Skip to contentSkip to conversation

8139cp dma-debug warning.

DISCUSSION Dave Jones 3 participants 6 messages Aug 6, 2009 – Aug 13, 2009

Subsystems (2)
lore
Share & tools
lore

Dave Jones

author
#0 lore raw

I'm chasing yet another dma-debug warning where we're unmapping a different size to what we mapped.

Quoted 30 lines
WARNING: at lib/dma-debug.c:803 check_unmap+0x1f5/0x509() (Not tainted)
Hardware name: 
8139cp 0000:00:03.0: DMA-API: device driver frees DMA memory with different
size [device address=0x000000001e9f8852] [map size=1536 bytes] [unmap size=1538
bytes]
Modules linked in: ipv6 dm_multipath uinput joydev 8139too virtio_balloon
8139cp mii i2c_piix4 virtio_pci i2c_core floppy squashfs pata_acpi ata_generic
[last unloaded: scsi_wait_scan]
Pid: 0, comm: swapper Not tainted 2.6.31-0.125.rc5.git2.fc12.i686 #1
Call Trace:
 [<c0444408>] warn_slowpath_common+0x7b/0xa3
 [<c0606746>] ? check_unmap+0x1f5/0x509
 [<c0444499>] warn_slowpath_fmt+0x34/0x48
 [<c0606746>] check_unmap+0x1f5/0x509
 [<c04ecb43>] ? check_valid_pointer+0x2c/0x6c
 [<c0606c00>] debug_dma_unmap_page+0x62/0x7b
 [<e0dc80be>] dma_unmap_single_attrs.clone.2+0x5a/0x75 [8139cp]
 [<e0dc8220>] cp_rx_poll+0x147/0x301 [8139cp]
 [<c077e5d4>] net_rx_action+0xa7/0x1d3
 [<c044accf>] __do_softirq+0xc8/0x192
 [<c044ade2>] do_softirq+0x49/0x7f
 [<c044af36>] irq_exit+0x48/0x8c
 [<c041c302>] smp_apic_timer_interrupt+0x7a/0x99
 [<c0404416>] apic_timer_interrupt+0x36/0x3c
 [<c0425e98>] ? native_safe_halt+0xa/0xc
 [<c040ad6f>] default_idle+0x55/0x98
 [<c046e44d>] ? trace_hardirqs_off+0x19/0x2c
 [<c04029e3>] cpu_idle+0xac/0xcd
 [<c08080ee>] rest_init+0x66/0x79
 [<c0a72ae0>] start_kernel+0x36f/0x385
Quoted 8 lines
 [<c0a7207e>] __init_begin+0x7e/0x96
---[ end trace f3c3298e5df24f15 ]---
Mapped at:
 [<c0606f6b>] debug_dma_map_page+0x6b/0x13b
 [<e0dc7c22>] dma_map_single_attrs.clone.1+0x78/0x93 [8139cp]
 [<e0dc86d0>] cp_init_rings+0xaa/0x12c [8139cp]
 [<e0dc87d6>] cp_open+0x84/0x154 [8139cp]
 [<c077ff2b>] dev_open+0x99/0xe4

Looking at 8139cp.c I see this code in cp_rx_poll ..

 552                 buflen = cp->rx_buf_sz + NET_IP_ALIGN;
 553                 new_skb = netdev_alloc_skb(dev, buflen);
 554                 if (!new_skb) {

555 dev->stats.rx_dropped++;
556 goto rx_next;
557 }
558
559 skb_reserve(new_skb, NET_IP_ALIGN);

Aren't we padding the alignment twice here? or am I missing something? The warning was a difference of two bytes, which is NET_IP_ALIGN, so I'm wondering if that skb_reserve line needs to be nuked ?

Dave

author
#1 lore raw
Quoted 54 lines · Dave Jones

I'm chasing yet another dma-debug warning where we're unmapping a different size to what we mapped.

Quoted 30 lines
WARNING: at lib/dma-debug.c:803 check_unmap+0x1f5/0x509() (Not tainted)
Hardware name: 
8139cp 0000:00:03.0: DMA-API: device driver frees DMA memory with different
size [device address=0x000000001e9f8852] [map size=1536 bytes] [unmap size=1538
bytes]
Modules linked in: ipv6 dm_multipath uinput joydev 8139too virtio_balloon
8139cp mii i2c_piix4 virtio_pci i2c_core floppy squashfs pata_acpi ata_generic
[last unloaded: scsi_wait_scan]
Pid: 0, comm: swapper Not tainted 2.6.31-0.125.rc5.git2.fc12.i686 #1
Call Trace:
 [<c0444408>] warn_slowpath_common+0x7b/0xa3
 [<c0606746>] ? check_unmap+0x1f5/0x509
 [<c0444499>] warn_slowpath_fmt+0x34/0x48
 [<c0606746>] check_unmap+0x1f5/0x509
 [<c04ecb43>] ? check_valid_pointer+0x2c/0x6c
 [<c0606c00>] debug_dma_unmap_page+0x62/0x7b
 [<e0dc80be>] dma_unmap_single_attrs.clone.2+0x5a/0x75 [8139cp]
 [<e0dc8220>] cp_rx_poll+0x147/0x301 [8139cp]
 [<c077e5d4>] net_rx_action+0xa7/0x1d3
 [<c044accf>] __do_softirq+0xc8/0x192
 [<c044ade2>] do_softirq+0x49/0x7f
 [<c044af36>] irq_exit+0x48/0x8c
 [<c041c302>] smp_apic_timer_interrupt+0x7a/0x99
 [<c0404416>] apic_timer_interrupt+0x36/0x3c
 [<c0425e98>] ? native_safe_halt+0xa/0xc
 [<c040ad6f>] default_idle+0x55/0x98
 [<c046e44d>] ? trace_hardirqs_off+0x19/0x2c
 [<c04029e3>] cpu_idle+0xac/0xcd
 [<c08080ee>] rest_init+0x66/0x79
 [<c0a72ae0>] start_kernel+0x36f/0x385
Quoted 8 lines
 [<c0a7207e>] __init_begin+0x7e/0x96
---[ end trace f3c3298e5df24f15 ]---
Mapped at:
 [<c0606f6b>] debug_dma_map_page+0x6b/0x13b
 [<e0dc7c22>] dma_map_single_attrs.clone.1+0x78/0x93 [8139cp]
 [<e0dc86d0>] cp_init_rings+0xaa/0x12c [8139cp]
 [<e0dc87d6>] cp_open+0x84/0x154 [8139cp]
 [<c077ff2b>] dev_open+0x99/0xe4

Looking at 8139cp.c I see this code in cp_rx_poll ..

 552                 buflen = cp->rx_buf_sz + NET_IP_ALIGN;
 553                 new_skb = netdev_alloc_skb(dev, buflen);
 554                 if (!new_skb) {

555 dev->stats.rx_dropped++;
556 goto rx_next;
557 }
558
559 skb_reserve(new_skb, NET_IP_ALIGN);

Quoted 3 lines

Aren't we padding the alignment twice here? or am I missing something? The warning was a difference of two bytes, which is NET_IP_ALIGN, so I'm wondering if that skb_reserve line needs to be nuked ?

There's another instance of the same further in the file.

Does this look right?

Dave

drivers/net/8139cp.c +0 -4
index 07919d0..065c9c3 100644
@@ -556,8 +556,6 @@ rx_status_loop:
 			goto rx_next;
 		}
 
-		skb_reserve(new_skb, NET_IP_ALIGN);
-
 		dma_unmap_single(&cp->pdev->dev, mapping,
 				 buflen, PCI_DMA_FROMDEVICE);
 
@@ -1061,8 +1059,6 @@ static int cp_refill_rx(struct cp_private *cp)
 		if (!skb)
 			goto err_out;
 
-		skb_reserve(skb, NET_IP_ALIGN);
-
 		mapping = dma_map_single(&cp->pdev->dev, skb->data,
 					 cp->rx_buf_sz, PCI_DMA_FROMDEVICE);
 		cp->rx_skb[i] = skb;
maintainer
#2 lore raw

From: Dave Jones <[email protected]>
Date: Wed, 12 Aug 2009 13:13:33 -0400

Quoted 3 lines

There's another instance of the same further in the file.

Does this look right?

Dave, Francois posted the following fix to lkml (he forgot to CC: netdev and the people reporting this bug, oh well)

Did you see it? It should fix the bug.

8139cp: balance dma_map_single vs dma_unmap_single pair

The driver always:
1. allocate cp->rx_buf_sz + NET_IP_ALIGN
2. map cp->rx_buf_sz

---
 drivers/net/8139cp.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)
drivers/net/8139cp.c +2 -3
index 50efde1..d0dbbf3 100644
@@ -515,7 +515,7 @@ rx_status_loop:
 		dma_addr_t mapping;
 		struct sk_buff *skb, *new_skb;
 		struct cp_desc *desc;
-		unsigned buflen;
+		const unsigned buflen = cp->rx_buf_sz;
 
 		skb = cp->rx_skb[rx_tail];
 		BUG_ON(!skb);
@@ -549,8 +549,7 @@ rx_status_loop:
 			pr_debug("%s: rx slot %d status 0x%x len %d\n",
 			       dev->name, rx_tail, status, len);
 
-		buflen = cp->rx_buf_sz + NET_IP_ALIGN;
-		new_skb = netdev_alloc_skb(dev, buflen);
+		new_skb = netdev_alloc_skb(dev, buflen + NET_IP_ALIGN);
 		if (!new_skb) {
 			dev->stats.rx_dropped++;
 			goto rx_next;
Signature · 1 line
1.6.3.4
author
#3 lore raw
Quoted 9 lines · David Miller

From: Dave Jones <[email protected]>
Date: Wed, 12 Aug 2009 13:13:33 -0400

Quoted 3 lines

There's another instance of the same further in the file.

Does this look right?

Dave, Francois posted the following fix to lkml (he forgot to CC: netdev and the people reporting this bug, oh well)

Ah, I missed it (disk crashed on friday, so I've been backlogged since then). Thanks for the forward.

Quoted 28 lines

Did you see it? It should fix the bug.

8139cp: balance dma_map_single vs dma_unmap_single pair

The driver always:
1. allocate cp->rx_buf_sz + NET_IP_ALIGN
2. map cp->rx_buf_sz

---
 drivers/net/8139cp.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)
drivers/net/8139cp.c +1 -1
index 50efde1..d0dbbf3 100644
@@ -515,7 +515,7 @@ rx_status_loop:
 		dma_addr_t mapping;
 		struct sk_buff *skb, *new_skb;
 		struct cp_desc *desc;
-		unsigned buflen;
+		const unsigned buflen = cp->rx_buf_sz;
 
 		skb = cp->rx_skb[rx_tail];
 		BUG_ON(!skb);

drivers/net/8139cp.c rx_status_loop: · line 555

Quoted 9 lines
drivers/net/8139cp.c +1 -2
@@ -549,8 +549,7 @@ rx_status_loop:
 			pr_debug("%s: rx slot %d status 0x%x len %d\n",
 			       dev->name, rx_tail, status, len);
 
-		buflen = cp->rx_buf_sz + NET_IP_ALIGN;
-		new_skb = netdev_alloc_skb(dev, buflen);
+		new_skb = netdev_alloc_skb(dev, buflen + NET_IP_ALIGN);
 		if (!new_skb) {
 			dev->stats.rx_dropped++;
 			goto rx_next;

Below this, we're still doing an skb_reserve(NET_IP_ALIGN) on new_skb. Although the mapping is now constantly sized, aren't we still wastefully bumping the data/tail of the skb twice ?

Dave

#4 lore raw

Dave Jones <[email protected]> :
[...]

Quoted 4 lines
Quoted 4 lines
drivers/net/8139cp.c +0 -0
@@ -549,8 +549,7 @@ rx_status_loop:
 			pr_debug("%s: rx slot %d status 0x%x len %d\n",
 			       dev->name, rx_tail, status, len);
 
Quoted 10 lines

drivers/net/8139cp.c rx_status_loop: · line 555

Quoted 6 lines
drivers/net/8139cp.c +1 -2
-		buflen = cp->rx_buf_sz + NET_IP_ALIGN;
-		new_skb = netdev_alloc_skb(dev, buflen);
+		new_skb = netdev_alloc_skb(dev, buflen + NET_IP_ALIGN);
 		if (!new_skb) {
 			dev->stats.rx_dropped++;
 			goto rx_next;

Below this, we're still doing an skb_reserve(NET_IP_ALIGN) on new_skb. Although the mapping is now constantly sized, aren't we still wastefully bumping the data/tail of the skb twice ?

$ grep -n NET_IP_ALIGN drivers/net/8139cp.c
552:		new_skb = netdev_alloc_skb(dev, buflen + NET_IP_ALIGN);
558:		skb_reserve(new_skb, NET_IP_ALIGN);
1059:		skb = netdev_alloc_skb(dev, cp->rx_buf_sz + NET_IP_ALIGN);
1063:		skb_reserve(skb, NET_IP_ALIGN);

I do not get it : netdev_alloc_skb allocates but it does not "bump" as skb_reserve does (and skb_reserve does not allocate). Where would the double bump come from ?

The mapping is constantly sized but we want the ethernet header following IP data to be evenly aligned and thus the mapping to be oddly aligned.

Signature · 1 line
Ueimor
author
#5 lore raw
Quoted 10 lines · Francois Romieu
Quoted 3 lines

Below this, we're still doing an skb_reserve(NET_IP_ALIGN) on new_skb. Although the mapping is now constantly sized, aren't we still wastefully bumping the data/tail of the skb twice ?

$ grep -n NET_IP_ALIGN drivers/net/8139cp.c
552:		new_skb = netdev_alloc_skb(dev, buflen + NET_IP_ALIGN);
558:		skb_reserve(new_skb, NET_IP_ALIGN);
1059:		skb = netdev_alloc_skb(dev, cp->rx_buf_sz + NET_IP_ALIGN);
1063:		skb_reserve(skb, NET_IP_ALIGN);
Quoted 6 lines

I do not get it : netdev_alloc_skb allocates but it does not "bump" as skb_reserve does (and skb_reserve does not allocate). Where would the double bump come from ?

The mapping is constantly sized but we want the ethernet header following IP data to be evenly aligned and thus the mapping to be oddly aligned.

Ah, now I get it. Thanks for the explanation :)

Dave