Skip to contentSkip to conversation

VIA velocity skb leak.

DISCUSSION Dave Jones 4 participants 13 messages Mar 12, 2009 – Mar 19, 2009

Subsystems (2)
lore
Share & tools
lore

Dave Jones

author
#0 lore raw

I think I've found a bug in the via-velocity driver, but I'm not sure how to fix it.
We've been carrying some debug patches in the Fedora kernel that dwmw2 wrote, which do some sanity checks on dma allocations. Velocity triggered this trace..

[  324.777540] ------------[ cut here ]------------
[  324.777559] WARNING: at lib/dma-debug.c:470 check_unmap+0x196/0x3e4() (Not tainted)
[  324.777575] Hardware name:  
[  324.777591] via-velocity 0000:00:0e.0: DMA-API: device driver frees DMA memory with different size [device address=0x000000001a4c8ca2] [map size=60 bytes] [unmap size=54 bytes]
[  324.777614] Modules linked in: ipt_MASQUERADE iptable_nat nf_nat ipt_LOG xt_limit sunrpc nf_conntrack_netbios_ns ip6t_REJECT ip6table_filter ip6_tables ipv6 cpufreq_ondemand serio_raw usb_storage pcspkr i2c_viapro 3c59x mii i2c_core firewire_ohci via_velocity firewire_core crc_ccitt crc_itu_t ata_generic pata_acpi pata_via ext2 [last unloaded: scsi_wait_scan]
[  324.777792] Pid: 0, comm: swapper Not tainted 2.6.29-0.179.rc6.git5.fc11.i686.PAE #1
[  324.777807] Call Trace:
[  324.777839]  [<c0437a9f>] warn_slowpath+0x7c/0xa7
[  324.777870]  [<c0701800>] ? _spin_unlock_irq+0x1c/0x34
[  324.777899]  [<c042ae55>] ? account_group_exec_runtime+0x4d/0x54
[  324.777929]  [<c045783c>] ? mark_lock+0x1e/0x30b
[  324.777952]  [<c045783c>] ? mark_lock+0x1e/0x30b
[  324.777973]  [<c0554765>] ? get_hash_bucket+0x26/0x2f
[  324.777995]  [<c0554fc3>] check_unmap+0x196/0x3e4
[  324.778021]  [<c040de4d>] ? sched_clock+0x9/0xd
[  324.778042]  [<c05553bf>] debug_dma_unmap_page+0x68/0x70
[  324.778064]  [<c05553bf>] ? debug_dma_unmap_page+0x68/0x70
[  324.778111]  [<dcfc251c>] pci_unmap_single+0x65/0x70 [via_velocity]
[  324.778148]  [<dcfc25f5>] velocity_tx_srv+0xce/0x183 [via_velocity]
[  324.778188]  [<dcfc39a7>] velocity_intr+0x52f/0x594 [via_velocity]
[  324.778211]  [<c0457023>] ? trace_hardirqs_off+0xb/0xd
[  324.778232]  [<c067d15d>] ? netif_rx+0x9d/0x11a
[  324.778277]  [<dc861f8e>] ? boomerang_interrupt+0x324/0x331 [3c59x]
[  324.778300]  [<c045783c>] ? mark_lock+0x1e/0x30b
[  324.778330]  [<c047a6a4>] ? handle_fasteoi_irq+0x76/0xb7
[  324.778351]  [<c04566d8>] ? lock_release_holdtime+0x2b/0x123
[  324.778374]  [<c0479284>] handle_IRQ_event+0x22/0x58
[  324.778394]  [<c0479284>] ? handle_IRQ_event+0x22/0x58
[  324.778417]  [<c047a6ae>] handle_fasteoi_irq+0x80/0xb7
[  324.778437]  [<c047a62e>] ? handle_fasteoi_irq+0x0/0xb7
[  324.778450]  <IRQ>  [<c0409cac>] ? common_interrupt+0x2c/0x34
[  324.778490]  [<c0457d78>] ? trace_hardirqs_on+0xb/0xd
[  324.778501]  [<c059fcca>] ? acpi_idle_enter_simple+0x13a/0x179
[  324.778501]  [<c059f9b4>] ? acpi_idle_enter_bm+0xd1/0x2ad
[  324.778501]  [<c0660105>] ? menu_select+0x3d/0xa4
[  324.778501]  [<c065f670>] ? cpuidle_idle_call+0x65/0x9d
[  324.778501]  [<c065f670>] ? cpuidle_idle_call+0x65/0x9d
[  324.778501]  [<c040871e>] ? cpu_idle+0x72/0x90
[  324.778501]  [<c06efa03>] ? rest_init+0x53/0x55
[  324.778501] ---[ end trace 57ea7e11ea46a364 ]---

It looks like in velocity_xmit we do the skb_padto because the length is below the minimum frame size (56 bytes) and pci_map the resulting 60 byte skb. But later, when we free it, we forget that we did that size adjustment. Does skb_padto adjust the skb->len somewhere, or is this responsibility of the caller?

Dave

Signature · 1 line
http://www.codemonkey.org.uk
maintainer
#1 lore raw

velocity_xmit() needs to set 'pktlen = skb->len;' after, not before, the skb_padto() call.

maintainer
#2 lore raw

From: David Miller <[email protected]>
Date: Wed, 11 Mar 2009 21:17:06 -0700 (PDT)

Quoted 3 lines

velocity_xmit() needs to set 'pktlen = skb->len;' after, not before, the skb_padto() call.

Actually that won't work since, as you suggest, skb->len isn't updated by skb_padto().

So the transmit needs something like:

	pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);

velocity_free_tx_buf() needs to make the same calculation instead of just plain skb->len

This bug probably exists in every other driver using skb_padto() :-)

author
#3 lore raw
Quoted 10 lines · David Miller

From: David Miller <[email protected]>
Date: Wed, 11 Mar 2009 21:17:06 -0700 (PDT)

Quoted 3 lines

velocity_xmit() needs to set 'pktlen = skb->len;' after, not before, the skb_padto() call.

Actually that won't work since, as you suggest, skb->len isn't updated by skb_padto().

Quoted 6 lines

So the transmit needs something like:

	pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);

velocity_free_tx_buf() needs to make the same calculation instead of just plain skb->len

Something like this ?
(It looks like the ZERO_COPY_SUPPORT is never enabled anywhere,
so I didn't dig into how that works).

drivers/net/via-velocity.c +7 -8
index c5691fd..cd34dda 100644
@@ -1838,6 +1838,7 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_
 {
 	struct sk_buff *skb = tdinfo->skb;
 	int i;
+	int pktlen;
 
 	/*
 	 *	Don't unmap the pre-allocated tx_bufs
@@ -1845,10 +1846,11 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_
 	if (tdinfo->skb_dma) {
 
+		pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);
 		for (i = 0; i < tdinfo->nskb_dma; i++) {
 #ifdef VELOCITY_ZERO_COPY_SUPPORT
 			pci_unmap_single(vptr->pdev, tdinfo->skb_dma[i], le16_to_cpu(td->tdesc1.len), PCI_DMA_TODEVICE);
 #else
-			pci_unmap_single(vptr->pdev, tdinfo->skb_dma[i], skb->len, PCI_DMA_TODEVICE);
+			pci_unmap_single(vptr->pdev, tdinfo->skb_dma[i], pktlen, PCI_DMA_TODEVICE);
 #endif
 			tdinfo->skb_dma[i] = 0;
 		}
@@ -2080,17 +2083,14 @@ static int velocity_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct tx_desc *td_ptr;
 	struct velocity_td_info *tdinfo;
 	unsigned long flags;
-	int pktlen = skb->len;
+	int pktlen;
 	__le16 len;
 	int index;
 
 
-
-	if (skb->len < ETH_ZLEN) {
-		if (skb_padto(skb, ETH_ZLEN))
-			goto out;
-		pktlen = ETH_ZLEN;
-	}
+	if (skb_padto(skb, ETH_ZLEN))
+		goto out;
+	pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);
 
 	len = cpu_to_le16(pktlen);
 
Quoted 2 lines

This bug probably exists in every other driver using skb_padto() :-)

Once I've tested this (tomorrow), I'll do a sweep through some of the others. I expect the dma-debug stuff will pick them up eventually when that hits mainline, so it'd be good to fix up the lower hanging fruit.

The dma-debug patches are kinda neat, this is just one type of bug class it picks up. I forwarded another type to the e1000 list earlier today.

Dave

Signature · 1 line
http://www.codemonkey.org.uk
#4 lore raw
Quoted 35 lines · Dave Jones
Quoted 10 lines · David Miller

From: David Miller <[email protected]>
Date: Wed, 11 Mar 2009 21:17:06 -0700 (PDT)

Quoted 3 lines

velocity_xmit() needs to set 'pktlen = skb->len;' after, not before, the skb_padto() call.

Actually that won't work since, as you suggest, skb->len isn't updated by skb_padto().

Quoted 6 lines

So the transmit needs something like:

	pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);

velocity_free_tx_buf() needs to make the same calculation instead of just plain skb->len

Something like this ?
(It looks like the ZERO_COPY_SUPPORT is never enabled anywhere,
so I didn't dig into how that works).

drivers/net/via-velocity.c +1 -0
index c5691fd..cd34dda 100644
@@ -1838,6 +1838,7 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_
 {
 	struct sk_buff *skb = tdinfo->skb;
 	int i;
+	int pktlen;
 
 	/*
 	 *	Don't unmap the pre-allocated tx_bufs

drivers/net/via-velocity.c static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_ · line 1848

Quoted 3 lines
drivers/net/via-velocity.c +1 -0
@@ -1845,10 +1846,11 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_
 	if (tdinfo->skb_dma) {
 
+		pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);

I personally find better to use max(skb->len, ETH_ZLEN) macro, but YMMV ;)

It actually can avoid you a bug ;)

author
#5 lore raw
Quoted 8 lines · Eric Dumazet

Review static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_ · line 1848

Quoted 4 lines
(quoted diff) +1 -0
@@ -1845,10 +1846,11 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_
 	if (tdinfo->skb_dma) {
 
+		pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);

I personally find better to use max(skb->len, ETH_ZLEN) macro, but YMMV ;)

It actually can avoid you a bug ;)

I prefer that too, but it makes a warning.

drivers/net/via-velocity.c:2093: warning: comparison of distinct pointer types lacks a cast

We can fix this by either casting ETH_ZLEN to an unsigned int, or we could just do the diff below..

Or did I overlook something?

(if this looks ok, perhaps the other defines could use the same treatment?)

Dave

The minimum frame length is never signed, define it as such so we don't need excessive casts in comparisons.

include/linux/if_ether.h +1 -1
index 7f3c735..c41183e 100644
@@ -30,7 +30,7 @@
 
 #define ETH_ALEN	6		/* Octets in one ethernet addr	 */
 #define ETH_HLEN	14		/* Total octets in header.	 */
-#define ETH_ZLEN	60		/* Min. octets in frame sans FCS */
+#define ETH_ZLEN	60U		/* Min. octets in frame sans FCS */
 #define ETH_DATA_LEN	1500		/* Max. octets in payload	 */
 #define ETH_FRAME_LEN	1514		/* Max. octets in frame sans FCS */
 #define ETH_FCS_LEN	4		/* Octets in the FCS		 */
Signature · 1 line
http://www.codemonkey.org.uk
#6 lore raw
Quoted 35 lines · Dave Jones
Quoted 8 lines · Eric Dumazet

Review static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_ · line 1848

Quoted 4 lines
(quoted diff) +1 -0
@@ -1845,10 +1846,11 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_
 	if (tdinfo->skb_dma) {
 
+		pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);

I personally find better to use max(skb->len, ETH_ZLEN) macro, but YMMV ;)

It actually can avoid you a bug ;)

I prefer that too, but it makes a warning.

drivers/net/via-velocity.c:2093: warning: comparison of distinct pointer types lacks a cast

We can fix this by either casting ETH_ZLEN to an unsigned int, or we could just do the diff below..

Or did I overlook something?

(if this looks ok, perhaps the other defines could use the same treatment?)

Dave

The minimum frame length is never signed, define it as such so we don't need excessive casts in comparisons.

include/linux/if_ether.h +0 -0
index 7f3c735..c41183e 100644

include/linux/if_ether.h line 36

Quoted 8 lines
include/linux/if_ether.h +1 -1
@@ -30,7 +30,7 @@
 
 #define ETH_ALEN	6		/* Octets in one ethernet addr	 */
 #define ETH_HLEN	14		/* Total octets in header.	 */
-#define ETH_ZLEN	60		/* Min. octets in frame sans FCS */
+#define ETH_ZLEN	60U		/* Min. octets in frame sans FCS */
 #define ETH_DATA_LEN	1500		/* Max. octets in payload	 */
 #define ETH_FRAME_LEN	1514		/* Max. octets in frame sans FCS */
 #define ETH_FCS_LEN	4		/* Octets in the FCS		 */

or use max_t(unsigned, skb->len, ETH_ZLEN)

maintainer
#7 lore raw

From: Eric Dumazet <[email protected]>
Date: Thu, 12 Mar 2009 06:14:27 +0100

Quoted 1 line

or use max_t(unsigned, skb->len, ETH_ZLEN)

Right, I'll add Dave's patch, but using max_t() here.

Thanks everyone.

author
#8 lore raw

a2065 skb_padto cleanups

Whilst looking for other instances of the bug that was in velocity, I noticed a few possible cleanups in a2065. I don't have this hardware, so extra review appreciated.

Dave
---

Remove unnecessary check (skb_padto does the same check) Remove unnecessary duplicate variable
Remove unnecessary clearing of padded part of skb.

drivers/net/a2065.c +5 -14
index 7a60bdd..d0d0c2f 100644
@@ -552,18 +552,13 @@ static int lance_start_xmit (struct sk_buff *skb, struct net_device *dev)
 	struct lance_private *lp = netdev_priv(dev);
 	volatile struct lance_regs *ll = lp->ll;
 	volatile struct lance_init_block *ib = lp->init_block;
-	int entry, skblen, len;
+	int entry, skblen;
 	int status = 0;
 	unsigned long flags;
 
-	skblen = skb->len;
-	len = skblen;
-
-	if (len < ETH_ZLEN) {
-		len = ETH_ZLEN;
-		if (skb_padto(skb, ETH_ZLEN))
-			return 0;
-	}
+	if (skb_padto(skb, ETH_ZLEN))
+		return 0;
+	skblen = max_t(unsigned, skb->len, ETH_ZLEN);
 
 	local_irq_save(flags);
 
@@ -586,15 +581,11 @@ static int lance_start_xmit (struct sk_buff *skb, struct net_device *dev)
 	}
 #endif
 	entry = lp->tx_new & lp->tx_ring_mod_mask;
-	ib->btx_ring [entry].length = (-len) | 0xf000;
+	ib->btx_ring [entry].length = (-skblen) | 0xf000;
 	ib->btx_ring [entry].misc = 0;
 
 	skb_copy_from_linear_data(skb, (void *)&ib->tx_buf [entry][0], skblen);
 
-	/* Clear the slack of the packet, do I need this? */
-	if (len != skblen)
-		memset ((void *) &ib->tx_buf [entry][skblen], 0, len - skblen);
-
 	/* Now, give the packet to the lance */
 	ib->btx_ring [entry].tmd1_bits = (LE_T1_POK|LE_T1_OWN);
 	lp->tx_new = (lp->tx_new+1) & lp->tx_ring_mod_mask;
Signature · 1 line
http://www.codemonkey.org.uk
author
#10 lore raw

r8169 skb leak

In reply to David Miller

r8169 seems to have the same problem that the via-velocity did with skb padding. Found by code-inspection only, no hardware to test.

drivers/net/r8169.c +6 -6
index b347340..07b1de1 100644
@@ -3145,7 +3145,9 @@ err_out:
 static void rtl8169_unmap_tx_skb(struct pci_dev *pdev, struct ring_info *tx_skb,
 				 struct TxDesc *desc)
 {
-	unsigned int len = tx_skb->len;
+	unsigned int len;
+
+	len = max_t(unsigned, tx_skb->len, ETH_ZLEN);
 
 	pci_unmap_single(pdev, le64_to_cpu(desc->addr), len, PCI_DMA_TODEVICE);
 	desc->opts1 = 0x00;
@@ -3364,11 +3366,9 @@ static int rtl8169_start_xmit(struct sk_buff *skb, struct net_device *dev)
 	} else {
 		len = skb->len;
 
-		if (unlikely(len < ETH_ZLEN)) {
-			if (skb_padto(skb, ETH_ZLEN))
-				goto err_update_stats;
-			len = ETH_ZLEN;
-		}
+		if (skb_padto(skb, ETH_ZLEN))
+			goto err_update_stats;
+		len = max_t(unsigned, skb->len, ETH_ZLEN);
 
 		opts1 |= FirstFrag | LastFrag;
 		tp->tx_skb[entry].skb = skb;
Signature · 1 line
http://www.codemonkey.org.uk
#11 lore raw

Re: r8169 skb leak

Dave Jones <[email protected]> :

Quoted 2 lines

r8169 seems to have the same problem that the via-velocity did with skb padding. Found by code-inspection only, no hardware to test.

Davem, can you defer this one until now + 24h ?

The hardware is supposed to auto-pad: I'd rather check it does with a few adapters and remove this code altogether.

Signature · 1 line
Ueimor
maintainer
#12 lore raw

Re: r8169 skb leak

From: Francois Romieu <[email protected]>
Date: Fri, 13 Mar 2009 23:26:18 +0100

Quoted 8 lines

Dave Jones <[email protected]> :

Quoted 2 lines

r8169 seems to have the same problem that the via-velocity did with skb padding. Found by code-inspection only, no hardware to test.

Davem, can you defer this one until now + 24h ?

The hardware is supposed to auto-pad: I'd rather check it does with a few adapters and remove this code altogether.

Sure.