Re: [PATCH v2 3/4] RDMA/rxe: support perf mgmt GET method

[Zhu Yanjun <yanjun.zhu@linux.dev>]

在 2026/3/28 19:55, zhenwei pi 写道:
> In RXE, hardware counters are already supported, but not in a
> standardized manner. For instance, user-space monitoring tools like
> atop only read from the *counters* directory. Therefore, it is
> necessary to add perf management support to RXE.
> 
> Signed-off-by: zhenwei pi <zhenwei.pi@linux.dev>
> ---
>   drivers/infiniband/sw/rxe/Makefile    |  1 +
>   drivers/infiniband/sw/rxe/rxe_loc.h   |  6 ++
>   drivers/infiniband/sw/rxe/rxe_mad.c   | 93 +++++++++++++++++++++++++++
>   drivers/infiniband/sw/rxe/rxe_verbs.c |  1 +
>   drivers/infiniband/sw/rxe/rxe_verbs.h |  5 ++
>   5 files changed, 106 insertions(+)
>   create mode 100644 drivers/infiniband/sw/rxe/rxe_mad.c
> 
> diff --git a/drivers/infiniband/sw/rxe/Makefile b/drivers/infiniband/sw/rxe/Makefile
> index 93134f1d1d0c..3c47e5b982c2 100644
> --- a/drivers/infiniband/sw/rxe/Makefile
> +++ b/drivers/infiniband/sw/rxe/Makefile
> @@ -22,6 +22,7 @@ rdma_rxe-y := \
>   	rxe_mcast.o \
>   	rxe_task.o \
>   	rxe_net.o \
> +	rxe_mad.o \
>   	rxe_hw_counters.o
>   
>   rdma_rxe-$(CONFIG_INFINIBAND_ON_DEMAND_PAGING) += rxe_odp.o
> diff --git a/drivers/infiniband/sw/rxe/rxe_loc.h b/drivers/infiniband/sw/rxe/rxe_loc.h
> index 7992290886e1..a8ce85147c1f 100644
> --- a/drivers/infiniband/sw/rxe/rxe_loc.h
> +++ b/drivers/infiniband/sw/rxe/rxe_loc.h
> @@ -245,4 +245,10 @@ static inline int rxe_ib_advise_mr(struct ib_pd *pd,
>   
>   #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
>   
> +/* rxe-mad.c */
> +int rxe_process_mad(struct ib_device *ibdev, int mad_flags, u32 port_num,
> +		    const struct ib_wc *in_wc, const struct ib_grh *in_grh,
> +		    const struct ib_mad *in, struct ib_mad *out,
> +		    size_t *out_mad_size, u16 *out_mad_pkey_index);
> +
>   #endif /* RXE_LOC_H */
> diff --git a/drivers/infiniband/sw/rxe/rxe_mad.c b/drivers/infiniband/sw/rxe/rxe_mad.c
> new file mode 100644
> index 000000000000..5e0567806c02
> --- /dev/null
> +++ b/drivers/infiniband/sw/rxe/rxe_mad.c
> @@ -0,0 +1,93 @@
> +// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
> +/*
> + * Copyright (c) 2026 zhenwei pi <zhenwei.pi@linux.dev>
> + */
> +
> +#include <rdma/ib_pma.h>
> +#include "rxe.h"
> +#include "rxe_hw_counters.h"
> +
> +static int rxe_get_pma_info(struct ib_mad *out)
> +{
> +	struct ib_class_port_info cpi = {};
> +
> +	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
> +	memcpy((out->data + 40), &cpi, sizeof(cpi));
> +
> +	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
> +}
> +
> +static int rxe_get_pma_counters(struct rxe_dev *rxe, struct ib_mad *out)
> +{
> +	struct ib_pma_portcounters *pma_cnt = (struct ib_pma_portcounters *)(out->data + 40);
> +	s64 val;
> +
> +	/* IBA release 1.8, 16.1.3.5: During operation, instead of overflowing, they shall stop
> +	 * at all ones.
> +	 */
> +	val = rxe_counter_get(rxe, RXE_CNT_LINK_DOWNED);
> +	if (val > U8_MAX)
> +		pma_cnt->link_downed_counter = U8_MAX;
> +	else
> +		pma_cnt->link_downed_counter = (u8)val;
> +
> +	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
> +}
> +
> +static int rxe_get_pma_counters_ext(struct rxe_dev *rxe, struct ib_mad *out)
> +{
> +	struct ib_pma_portcounters_ext *pma_cnt_ext =
> +		(struct ib_pma_portcounters_ext *)(out->data + 40);
> +
> +	pma_cnt_ext->port_xmit_data = cpu_to_be64(rxe_counter_get(rxe, RXE_CNT_SENT_BYTES) >> 2);
> +	pma_cnt_ext->port_rcv_data = cpu_to_be64(rxe_counter_get(rxe, RXE_CNT_RCVD_BYTES) >> 2);
> +	pma_cnt_ext->port_xmit_packets = cpu_to_be64(rxe_counter_get(rxe, RXE_CNT_SENT_PKTS));
> +	pma_cnt_ext->port_rcv_packets = cpu_to_be64(rxe_counter_get(rxe, RXE_CNT_RCVD_PKTS));
> +
> +	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
> +}
> +
> +static int rxe_get_perf_mgmt(struct rxe_dev *rxe, const struct ib_mad *in,
> +				 struct ib_mad *out)
> +{
> +	switch (in->mad_hdr.attr_id) {
> +	case IB_PMA_CLASS_PORT_INFO:
> +		return rxe_get_pma_info(out);
> +
> +	case IB_PMA_PORT_COUNTERS:
> +		return rxe_get_pma_counters(rxe, out);
> +
> +	case IB_PMA_PORT_COUNTERS_EXT:
> +		return rxe_get_pma_counters_ext(rxe, out);
> +
> +	default:
> +		break;
> +	}
> +
> +	return IB_MAD_RESULT_FAILURE;

In rxe_get_perf_mgmt, the default case returns IB_MAD_RESULT_FAILURE.

Returning FAILURE in the MAD stack often results in the packet being 
dropped without a response. This causes management tools (like ibstat or 
perfquery) to hang until a timeout occurs.

According to IBA specs, the agent should return a reply with a proper 
error status.

Return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY and set 
out->mad_hdr.status = cpu_to_be16(IB_MGMT_MAD_STATUS_BAD_ATTR);.

The source code in default branch should be:

"
default:
out->mad_hdr.status = cpu_to_be16(IB_MGMT_MAD_STATUS_BAD_ATTR);
return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
"

> +}
> +
> +int rxe_process_mad(struct ib_device *ibdev, int mad_flags, u32 port_num,
> +		    const struct ib_wc *in_wc, const struct ib_grh *in_grh,
> +		    const struct ib_mad *in, struct ib_mad *out,
> +		    size_t *out_mad_size, u16 *out_mad_pkey_index)
> +{
> +	struct rxe_dev *rxe = to_rdev(ibdev);
> +	u8 mgmt_class = in->mad_hdr.mgmt_class;
> +	u8 method = in->mad_hdr.method;
> +
> +	if (port_num != RXE_PORT)
> +		return IB_MAD_RESULT_FAILURE;
> +

In the IB Management Datagram (MAD) spec, many Reserved fields or unused 
counter fields are strictly required to be zero-filled.

By using memset(out, 0, sizeof(*out)), you ensure that all fields in the 
response packet—including those not currently supported by the 
driver—have a predefined initial value. This prevents the transmission 
of packets containing "random noise" (residual kernel data) to the 
querying entity.

Add "memset(out, 0, sizeof(*out));" here.

Zhu Yanjun

> +	switch (mgmt_class) {
> +	case IB_MGMT_CLASS_PERF_MGMT:
> +		if (method == IB_MGMT_METHOD_GET)
> +			return rxe_get_perf_mgmt(rxe, in, out);
> +		break;
> +
> +	default:
> +		break;
> +	}
> +
> +	return IB_MAD_RESULT_FAILURE;
> +}
> diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.c b/drivers/infiniband/sw/rxe/rxe_verbs.c
> index bcd486e8668b..7df0cb5a09a3 100644
> --- a/drivers/infiniband/sw/rxe/rxe_verbs.c
> +++ b/drivers/infiniband/sw/rxe/rxe_verbs.c
> @@ -1509,6 +1509,7 @@ static const struct ib_device_ops rxe_dev_ops = {
>   	.post_recv = rxe_post_recv,
>   	.post_send = rxe_post_send,
>   	.post_srq_recv = rxe_post_srq_recv,
> +	.process_mad = rxe_process_mad,
>   	.query_ah = rxe_query_ah,
>   	.query_device = rxe_query_device,
>   	.query_pkey = rxe_query_pkey,
> diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.h b/drivers/infiniband/sw/rxe/rxe_verbs.h
> index 2bcfb919a40b..1c4fa8eaa733 100644
> --- a/drivers/infiniband/sw/rxe/rxe_verbs.h
> +++ b/drivers/infiniband/sw/rxe/rxe_verbs.h
> @@ -466,6 +466,11 @@ static inline void rxe_counter_add(struct rxe_dev *rxe, enum rxe_counters index,
>   	atomic64_add(val, &rxe->stats_counters[index]);
>   }
>   
> +static inline s64 rxe_counter_get(struct rxe_dev *rxe, enum rxe_counters index)
> +{
> +	return atomic64_read(&rxe->stats_counters[index]);
> +}
> +
>   static inline struct rxe_dev *to_rdev(struct ib_device *dev)
>   {
>   	return dev ? container_of(dev, struct rxe_dev, ib_dev) : NULL;