Skip to contentSkip to conversation

bpf: cpumap: report queue_index to xdp_rxq_info

RFC PATCH In review Jose A. Perez de Azpillaga 3 participants 4 messages Apr 11

Subsystems (5)
Networking Network drivers BPF XDP +1 more subsystems
lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 bpf: cpumap: report queue_index to xdp_rxq_info

When a packet is redirected to a CPU map entry, cpu_map_bpf_prog_run_xdp() reconstructs a minimal xdp_rxq_info from xdp_frame fields (dev_rx and mem_type) before re-running the BPF program on the target CPU. However, queue_index was never preserved across the CPU boundary, so BPF programs running in cpumap context always observe ctx->rx_queue_index == 0, regardless of which hardware queue originally received the packet.

Fix this by storing the originating queue_index in struct xdp_frame, following the same pattern already established for dev_rx and mem_type. The field is populated from rxq->queue_index in xdp_convert_buff_to_frame() during NAPI context, when the rxq_info is still valid, and restored into the reconstructed rxq_info in cpu_map_bpf_prog_run_xdp().

Also use xdpf->queue_index in __xdp_build_skb_from_frame() to call skb_record_rx_queue(), which was previously listed as missing information in that function's comment.

Also propagate queue_index in dpaa_a050385_wa_xdpf(), which manually constructs a new xdp_frame from an uninitialized page. Without this, queue_index would contain stale data from the page allocator.

---
Note: this patch was only compiled, not tested.

 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 1 +
 include/net/xdp.h                              | 4 +++-
 kernel/bpf/cpumap.c                            | 2 +-
 net/core/xdp.c                                 | 2 +-
 4 files changed, 6 insertions(+), 3 deletions(-)
drivers/net/ethernet/freescale/dpaa/dpaa_eth.c +1 -0
index 3edc8d142dd5..00e36b0ac74d 100644
@@ -2281,6 +2281,7 @@ static int dpaa_a050385_wa_xdpf(struct dpaa_priv *priv,
 	new_xdpf->headroom = priv->tx_headroom;
 	new_xdpf->frame_sz = DPAA_BP_RAW_SIZE;
 	new_xdpf->mem_type = MEM_TYPE_PAGE_ORDER0;
+	new_xdpf->queue_index = xdpf->queue_index;
 
 	/* Release the initial buffer */
 	xdp_return_frame_rx_napi(xdpf);
include/net/xdp.h +3 -1
index aa742f413c35..6db10e6a8864 100644
@@ -297,10 +297,11 @@ struct xdp_frame {
 	u32 headroom;
 	u32 metasize; /* uses lower 8-bits */
 	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
-	 * while mem_type is valid on remote CPU.
+	 * while mem_type and queue_index are valid on remote CPU.
 	 */
 	enum xdp_mem_type mem_type:32;
 	struct net_device *dev_rx; /* used by cpumap */
+	u32 queue_index; /* used by cpumap */
 	u32 frame_sz;
 	u32 flags; /* supported values defined in xdp_buff_flags */
 };
@@ -441,6 +442,7 @@ struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
 
 	/* rxq only valid until napi_schedule ends, convert to xdp_mem_type */
 	xdp_frame->mem_type = xdp->rxq->mem.type;
+	xdp_frame->queue_index = xdp->rxq->queue_index;
 
 	return xdp_frame;
 }
kernel/bpf/cpumap.c +1 -1
index 5e59ab896f05..448da572de9a 100644
@@ -197,7 +197,7 @@ static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 
 		rxq.dev = xdpf->dev_rx;
 		rxq.mem.type = xdpf->mem_type;
-		/* TODO: report queue_index to xdp_rxq_info */
+		rxq.queue_index = xdpf->queue_index;
 
 		xdp_convert_frame_to_buff(xdpf, &xdp);
 
net/core/xdp.c +1 -1
index 9890a30584ba..326e3057ed7f 100644
@@ -829,11 +829,11 @@ struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
 
 	/* Essential SKB info: protocol and skb->dev */
 	skb->protocol = eth_type_trans(skb, dev);
+	skb_record_rx_queue(skb, xdpf->queue_index);
 
 	/* Optional SKB info, currently missing:
 	 * - HW checksum info		(skb->ip_summed)
 	 * - HW RX hash			(skb_set_hash)
-	 * - RX ring dev queue index	(skb_record_rx_queue)
 	 */
 
 	if (xdpf->mem_type == MEM_TYPE_PAGE_POOL)

--
2.53.0

#1 lore raw
Quoted 76 lines · Jose A. Perez de Azpillaga

When a packet is redirected to a CPU map entry, cpu_map_bpf_prog_run_xdp() reconstructs a minimal xdp_rxq_info from xdp_frame fields (dev_rx and mem_type) before re-running the BPF program on the target CPU. However, queue_index was never preserved across the CPU boundary, so BPF programs running in cpumap context always observe ctx->rx_queue_index == 0, regardless of which hardware queue originally received the packet.

Fix this by storing the originating queue_index in struct xdp_frame, following the same pattern already established for dev_rx and mem_type. The field is populated from rxq->queue_index in xdp_convert_buff_to_frame() during NAPI context, when the rxq_info is still valid, and restored into the reconstructed rxq_info in cpu_map_bpf_prog_run_xdp().

Also use xdpf->queue_index in __xdp_build_skb_from_frame() to call skb_record_rx_queue(), which was previously listed as missing information in that function's comment.

Also propagate queue_index in dpaa_a050385_wa_xdpf(), which manually constructs a new xdp_frame from an uninitialized page. Without this, queue_index would contain stale data from the page allocator.

---
Note: this patch was only compiled, not tested.

 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 1 +
 include/net/xdp.h                              | 4 +++-
 kernel/bpf/cpumap.c                            | 2 +-
 net/core/xdp.c                                 | 2 +-
 4 files changed, 6 insertions(+), 3 deletions(-)
drivers/net/ethernet/freescale/dpaa/dpaa_eth.c +1 -0
index 3edc8d142dd5..00e36b0ac74d 100644
@@ -2281,6 +2281,7 @@ static int dpaa_a050385_wa_xdpf(struct dpaa_priv *priv,
        new_xdpf->headroom = priv->tx_headroom;
        new_xdpf->frame_sz = DPAA_BP_RAW_SIZE;
        new_xdpf->mem_type = MEM_TYPE_PAGE_ORDER0;
+       new_xdpf->queue_index = xdpf->queue_index;
 
        /* Release the initial buffer */
        xdp_return_frame_rx_napi(xdpf);
include/net/xdp.h +3 -1
index aa742f413c35..6db10e6a8864 100644
@@ -297,10 +297,11 @@ struct xdp_frame {
        u32 headroom;
        u32 metasize; /* uses lower 8-bits */
        /* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
-        * while mem_type is valid on remote CPU.
+        * while mem_type and queue_index are valid on remote CPU.
         */
        enum xdp_mem_type mem_type:32;
        struct net_device *dev_rx; /* used by cpumap */
+       u32 queue_index; /* used by cpumap */
        u32 frame_sz;
        u32 flags; /* supported values defined in xdp_buff_flags */
 };
@@ -441,6 +442,7 @@ struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
 
        /* rxq only valid until napi_schedule ends, convert to xdp_mem_type */
        xdp_frame->mem_type = xdp->rxq->mem.type;
+       xdp_frame->queue_index = xdp->rxq->queue_index;
 
        return xdp_frame;
 }
kernel/bpf/cpumap.c +0 -0
index 5e59ab896f05..448da572de9a 100644

kernel/bpf/cpumap.c static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu, · line 200

Quoted 5 lines
kernel/bpf/cpumap.c +1 -1
@@ -197,7 +197,7 @@ static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 
                rxq.dev = xdpf->dev_rx;
                rxq.mem.type = xdpf->mem_type;
-               /* TODO: report queue_index to xdp_rxq_info */
+               rxq.queue_index = xdpf->queue_index;

This is like 5th time people attempt to address this TODO.

Just remove that comment. Don't send broken patches.

pw-bot: cr

#2 lore raw
Quoted 79 lines · Alexei Starovoitov
Quoted 77 lines · Jose A. Perez de Azpillaga

When a packet is redirected to a CPU map entry, cpu_map_bpf_prog_run_xdp() reconstructs a minimal xdp_rxq_info from xdp_frame fields (dev_rx and mem_type) before re-running the BPF program on the target CPU. However, queue_index was never preserved across the CPU boundary, so BPF programs running in cpumap context always observe ctx->rx_queue_index == 0, regardless of which hardware queue originally received the packet.

Fix this by storing the originating queue_index in struct xdp_frame, following the same pattern already established for dev_rx and mem_type. The field is populated from rxq->queue_index in xdp_convert_buff_to_frame() during NAPI context, when the rxq_info is still valid, and restored into the reconstructed rxq_info in cpu_map_bpf_prog_run_xdp().

Also use xdpf->queue_index in __xdp_build_skb_from_frame() to call skb_record_rx_queue(), which was previously listed as missing information in that function's comment.

Also propagate queue_index in dpaa_a050385_wa_xdpf(), which manually constructs a new xdp_frame from an uninitialized page. Without this, queue_index would contain stale data from the page allocator.

---
Note: this patch was only compiled, not tested.

 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 1 +
 include/net/xdp.h                              | 4 +++-
 kernel/bpf/cpumap.c                            | 2 +-
 net/core/xdp.c                                 | 2 +-
 4 files changed, 6 insertions(+), 3 deletions(-)
drivers/net/ethernet/freescale/dpaa/dpaa_eth.c +1 -0
index 3edc8d142dd5..00e36b0ac74d 100644
@@ -2281,6 +2281,7 @@ static int dpaa_a050385_wa_xdpf(struct dpaa_priv *priv,
        new_xdpf->headroom = priv->tx_headroom;
        new_xdpf->frame_sz = DPAA_BP_RAW_SIZE;
        new_xdpf->mem_type = MEM_TYPE_PAGE_ORDER0;
+       new_xdpf->queue_index = xdpf->queue_index;
 
        /* Release the initial buffer */
        xdp_return_frame_rx_napi(xdpf);
include/net/xdp.h +3 -1
index aa742f413c35..6db10e6a8864 100644
@@ -297,10 +297,11 @@ struct xdp_frame {
        u32 headroom;
        u32 metasize; /* uses lower 8-bits */
        /* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
-        * while mem_type is valid on remote CPU.
+        * while mem_type and queue_index are valid on remote CPU.
         */
        enum xdp_mem_type mem_type:32;
        struct net_device *dev_rx; /* used by cpumap */
+       u32 queue_index; /* used by cpumap */
        u32 frame_sz;
        u32 flags; /* supported values defined in xdp_buff_flags */
 };
@@ -441,6 +442,7 @@ struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
 
        /* rxq only valid until napi_schedule ends, convert to xdp_mem_type */
        xdp_frame->mem_type = xdp->rxq->mem.type;
+       xdp_frame->queue_index = xdp->rxq->queue_index;
 
        return xdp_frame;
 }
kernel/bpf/cpumap.c +0 -0
index 5e59ab896f05..448da572de9a 100644
@@ -197,7 +197,7 @@ static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 
Quoted 8 lines

kernel/bpf/cpumap.c static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu, · line 200

Quoted 4 lines
kernel/bpf/cpumap.c +1 -1
                rxq.dev = xdpf->dev_rx;
                rxq.mem.type = xdpf->mem_type;
-               /* TODO: report queue_index to xdp_rxq_info */
+               rxq.queue_index = xdpf->queue_index;

This is like 5th time people attempt to address this TODO.

Just remove that comment. Don't send broken patches.

oh... okay. but I have a question, since the bot detected something I didn't and queue_index should be propagated in xdp_convert_zc_to_xdp_frame(), or maybe intentional?

is it better to do as you said, removing the comment, or doing what the bot said with proper test?

--
regards,
jose a. p-a

#3 lore raw
In reply to Jose A. Perez de Azpillaga
Is the updated comment accurate for all paths?
Quoted 6 lines
include/net/xdp.h +0 -0
@@ -297,10 +297,11 @@ struct xdp_frame {
 	u32 headroom;
 	u32 metasize; /* uses lower 8-bits */

include/net/xdp.h struct xdp_frame { · line 307

Quoted 10 lines
include/net/xdp.h +2 -1
 	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
-	 * while mem_type is valid on remote CPU.
+	 * while mem_type and queue_index are valid on remote CPU.
 	 */
 	enum xdp_mem_type mem_type:32;
 	struct net_device *dev_rx; /* used by cpumap */
+	u32 queue_index; /* used by cpumap */
 	u32 frame_sz;
 	u32 flags; /* supported values defined in xdp_buff_flags */
 };

Is the updated comment accurate for all paths?

xdp_convert_buff_to_frame() has an early return for the zero-copy path:

    if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL)
        return xdp_convert_zc_to_xdp_frame(xdp);

xdp_convert_zc_to_xdp_frame() zeroes the frame with memset and explicitly sets mem_type but never sets queue_index:

net/core/xdp.c:xdp_convert_zc_to_xdp_frame() {
    ...
    memset(xdpf, 0, sizeof(*xdpf));
    ...
    xdpf->mem_type = MEM_TYPE_PAGE_ORDER0;
    /* queue_index is not set -- remains 0 from memset */
    ...
}

The comment says queue_index is "valid on remote CPU" but for zero-copy converted frames it will always be 0 regardless of which hardware queue originated the packet.

include/net/xdp.h struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp) · line 447

Quoted 7 lines
include/net/xdp.h +1 -0
@@ -441,6 +442,7 @@ struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
 
 	/* rxq only valid until napi_schedule ends, convert to xdp_mem_type */
 	xdp_frame->mem_type = xdp->rxq->mem.type;
+	xdp_frame->queue_index = xdp->rxq->queue_index;
 
 	return xdp_frame;

This assignment is only reached for the non-ZC path. Should xdp_convert_zc_to_xdp_frame() also set queue_index from xdp->rxq->queue_index to match?

---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

CI run summary: https://github.com/kernel-patches/bpf/actions/runs/24288456014