcrdma, rely on drain_qp() to ensure all outstanding
requests are completed before releasing related memory. If drain_qp()
fails, ULPs may release memory directly, and in-flight WRs may later be
flushed after the memory is freed, potentially leading to UAF.

drain_qp() failures can happen when HW enters an error state or is
reset. Add support to drain SQ and RQ in such cases by posting a
fake WR during reset, so the driver can process all remaining WRs in
sequence and generate corresponding completions.

Always invoke comp_handler() in drain process to ensure completions
are not lost under concurrency (e.g. concurrent post_send() and
reset, or QPs created during reset). If the CQ is already processed,
cancel any already scheduled comp_handler() to avoid concurrency
issues.

Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20260108113032.856306-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
This commit is contained in:
Chengchang Tang 2026-01-08 19:30:32 +08:00 committed by Leon Romanovsky
parent 5c3f795d17
commit 354e7a6d44
1 changed files with 166 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

166
drivers/infiniband/hw/hns/hns_roce_hw_v2.c
View File

@ -876,6 +876,170 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
return ret;
}
static int hns_roce_push_drain_wr(struct hns_roce_wq *wq, struct ib_cq *cq,
u64 wr_id)
{
unsigned long flags;
int ret = 0;
spin_lock_irqsave(&wq->lock, flags);
if (hns_roce_wq_overflow(wq, 1, cq)) {
ret = -ENOMEM;
goto out;
}
wq->wrid[wq->head & (wq->wqe_cnt - 1)] = wr_id;
wq->head++;
out:
spin_unlock_irqrestore(&wq->lock, flags);
return ret;
}
struct hns_roce_drain_cqe {
struct ib_cqe cqe;
struct completion done;
};
static void hns_roce_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
{
struct hns_roce_drain_cqe *cqe = container_of(wc->wr_cqe,
struct hns_roce_drain_cqe,
cqe);
complete(&cqe->done);
}
static void handle_drain_completion(struct ib_cq *ibcq,
struct hns_roce_drain_cqe *drain,
struct hns_roce_dev *hr_dev)
{
#define TIMEOUT (HZ / 10)
struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
unsigned long flags;
bool triggered;
if (ibcq->poll_ctx == IB_POLL_DIRECT) {
while (wait_for_completion_timeout(&drain->done, TIMEOUT) <= 0)
ib_process_cq_direct(ibcq, -1);
return;
}
if (hr_dev->state < HNS_ROCE_DEVICE_STATE_RST_DOWN)
goto waiting_done;
spin_lock_irqsave(&hr_cq->lock, flags);
triggered = hr_cq->is_armed;
hr_cq->is_armed = 1;
spin_unlock_irqrestore(&hr_cq->lock, flags);
/* Triggered means this cq is processing or has been processed
* by hns_roce_handle_device_err() or this function. We need to
* cancel the already invoked comp_handler() to avoid concurrency.
* If it has not been triggered, we can directly invoke
* comp_handler().
*/
if (triggered) {
switch (ibcq->poll_ctx) {
case IB_POLL_SOFTIRQ:
irq_poll_disable(&ibcq->iop);
irq_poll_enable(&ibcq->iop);
break;
case IB_POLL_WORKQUEUE:
case IB_POLL_UNBOUND_WORKQUEUE:
cancel_work_sync(&ibcq->work);
break;
default:
WARN_ON_ONCE(1);
}
}
if (ibcq->comp_handler)
ibcq->comp_handler(ibcq, ibcq->cq_context);
waiting_done:
if (ibcq->comp_handler)
wait_for_completion(&drain->done);
}
static void hns_roce_v2_drain_rq(struct ib_qp *ibqp)
{
struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
struct hns_roce_drain_cqe rdrain = {};
const struct ib_recv_wr *bad_rwr;
struct ib_cq *cq = ibqp->recv_cq;
struct ib_recv_wr rwr = {};
int ret;
ret = ib_modify_qp(ibqp, &attr, IB_QP_STATE);
if (ret && hr_dev->state < HNS_ROCE_DEVICE_STATE_RST_DOWN) {
ibdev_err_ratelimited(&hr_dev->ib_dev,
"failed to modify qp during drain rq, ret = %d.\n",
ret);
return;
}
rwr.wr_cqe = &rdrain.cqe;
rdrain.cqe.done = hns_roce_drain_qp_done;
init_completion(&rdrain.done);
if (hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN)
ret = hns_roce_push_drain_wr(&hr_qp->rq, cq, rwr.wr_id);
else
ret = hns_roce_v2_post_recv(ibqp, &rwr, &bad_rwr);
if (ret) {
ibdev_err_ratelimited(&hr_dev->ib_dev,
"failed to post recv for drain rq, ret = %d.\n",
ret);
return;
}
handle_drain_completion(cq, &rdrain, hr_dev);
}
static void hns_roce_v2_drain_sq(struct ib_qp *ibqp)
{
struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
struct hns_roce_drain_cqe sdrain = {};
const struct ib_send_wr *bad_swr;
struct ib_cq *cq = ibqp->send_cq;
struct ib_rdma_wr swr = {
.wr = {
.next = NULL,
{ .wr_cqe = &sdrain.cqe, },
.opcode = IB_WR_RDMA_WRITE,
},
};
int ret;
ret = ib_modify_qp(ibqp, &attr, IB_QP_STATE);
if (ret && hr_dev->state < HNS_ROCE_DEVICE_STATE_RST_DOWN) {
ibdev_err_ratelimited(&hr_dev->ib_dev,
"failed to modify qp during drain sq, ret = %d.\n",
ret);
return;
}
sdrain.cqe.done = hns_roce_drain_qp_done;
init_completion(&sdrain.done);
if (hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN)
ret = hns_roce_push_drain_wr(&hr_qp->sq, cq, swr.wr.wr_id);
else
ret = hns_roce_v2_post_send(ibqp, &swr.wr, &bad_swr);
if (ret) {
ibdev_err_ratelimited(&hr_dev->ib_dev,
"failed to post send for drain sq, ret = %d.\n",
ret);
return;
}
handle_drain_completion(cq, &sdrain, hr_dev);
}
static void *get_srq_wqe_buf(struct hns_roce_srq *srq, u32 n)
{
return hns_roce_buf_offset(srq->buf_mtr.kmem, n << srq->wqe_shift);
@ -7040,6 +7204,8 @@ static const struct ib_device_ops hns_roce_v2_dev_ops = {
.post_send = hns_roce_v2_post_send,
.query_qp = hns_roce_v2_query_qp,
.req_notify_cq = hns_roce_v2_req_notify_cq,
.drain_rq = hns_roce_v2_drain_rq,
.drain_sq = hns_roce_v2_drain_sq,
};
static const struct ib_device_ops hns_roce_v2_dev_srq_ops = {