Skip to contentSkip to conversation

kni: fix possible kernel crash with va2pa

PATCH Superseded Yangchao Zhou 6 participants 17 messages Feb 28, 2019 – Jul 15, 2019

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kni: fix possible kernel crash with va2pa

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

---
 kernel/linux/kni/kni_net.c                       | 16 ++--------------
 .../eal/include/exec-env/rte_kni_common.h        |  4 ++++
 lib/librte_kni/rte_kni.c                         | 15 ++++++++++++++-
 3 files changed, 20 insertions(+), 15 deletions(-)
kernel/linux/kni/kni_net.c +2 -14
index 7371b6d58..caef8754f 100644
@@ -61,18 +61,6 @@ kva2data_kva(struct rte_kni_mbuf *m)
 	return phys_to_virt(m->buf_physaddr + m->data_off);
 }
 
-/* virtual address to physical address */
-static void *
-va2pa(void *va, struct rte_kni_mbuf *m)
-{
-	void *pa;
-
-	pa = (void *)((unsigned long)va -
-			((unsigned long)m->buf_addr -
-			 (unsigned long)m->buf_physaddr));
-	return pa;
-}
-
 /*
  * It can be called to process the request.
  */
@@ -363,7 +351,7 @@ kni_net_rx_normal(struct kni_dev *kni)
 				if (!kva->next)
 					break;
 
-				kva = pa2kva(va2pa(kva->next, kva));
+				kva = pa2kva(kva->next_pa);
 				data_kva = kva2data_kva(kva);
 			}
 		}
@@ -545,7 +533,7 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 				if (!kva->next)
 					break;
 
-				kva = pa2kva(va2pa(kva->next, kva));
+				kva = pa2kva(kva->next_pa);
 				data_kva = kva2data_kva(kva);
 			}
 		}
lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h +4 -0
index 5afa08713..608f5c13f 100644
@@ -87,6 +87,10 @@ struct rte_kni_mbuf {
 	char pad3[8] __attribute__((__aligned__(RTE_CACHE_LINE_MIN_SIZE)));
 	void *pool;
 	void *next;
+	union {
+		uint64_t tx_offload;
+		void *next_pa;      /**< Physical address of next mbuf. */
+	};
 };
 
 /*
lib/librte_kni/rte_kni.c +14 -1
index 73aeccccf..1aaebcfa1 100644
@@ -353,6 +353,17 @@ va2pa(struct rte_mbuf *m)
 			 (unsigned long)m->buf_iova));
 }
 
+static void *
+va2pa_all(struct rte_mbuf *m)
+{
+	struct rte_kni_mbuf *mbuf = (struct rte_kni_mbuf *)m;
+	while (mbuf->next) {
+		mbuf->next_pa = va2pa(mbuf->next);
+		mbuf = mbuf->next;
+	}
+	return va2pa(m);
+}
+
 static void
 obj_free(struct rte_mempool *mp __rte_unused, void *opaque, void *obj,
 		unsigned obj_idx __rte_unused)
@@ -550,7 +561,7 @@ rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
 	unsigned int i;
 
 	for (i = 0; i < num; i++)
-		phy_mbufs[i] = va2pa(mbufs[i]);
+		phy_mbufs[i] = va2pa_all(mbufs[i]);
 
 	ret = kni_fifo_put(kni->rx_q, phy_mbufs, num);
 
@@ -607,6 +618,8 @@ kni_allocate_mbufs(struct rte_kni *kni)
 			 offsetof(struct rte_kni_mbuf, pkt_len));
 	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
 			 offsetof(struct rte_kni_mbuf, ol_flags));
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, tx_offload) !=
+			 offsetof(struct rte_kni_mbuf, tx_offload));
 
 	/* Check if pktmbuf pool has been configured */
 	if (kni->pktmbuf_pool == NULL) {
Signature · 1 line
2.19.1.windows.1
#1 lore raw
Quoted 3 lines · Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

Hi Yangchao,

The problem you described seems valid, when current mbuf and the mbuf pointed bu next pointer from different (huge)pages, address calculation will be wrong.

Can you able to reproduce the issue, or recognized the problem theoretically?

Quoted 54 lines
---
 kernel/linux/kni/kni_net.c                       | 16 ++--------------
 .../eal/include/exec-env/rte_kni_common.h        |  4 ++++
 lib/librte_kni/rte_kni.c                         | 15 ++++++++++++++-
 3 files changed, 20 insertions(+), 15 deletions(-)
kernel/linux/kni/kni_net.c +2 -14
index 7371b6d58..caef8754f 100644
@@ -61,18 +61,6 @@ kva2data_kva(struct rte_kni_mbuf *m)
 	return phys_to_virt(m->buf_physaddr + m->data_off);
 }
 
-/* virtual address to physical address */
-static void *
-va2pa(void *va, struct rte_kni_mbuf *m)
-{
-	void *pa;
-
-	pa = (void *)((unsigned long)va -
-			((unsigned long)m->buf_addr -
-			 (unsigned long)m->buf_physaddr));
-	return pa;
-}
-
 /*
  * It can be called to process the request.
  */
@@ -363,7 +351,7 @@ kni_net_rx_normal(struct kni_dev *kni)
 				if (!kva->next)
 					break;
 
-				kva = pa2kva(va2pa(kva->next, kva));
+				kva = pa2kva(kva->next_pa);
 				data_kva = kva2data_kva(kva);
 			}
 		}
@@ -545,7 +533,7 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 				if (!kva->next)
 					break;
 
-				kva = pa2kva(va2pa(kva->next, kva));
+				kva = pa2kva(kva->next_pa);
 				data_kva = kva2data_kva(kva);
 			}
 		}
lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h +0 -0
index 5afa08713..608f5c13f 100644

lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h struct rte_kni_mbuf { · line 93

Quoted 7 lines
lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h +4 -0
@@ -87,6 +87,10 @@ struct rte_kni_mbuf {
 	char pad3[8] __attribute__((__aligned__(RTE_CACHE_LINE_MIN_SIZE)));
 	void *pool;
 	void *next;
+	union {
+		uint64_t tx_offload;
+		void *next_pa;      /**< Physical address of next mbuf. */
+	};

This will cause overwrite the 'tx_offload' via 'next_pa', we don't use tx_offload in KNI but not sure about removing potential use for future.

What do you think about converting 'm->next' to physical address before putting them into 'rx_q', and in kernel side after data copied to skb convert 'm->next' back to virtual address before putting it into 'free_q' ?
I think both address conversion can be possible to do, a little tricky because address conversion should be calculated in next mbuf and previous mbuf->next in the chain should be updated.

Quoted 44 lines
lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h +0 -0
 };
 
 /*
lib/librte_kni/rte_kni.c +14 -1
index 73aeccccf..1aaebcfa1 100644
@@ -353,6 +353,17 @@ va2pa(struct rte_mbuf *m)
 			 (unsigned long)m->buf_iova));
 }
 
+static void *
+va2pa_all(struct rte_mbuf *m)
+{
+	struct rte_kni_mbuf *mbuf = (struct rte_kni_mbuf *)m;
+	while (mbuf->next) {
+		mbuf->next_pa = va2pa(mbuf->next);
+		mbuf = mbuf->next;
+	}
+	return va2pa(m);
+}
+
 static void
 obj_free(struct rte_mempool *mp __rte_unused, void *opaque, void *obj,
 		unsigned obj_idx __rte_unused)
@@ -550,7 +561,7 @@ rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
 	unsigned int i;
 
 	for (i = 0; i < num; i++)
-		phy_mbufs[i] = va2pa(mbufs[i]);
+		phy_mbufs[i] = va2pa_all(mbufs[i]);
 
 	ret = kni_fifo_put(kni->rx_q, phy_mbufs, num);
 
@@ -607,6 +618,8 @@ kni_allocate_mbufs(struct rte_kni *kni)
 			 offsetof(struct rte_kni_mbuf, pkt_len));
 	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
 			 offsetof(struct rte_kni_mbuf, ol_flags));
+	RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, tx_offload) !=
+			 offsetof(struct rte_kni_mbuf, tx_offload));
 
 	/* Check if pktmbuf pool has been configured */
 	if (kni->pktmbuf_pool == NULL) {
#2 lore raw

Was there any update to this patch , I am also seeing kernel crash in kni_net_rx_normal dueing skb_put which is happening for chained mbufs.

--
Regards,
Souvik

From: dev <[email protected]> On Behalf Of Ferruh Yigit
Sent: Wednesday, March 6, 2019 12:31 PM
To: Yangchao Zhou <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH] kni: fix possible kernel crash with va2pa

________________________________
NOTICE: This email was received from an EXTERNAL sender
________________________________

Quoted 3 lines · Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

Hi Yangchao,

The problem you described seems valid, when current mbuf and the mbuf pointed bu next pointer from different (huge)pages, address calculation will be wrong.

Can you able to reproduce the issue, or recognized the problem theoretically?

Quoted 54 lines

---
kernel/linux/kni/kni_net.c | 16 ++-------------- .../eal/include/exec-env/rte_kni_common.h | 4 ++++ lib/librte_kni/rte_kni.c | 15 ++++++++++++++-

3 files changed, 20 insertions(+), 15 deletions(-)
kernel/linux/kni/kni_net.c +2 -14
index 7371b6d58..caef8754f 100644
@@ -61,18 +61,6 @@ kva2data_kva(struct rte_kni_mbuf *m)
 return phys_to_virt(m->buf_physaddr + m->data_off);
 }
 
-/* virtual address to physical address */
-static void *
-va2pa(void *va, struct rte_kni_mbuf *m)
-{
- void *pa;
-
- pa = (void *)((unsigned long)va -
- ((unsigned long)m->buf_addr -
- (unsigned long)m->buf_physaddr));
- return pa;
-}
-
 /*
 * It can be called to process the request.
 */
@@ -363,7 +351,7 @@ kni_net_rx_normal(struct kni_dev *kni)
 if (!kva->next)
 break;
 
- kva = pa2kva(va2pa(kva->next, kva));
+ kva = pa2kva(kva->next_pa);
 data_kva = kva2data_kva(kva);
 }
 }
@@ -545,7 +533,7 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 if (!kva->next)
 break;
 
- kva = pa2kva(va2pa(kva->next, kva));
+ kva = pa2kva(kva->next_pa);
 data_kva = kva2data_kva(kva);
 }
 }
lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h +0 -0
index 5afa08713..608f5c13f 100644

lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h struct rte_kni_mbuf { · line 93

Quoted 7 lines
lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h +4 -0
@@ -87,6 +87,10 @@ struct rte_kni_mbuf {
 char pad3[8] __attribute__((__aligned__(RTE_CACHE_LINE_MIN_SIZE)));
 void *pool;
 void *next;
+ union {
+ uint64_t tx_offload;
+ void *next_pa; /**< Physical address of next mbuf. */
+ };

This will cause overwrite the 'tx_offload' via 'next_pa', we don't use tx_offload in KNI but not sure about removing potential use for future.

What do you think about converting 'm->next' to physical address before putting them into 'rx_q', and in kernel side after data copied to skb convert 'm->next' back to virtual address before putting it into 'free_q' ?
I think both address conversion can be possible to do, a little tricky because address conversion should be calculated in next mbuf and previous mbuf->next in the chain should be updated.

Quoted 41 lines
lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h +0 -0
 };
 
 /*
lib/librte_kni/rte_kni.c +11 -0
index 73aeccccf..1aaebcfa1 100644
@@ -353,6 +353,17 @@ va2pa(struct rte_mbuf *m)
 (unsigned long)m->buf_iova));
 }
 
+static void *
+va2pa_all(struct rte_mbuf *m)
+{
+ struct rte_kni_mbuf *mbuf = (struct rte_kni_mbuf *)m;
+ while (mbuf->next) {
+ mbuf->next_pa = va2pa(mbuf->next);
+ mbuf = mbuf->next;
+ }
+ return va2pa(m);
+}
+

static void
obj_free(struct rte_mempool *mp __rte_unused, void *opaque, void *obj, unsigned obj_idx __rte_unused)

lib/librte_kni/rte_kni.c +1 -1
@@ -550,7 +561,7 @@ rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
 unsigned int i;
 
 for (i = 0; i < num; i++)
- phy_mbufs[i] = va2pa(mbufs[i]);
+ phy_mbufs[i] = va2pa_all(mbufs[i]);
 
 ret = kni_fifo_put(kni->rx_q, phy_mbufs, num);
 
@@ -607,6 +618,8 @@ kni_allocate_mbufs(struct rte_kni *kni)
 offsetof(struct rte_kni_mbuf, pkt_len));

RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) != offsetof(struct rte_kni_mbuf, ol_flags));
+ RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, tx_offload) !=
+ offsetof(struct rte_kni_mbuf, tx_offload));

Quoted 3 lines

/* Check if pktmbuf pool has been configured */
if (kni->pktmbuf_pool == NULL) {

-----------------------------------------------------------------------------------------------------------------------
Notice: This e-mail together with any attachments may contain information of Ribbon Communications Inc. that is confidential and/or proprietary for the sole use of the intended recipient. Any review, disclosure, reliance or distribution by others or forwarding without express permission is strictly prohibited. If you are not the intended recipient, please notify the sender immediately and then delete all copies, including any attachments.
-----------------------------------------------------------------------------------------------------------------------

#3 lore raw

PATCH v2 1/1 kni: fix possible kernel crash with va2pa

In reply to Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

---
v2: Add an explanation that causes this problem.

    Use m->next to store physical address.
---
 kernel/linux/kni/kni_net.c                    | 42 +++++++++++--------
 .../eal/include/exec-env/rte_kni_common.h     |  2 +-
 lib/librte_kni/rte_kni.c                      | 15 ++++++-
 3 files changed, 40 insertions(+), 19 deletions(-)
kernel/linux/kni/kni_net.c +25 -17
index 7371b6d58..106b5153f 100644
@@ -61,18 +61,6 @@ kva2data_kva(struct rte_kni_mbuf *m)
 	return phys_to_virt(m->buf_physaddr + m->data_off);
 }
 
-/* virtual address to physical address */
-static void *
-va2pa(void *va, struct rte_kni_mbuf *m)
-{
-	void *pa;
-
-	pa = (void *)((unsigned long)va -
-			((unsigned long)m->buf_addr -
-			 (unsigned long)m->buf_physaddr));
-	return pa;
-}
-
 /*
  * It can be called to process the request.
  */
@@ -173,7 +161,10 @@ kni_fifo_trans_pa2va(struct kni_dev *kni,
 	struct rte_kni_fifo *src_pa, struct rte_kni_fifo *dst_va)
 {
 	uint32_t ret, i, num_dst, num_rx;
-	void *kva;
+	struct rte_kni_mbuf *kva, *_kva;
+	int nb_segs;
+	int kva_nb_segs;
+
 	do {
 		num_dst = kni_fifo_free_count(dst_va);
 		if (num_dst == 0)
@@ -188,6 +179,17 @@ kni_fifo_trans_pa2va(struct kni_dev *kni,
 		for (i = 0; i < num_rx; i++) {
 			kva = pa2kva(kni->pa[i]);
 			kni->va[i] = pa2va(kni->pa[i], kva);
+
+			kva_nb_segs = kva->nb_segs;
+			for (nb_segs = 0; nb_segs < kva_nb_segs; nb_segs++) {
+				if (!kva->next)
+					break;
+
+				_kva = kva;
+				kva = pa2kva(kva->next);
+				/* Convert physical address to virtual address */
+				_kva->next = pa2va(_kva->next, kva);
+			}
 		}
 
 		ret = kni_fifo_put(dst_va, kni->va, num_rx);
@@ -313,7 +315,7 @@ kni_net_rx_normal(struct kni_dev *kni)
 	uint32_t ret;
 	uint32_t len;
 	uint32_t i, num_rx, num_fq;
-	struct rte_kni_mbuf *kva;
+	struct rte_kni_mbuf *kva, *_kva;
 	void *data_kva;
 	struct sk_buff *skb;
 	struct net_device *dev = kni->net_dev;
@@ -363,8 +365,11 @@ kni_net_rx_normal(struct kni_dev *kni)
 				if (!kva->next)
 					break;
 
-				kva = pa2kva(va2pa(kva->next, kva));
+				_kva = kva;
+				kva = pa2kva(kva->next);
 				data_kva = kva2data_kva(kva);
+				/* Convert physical address to virtual address */
+				_kva->next = pa2va(_kva->next, kva);
 			}
 		}
 
@@ -481,7 +486,7 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 	uint32_t ret;
 	uint32_t len;
 	uint32_t i, num_rq, num_fq, num;
-	struct rte_kni_mbuf *kva;
+	struct rte_kni_mbuf *kva, *_kva;
 	void *data_kva;
 	struct sk_buff *skb;
 	struct net_device *dev = kni->net_dev;
@@ -545,8 +550,11 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 				if (!kva->next)
 					break;
 
-				kva = pa2kva(va2pa(kva->next, kva));
+				_kva = kva;
+				kva = pa2kva(kva->next);
 				data_kva = kva2data_kva(kva);
+				/* Convert physical address to virtual address */
+				_kva->next = pa2va(_kva->next, kva);
 			}
 		}
 
lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h +1 -1
index 5afa08713..688db9758 100644
@@ -86,7 +86,7 @@ struct rte_kni_mbuf {
 	/* fields on second cache line */
 	char pad3[8] __attribute__((__aligned__(RTE_CACHE_LINE_MIN_SIZE)));
 	void *pool;
-	void *next;
+	void *next;             /**< Physical address of next mbuf in kernel. */
 };
 
 /*
lib/librte_kni/rte_kni.c +14 -1
index 73aeccccf..74b1ff5b6 100644
@@ -353,6 +353,19 @@ va2pa(struct rte_mbuf *m)
 			 (unsigned long)m->buf_iova));
 }
 
+static void *
+va2pa_all(struct rte_mbuf *mbuf)
+{
+	void *phy_mbuf = va2pa(mbuf);
+	struct rte_mbuf *next = mbuf->next;
+	while (next) {
+		mbuf->next = va2pa(next);
+		mbuf = next;
+		next = mbuf->next;
+	}
+	return phy_mbuf;
+}
+
 static void
 obj_free(struct rte_mempool *mp __rte_unused, void *opaque, void *obj,
 		unsigned obj_idx __rte_unused)
@@ -550,7 +563,7 @@ rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
 	unsigned int i;
 
 	for (i = 0; i < num; i++)
-		phy_mbufs[i] = va2pa(mbufs[i]);
+		phy_mbufs[i] = va2pa_all(mbufs[i]);
 
 	ret = kni_fifo_put(kni->rx_q, phy_mbufs, num);
 
Signature · 1 line
2.19.1.windows.1
#4 lore raw
Quoted 10 lines · Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

Quoted 4 lines

---
v2: Add an explanation that causes this problem.

    Use m->next to store physical address.

Overall code looks good to me, thanks for the work, it looks pretty clean. I would like to do some test before giving an ack.

Meanwhile, can you please update kni documentation, to document for the mbufs sent to kernel has mbuf->next fields as physical address. It is OK to send as a new version of the patch with documentation.

Thanks,
ferruh

#5 lore raw
In reply to Yangchao Zhou
Quoted 14 lines · Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

---
v2: Add an explanation that causes this problem.

    Use m->next to store physical address.

<...>

Quoted 12 lines
kernel/linux/kni/kni_net.c +1 -1
@@ -481,7 +486,7 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 	uint32_t ret;
 	uint32_t len;
 	uint32_t i, num_rq, num_fq, num;
-	struct rte_kni_mbuf *kva;
+	struct rte_kni_mbuf *kva, *_kva;
 	void *data_kva;
 	struct sk_buff *skb;
 	struct net_device *dev = kni->net_dev;
@@ -545,8 +550,11 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 				if (!kva->next)
 					break;

kernel/linux/kni/kni_net.c kni_net_rx_lo_fifo_skb(struct kni_dev *kni) · line 560

Quoted 10 lines
kernel/linux/kni/kni_net.c +4 -1
 
-				kva = pa2kva(va2pa(kva->next, kva));
+				_kva = kva;
+				kva = pa2kva(kva->next);
 				data_kva = kva2data_kva(kva);
+				/* Convert physical address to virtual address */
+				_kva->next = pa2va(_kva->next, kva);
 			}
 		}
 

Also need to update 'kni_net_rx_lo_fifo()', at worst to update 'next' fields because it fills 'kni->free_q', without conversion userspace will crash.

<...>

lib/librte_kni/rte_kni.c rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num) · line 568

Quoted 8 lines
lib/librte_kni/rte_kni.c +1 -1
@@ -550,7 +563,7 @@ rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
 	unsigned int i;
 
 	for (i = 0; i < num; i++)
-		phy_mbufs[i] = va2pa(mbufs[i]);
+		phy_mbufs[i] = va2pa_all(mbufs[i]);
 
 	ret = kni_fifo_put(kni->rx_q, phy_mbufs, num);

There is a problem here.

When fifo 'kni->rx_q' is full, 'rte_kni_tx_burst' will send less mbuf than requested, than the application needs to handle the remaining packages, most probably will free them, but now some packages has physical address in their 'next' field, which will cause app to crash.

I don't know really how to solve this.
Perhaps getting free count from 'kni->rx_q' and only convert that much (va2pa_all) to physical address can work, but I can't estimate performance effect of it.

#6 lore raw

Hi Yigit,
I was facing the kernel crash issue, at skb_over_put(), using dpdk using 18.11 on 4.19.28 kernel. On checking I did find this patch and this patch is solving the issue also. But then I saw you comment in the patch and that’s looks scary to have the patch. Is there any improvements/fixes planned for this issue and in which version? is there any performance impact of the below patch ? As this issues is blocking our release any inputs to this asap will be really appreciated.

--
Regards,
Souvik

From: dev <[email protected]> On Behalf Of Ferruh Yigit
Sent: Friday, March 22, 2019 4:49 PM
To: Yangchao Zhou <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH v2] kni: fix possible kernel crash with va2pa

________________________________
NOTICE: This email was received from an EXTERNAL sender
________________________________

Quoted 14 lines · Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

---
v2: Add an explanation that causes this problem.
Use m->next to store physical address.

<...>

Quoted 12 lines
kernel/linux/kni/kni_net.c +1 -1
@@ -481,7 +486,7 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 uint32_t ret;
 uint32_t len;
 uint32_t i, num_rq, num_fq, num;
- struct rte_kni_mbuf *kva;
+ struct rte_kni_mbuf *kva, *_kva;
 void *data_kva;
 struct sk_buff *skb;
 struct net_device *dev = kni->net_dev;
@@ -545,8 +550,11 @@ kni_net_rx_lo_fifo_skb(struct kni_dev *kni)
 if (!kva->next)
 break;

kernel/linux/kni/kni_net.c kni_net_rx_lo_fifo_skb(struct kni_dev *kni) · line 560

Quoted 10 lines
kernel/linux/kni/kni_net.c +4 -1
 
- kva = pa2kva(va2pa(kva->next, kva));
+ _kva = kva;
+ kva = pa2kva(kva->next);
 data_kva = kva2data_kva(kva);
+ /* Convert physical address to virtual address */
+ _kva->next = pa2va(_kva->next, kva);
 }
 }
 

Also need to update 'kni_net_rx_lo_fifo()', at worst to update 'next' fields because it fills 'kni->free_q', without conversion userspace will crash.

<...>

lib/librte_kni/rte_kni.c rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num) · line 568

Quoted 8 lines
lib/librte_kni/rte_kni.c +1 -1
@@ -550,7 +563,7 @@ rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
 unsigned int i;
 
 for (i = 0; i < num; i++)
- phy_mbufs[i] = va2pa(mbufs[i]);
+ phy_mbufs[i] = va2pa_all(mbufs[i]);
 
 ret = kni_fifo_put(kni->rx_q, phy_mbufs, num);

There is a problem here.

When fifo 'kni->rx_q' is full, 'rte_kni_tx_burst' will send less mbuf than requested, than the application needs to handle the remaining packages, most probably will free them, but now some packages has physical address in their 'next' field, which will cause app to crash.

I don't know really how to solve this.
Perhaps getting free count from 'kni->rx_q' and only convert that much (va2pa_all) to physical address can work, but I can't estimate performance effect of it.

-----------------------------------------------------------------------------------------------------------------------
Notice: This e-mail together with any attachments may contain information of Ribbon Communications Inc. that is confidential and/or proprietary for the sole use of the intended recipient. Any review, disclosure, reliance or distribution by others or forwarding without express permission is strictly prohibited. If you are not the intended recipient, please notify the sender immediately and then delete all copies, including any attachments.
-----------------------------------------------------------------------------------------------------------------------

#7 lore raw
In reply to Yangchao Zhou
Quoted 11 lines · On Tue, 12 Mar 2019 17:22:32 +0800 Yangchao Zhou <[email protected]> wrote:

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

Could you add a Fixes tag?

#8 lore raw

PATCH v3 1/1 kni: fix possible kernel crash with va2pa

In reply to Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

Fixes 8451269e6d7b ("kni: remove continuous memory restriction")
---
 .../prog_guide/kernel_nic_interface.rst       |  6 ++-
 kernel/linux/kni/kni_net.c                    | 51 ++++++++++++-------
 .../linux/eal/include/rte_kni_common.h        |  2 +-
 lib/librte_kni/rte_kni.c                      | 16 +++++-
 lib/librte_kni/rte_kni_fifo.h                 | 12 +++++
 5 files changed, 65 insertions(+), 22 deletions(-)
doc/guides/prog_guide/kernel_nic_interface.rst +4 -2
index daf87f4a8..2fd3b7983 100644
@@ -268,12 +268,14 @@ Use Case: Ingress
 -----------------
 
 On the DPDK RX side, the mbuf is allocated by the PMD in the RX thread context.
-This thread will enqueue the mbuf in the rx_q FIFO.
+This thread will enqueue the mbuf in the rx_q FIFO, and the next pointers in mbuf-chain will convert to physical address.
 The KNI thread will poll all KNI active devices for the rx_q.
 If an mbuf is dequeued, it will be converted to a sk_buff and sent to the net stack via netif_rx().
-The dequeued mbuf must be freed, so the same pointer is sent back in the free_q FIFO.
+The dequeued mbuf must be freed, so the same pointer is sent back in the free_q FIFO,
+and next pointers must convert back to virtual address if exists before put in the free_q FIFO.
 
 The RX thread, in the same main loop, polls this FIFO and frees the mbuf after dequeuing it.
+The address conversion of the next pointer is to prevent the chained mbuf in different hugepage segments from causing kernel crash.
 
 Use Case: Egress
 ----------------
kernel/linux/kni/kni_net.c +33 -18

107 lines not on this page · raw message

lib/librte_eal/linux/eal/include/rte_kni_common.h +1 -1
index 91a1c1408..37d9ee8f0 100644
@@ -86,7 +86,7 @@ struct rte_kni_mbuf {
 	/* fields on second cache line */
 	char pad3[8] __attribute__((__aligned__(RTE_CACHE_LINE_MIN_SIZE)));
 	void *pool;
-	void *next;
+	void *next;             /**< Physical address of next mbuf in kernel. */
 };
 
 /*
lib/librte_kni/rte_kni.c +15 -1
index e29d0cc7d..6fc36f744 100644
@@ -344,6 +344,19 @@ va2pa(struct rte_mbuf *m)
 			 (unsigned long)m->buf_iova));
 }
 
+static void *
+va2pa_all(struct rte_mbuf *mbuf)
+{
+	void *phy_mbuf = va2pa(mbuf);
+	struct rte_mbuf *next = mbuf->next;
+	while (next) {
+		mbuf->next = va2pa(next);
+		mbuf = next;
+		next = mbuf->next;
+	}
+	return phy_mbuf;
+}
+
 static void
 obj_free(struct rte_mempool *mp __rte_unused, void *opaque, void *obj,
 		unsigned obj_idx __rte_unused)
@@ -536,12 +549,13 @@ rte_kni_handle_request(struct rte_kni *kni)
 unsigned
 rte_kni_tx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned num)
 {
+	num = RTE_MIN(kni_fifo_free_count(kni->rx_q), num);
 	void *phy_mbufs[num];
 	unsigned int ret;
 	unsigned int i;
 
 	for (i = 0; i < num; i++)
-		phy_mbufs[i] = va2pa(mbufs[i]);
+		phy_mbufs[i] = va2pa_all(mbufs[i]);
 
 	ret = kni_fifo_put(kni->rx_q, phy_mbufs, num);
 
lib/librte_kni/rte_kni_fifo.h +12 -0
index 287d7deb2..11d7fd6bd 100644
@@ -104,3 +104,15 @@ kni_fifo_count(struct rte_kni_fifo *fifo)
 	unsigned fifo_read = __KNI_LOAD_ACQUIRE(&fifo->read);
 	return (fifo->len + fifo_write - fifo_read) & (fifo->len - 1);
 }
+
+/**
+ * Get the num of available elements in the fifo
+ */
+static inline uint32_t
+kni_fifo_free_count(struct rte_kni_fifo *fifo)
+{
+	uint32_t fifo_write = __KNI_LOAD_ACQUIRE(&fifo->write);
+	uint32_t fifo_read = __KNI_LOAD_ACQUIRE(&fifo->read);
+	return (fifo_read - fifo_write - 1) & (fifo->len - 1);
+}
+
Signature · 1 line
2.17.1
#9 lore raw

I am battling a related problem as reported on
https://bugs.dpdk.org/show_bug.cgi?id=183 and this patch seems relevant, so I applied this patch on 196a46fab6eeb3ce2039e3bcaca80f8ba43ffc8d

However, this patch does not work for me:
with CONFIG_RTE_LIBRTE_MBUF_DEBUG enabled, kni_free_mbufs's invocation of rte_pktmbuf_free throws "bad mbuf pool" error.

While all mbufs and segments in kni->rx_q now have physical addresses, the mbufs and segments placed back to kni->free_q still have (mis-)calculated virtual address. The pa2va function is not working properly.

Consequently, userspace side is passing wrong pointer to rte_pktmbuf_free, so that application crashes with CONFIG_RTE_LIBRTE_MBUF_DEBUG enabled.

#10 lore raw
Quoted 8 lines · Junxiao Shi

I am battling a related problem as reported on
https://bugs.dpdk.org/show_bug.cgi?id=183 and this patch seems relevant, so I applied this patch on 196a46fab6eeb3ce2039e3bcaca80f8ba43ffc8d

However, this patch does not work for me:
with CONFIG_RTE_LIBRTE_MBUF_DEBUG enabled, kni_free_mbufs's invocation of rte_pktmbuf_free throws "bad mbuf pool" error.

Quoted 7 lines

While all mbufs and segments in kni->rx_q now have physical addresses, the mbufs and segments placed back to kni->free_q still have (mis-)calculated virtual address. The pa2va function is not working properly.

Consequently, userspace side is passing wrong pointer to rte_pktmbuf_free, so that application crashes with CONFIG_RTE_LIBRTE_MBUF_DEBUG enabled.

Hi Junxiao,

I don't see any issue in the code, and as far as I tested not seeing any issue.

Can you please provide more information how to reproduce the issue, is it only enabling "CONFIG_RTE_LIBRTE_MBUF_DEBUG" config?

Thanks,
ferruh

#11 lore raw

Hi Ferruh

I've uploaded my test case to Bug 183.
You can determine whether it is a bug that should be covered by this patch, or it should be handled separately.

Yours, Junxiao

Quoted 27 lines · Ferruh Yigit
Quoted 8 lines · Junxiao Shi

I am battling a related problem as reported on
https://bugs.dpdk.org/show_bug.cgi?id=183 and this patch seems relevant, so I applied this patch on 196a46fab6eeb3ce2039e3bcaca80f8ba43ffc8d

However, this patch does not work for me:
with CONFIG_RTE_LIBRTE_MBUF_DEBUG enabled, kni_free_mbufs's invocation of rte_pktmbuf_free throws "bad mbuf pool" error.

Quoted 7 lines

While all mbufs and segments in kni->rx_q now have physical addresses, the mbufs and segments placed back to kni->free_q still have (mis-)calculated virtual address. The pa2va function is not working properly.

Consequently, userspace side is passing wrong pointer to rte_pktmbuf_free, so that application crashes with CONFIG_RTE_LIBRTE_MBUF_DEBUG enabled.

Hi Junxiao,

I don't see any issue in the code, and as far as I tested not seeing any issue.

Can you please provide more information how to reproduce the issue, is it only enabling "CONFIG_RTE_LIBRTE_MBUF_DEBUG" config?

Thanks,
ferruh

#12 lore raw
Quoted 5 lines · yoursunny

Hi Ferruh

I've uploaded my test case to Bug 183.
You can determine whether it is a bug that should be covered by this patch, or it should be handled separately.

Thanks for clarification,
your use case is about indirect mbufs, I can see why it is not working from your explanation in the bug report.
I think previously when all mbufs were coming from same physically continuous memory, indirect mbufs should work. Right now I don't know how to support them.

This defect fixes an issue for multi segment packages, caused by wrong address calculation for next segments.

So I am for continuing with this patch, independent from indirect mbufs issue.

Do you have any other issue, except than the indirect mbuf issue, if you have other KNI use cases?

Quoted 31 lines

Yours, Junxiao

Quoted 27 lines · Ferruh Yigit
Quoted 8 lines · Junxiao Shi

I am battling a related problem as reported on
https://bugs.dpdk.org/show_bug.cgi?id=183 and this patch seems relevant, so I applied this patch on 196a46fab6eeb3ce2039e3bcaca80f8ba43ffc8d

However, this patch does not work for me:
with CONFIG_RTE_LIBRTE_MBUF_DEBUG enabled, kni_free_mbufs's invocation of rte_pktmbuf_free throws "bad mbuf pool" error.

Quoted 7 lines

While all mbufs and segments in kni->rx_q now have physical addresses, the mbufs and segments placed back to kni->free_q still have (mis-)calculated virtual address. The pa2va function is not working properly.

Consequently, userspace side is passing wrong pointer to rte_pktmbuf_free, so that application crashes with CONFIG_RTE_LIBRTE_MBUF_DEBUG enabled.

Hi Junxiao,

I don't see any issue in the code, and as far as I tested not seeing any issue.

Can you please provide more information how to reproduce the issue, is it only enabling "CONFIG_RTE_LIBRTE_MBUF_DEBUG" config?

Thanks,
ferruh

#13 lore raw

Hi Ferruh

I have no objection of merging the current commit.
I do have some ideas on how to support indirect mbufs. Please review at https://bugs.dpdk.org/show_bug.cgi?id=183#c4

Yours, Junxiao

Quoted 5 lines · Ferruh Yigit

your use case is about indirect mbufs, I can see why it is not working from your explanation in the bug report.
I think previously when all mbufs were coming from same physically continuous memory, indirect mbufs should work. Right now I don't know how to support them.

#14 lore raw
In reply to Yangchao Zhou
Quoted 10 lines · Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

Quoted 3 lines
Fixes 8451269e6d7b ("kni: remove continuous memory restriction")

Overall looks good to me, not from this patch but can you please check below comment too.
Also there is a comment from Junxiao, lets clear it before the ack.

<...>

Quoted 10 lines
kernel/linux/kni/kni_net.c +1 -1
@@ -396,7 +401,7 @@ kni_net_rx_lo_fifo(struct kni_dev *kni)
 	uint32_t ret;
 	uint32_t len;
 	uint32_t i, num, num_rq, num_tq, num_aq, num_fq;
-	struct rte_kni_mbuf *kva;
+	struct rte_kni_mbuf *kva, *next_kva;
 	void *data_kva;
 	struct rte_kni_mbuf *alloc_kva;
 	void *alloc_data_kva;

kernel/linux/kni/kni_net.c kni_net_rx_lo_fifo(struct kni_dev *kni) · line 452

Quoted 9 lines
kernel/linux/kni/kni_net.c +6 -0
@@ -439,6 +444,13 @@ kni_net_rx_lo_fifo(struct kni_dev *kni)
 			data_kva = kva2data_kva(kva);
 			kni->va[i] = pa2va(kni->pa[i], kva);
 
+			while (kva->next) {
+				next_kva = pa2kva(kva->next);
+				/* Convert physical address to virtual address */
+				kva->next = pa2va(kva->next, next_kva);
+				kva = next_kva;
+			}

Not done in this patch, but in 'kni_net_rx_lo_fifo()' the len calculated as 'len = kva->pkt_len;'

But while copying 'data' to 'alloc_data' the segmentation is not taken into account and 'len' is used:
memcpy(alloc_data_kva, data_kva, len);

This may lead overflow 'alloc_data_kva' for some 'pkt_len' values.

Acked-by
#15 lore raw
Quoted 11 lines · Ferruh Yigit
Quoted 10 lines · Yangchao Zhou

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

Quoted 8 lines
Quoted 3 lines
Fixes 8451269e6d7b ("kni: remove continuous memory restriction")

Overall looks good to me, not from this patch but can you please check below comment too.
Also there is a comment from Junxiao, lets clear it before the ack.

Quoted 25 lines

<...>

Quoted 10 lines
kernel/linux/kni/kni_net.c +1 -1
@@ -396,7 +401,7 @@ kni_net_rx_lo_fifo(struct kni_dev *kni)
 	uint32_t ret;
 	uint32_t len;
 	uint32_t i, num, num_rq, num_tq, num_aq, num_fq;
-	struct rte_kni_mbuf *kva;
+	struct rte_kni_mbuf *kva, *next_kva;
 	void *data_kva;
 	struct rte_kni_mbuf *alloc_kva;
 	void *alloc_data_kva;

kernel/linux/kni/kni_net.c kni_net_rx_lo_fifo(struct kni_dev *kni) · line 452

Quoted 9 lines
kernel/linux/kni/kni_net.c +6 -0
@@ -439,6 +444,13 @@ kni_net_rx_lo_fifo(struct kni_dev *kni)
 			data_kva = kva2data_kva(kva);
 			kni->va[i] = pa2va(kni->pa[i], kva);
 
+			while (kva->next) {
+				next_kva = pa2kva(kva->next);
+				/* Convert physical address to virtual address */
+				kva->next = pa2va(kva->next, next_kva);
+				kva = next_kva;
+			}

Not done in this patch, but in 'kni_net_rx_lo_fifo()' the len calculated as 'len = kva->pkt_len;'

Quoted 6 lines

But while copying 'data' to 'alloc_data' the segmentation is not taken into account and 'len' is used:
memcpy(alloc_data_kva, data_kva, len);

This may lead overflow 'alloc_data_kva' for some 'pkt_len' values.

I will send separate patch for this.

#16 lore raw

11/07/2019 09:46, Ferruh Yigit:

Quoted 14 lines · Ferruh Yigit
Quoted 13 lines · Yangchao Zhou
Quoted 12 lines

va2pa depends on the physical address and virtual address offset of current mbuf. It may get the wrong physical address of next mbuf which allocated in another hugepage segment.

In rte_mempool_populate_default(), trying to allocate whole block of contiguous memory could be failed. Then, it would reserve memory in several memzones that have different physical address and virtual address offsets. The rte_mempool_populate_default() is used by rte_pktmbuf_pool_create().

Fixes 8451269e6d7b ("kni: remove continuous memory restriction")
Quoted 8 lines
Quoted 6 lines
Quoted 1 line

Overall looks good to me, not from this patch but can you please check below comment too.
Also there is a comment from Junxiao, lets clear it before the ack.

The commit log does not really explained neither the use case nor the solution. But as you acked it...
Applied, thanks