velocity driver unmaps incorrect size.
Subsystems (2)
Seen during boot when the interface comes up..
Dave
------------[ cut here ]------------ WARNING: at lib/dma-debug.c:505 check_unmap+0x1f8/0x4d4() Hardware name: via-velocity 0000:00:0e.0: DMA-API: device driver frees DMA memory with different size [device address=0x000000001a456242] [map size=90 bytes] [unmap size=1 bytes] Modules linked in: nf_conntrack_netbios_ns ip6t_REJECT ip6table_filter ip6_tables ipv6 cpufreq_ondemand acpi_cpufreq serio_raw pcspkr i2c_viapro 3c59x mii i2c_core usb_storage firewire_ohci via_velocity firewire_core crc_ccitt crc_itu_t ata_generic pata_acpi pata_via ext2 [last unloaded: scsi_wait_scan] Pid: 0, comm: swapper Tainted: G W 2.6.30-6.fc12.i686.PAE #1 Call Trace: [<c04434a2>] warn_slowpath_common+0x75/0x9d [<c05e1029>] ? check_unmap+0x1f8/0x4d4 [<c0443533>] warn_slowpath_fmt+0x34/0x48 [<c05e1029>] check_unmap+0x1f8/0x4d4 [<c05e15a8>] debug_dma_unmap_page+0x71/0x8a [<dcf88829>] pci_unmap_single+0x74/0x90 [via_velocity] [<dcf88922>] velocity_tx_srv+0xdd/0x1a0 [via_velocity] [<dcf89e76>] velocity_intr+0x52f/0x5a1 [via_velocity] [<dd0c60d6>] ? ip6_output_finish+0x0/0x104 [ipv6] [<dd0c6172>] ? ip6_output_finish+0x9c/0x104 [ipv6] [<c0494938>] ? handle_fasteoi_irq+0x27/0xcf [<c046c8fb>] ? print_lock_contention_bug+0x1f/0xd1 [<c0494938>] ? handle_fasteoi_irq+0x27/0xcf [<c0494938>] ? handle_fasteoi_irq+0x27/0xcf [<c0492c06>] handle_IRQ_event+0x56/0x112 [<c0492c06>] ? handle_IRQ_event+0x56/0x112 [<c0494997>] handle_fasteoi_irq+0x86/0xcf [<c040c350>] handle_irq+0x50/0x6c [<c040bc4f>] do_IRQ+0x52/0xa4 [<c040a715>] common_interrupt+0x35/0x3c [<c0757f3f>] ? netif_receive_skb+0x1b9/0x3ca [<c07581de>] process_backlog+0x8e/0xcf [<c0756262>] net_rx_action+0xa7/0x1b9 [<c0449064>] __do_softirq+0xaa/0x173 [<c0449176>] do_softirq+0x49/0x7f [<c04492ca>] irq_exit+0x48/0x87 [<c040bc7c>] do_IRQ+0x7f/0xa4 [<c040a715>] common_interrupt+0x35/0x3c [<c046b2ff>] ? trace_hardirqs_on+0x19/0x2c [<c06437a7>] ? acpi_idle_enter_simple+0x11d/0x161 [<c07316ff>] cpuidle_idle_call+0x73/0xbc [<c0409091>] cpu_idle+0xac/0xcb [<c07e024a>] rest_init+0x66/0x79 [<c0a38acd>] start_kernel+0x36b/0x383 [<c0a38090>] i386_start_kernel+0x7f/0x98 ---[ end trace 6db0eec1553c1cdc ]--- Mapped at: [<c05e196f>] debug_dma_map_page+0x85/0x168 [<dcf891fb>] pci_map_single+0x8e/0xab [via_velocity] [<dcf8979a>] velocity_xmit+0xae/0x1c2 [via_velocity] [<c0758d2d>] dev_hard_start_xmit+0x233/0x2a5 [<c076a13e>] __qdisc_run+0xd0/0x1c5
David Miller
maintainerFrom: Dave Jones <[email protected]>
Date: Sun, 21 Jun 2009 13:37:45 -0400
Quoted 4 lines
------------[ cut here ]------------ WARNING: at lib/dma-debug.c:505 check_unmap+0x1f8/0x4d4() Hardware name:via-velocity 0000:00:0e.0: DMA-API: device driver frees DMA memory with different size [device address=0x000000001a456242] [map size=90 bytes] [unmap size=1 bytes]
Ok, bad unmap size is "1".
Quoted 9 lines
Call Trace: [<c04434a2>] warn_slowpath_common+0x75/0x9d [<c05e1029>] ? check_unmap+0x1f8/0x4d4 [<c0443533>] warn_slowpath_fmt+0x34/0x48 [<c05e1029>] check_unmap+0x1f8/0x4d4 [<c05e15a8>] debug_dma_unmap_page+0x71/0x8a [<dcf88829>] pci_unmap_single+0x74/0x90 [via_velocity] [<dcf88922>] velocity_tx_srv+0xdd/0x1a0 [via_velocity] [<dcf89e76>] velocity_intr+0x52f/0x5a1 [via_velocity]
So since this is happening in velocity_tx_srv() it has to be velocity_free_tx_buf(). It has two cases, one for when VELOCITY_ZERO_COPY_SUPPORT is defined and one for when that is not defined.
There is no way to set that define that I can see in the tree, so we only need to consider the case where this macro is not defined:
static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_info *tdinfo) {
...
if (tdinfo->skb_dma) {
pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN);
for (i = 0; i < tdinfo->nskb_dma; i++) {
pci_unmap_single(vptr->pdev, tdinfo->skb_dma[i], pktlen, PCI_DMA_TODEVICE);
tdinfo->skb_dma[i] = 0;
}
} ...
}
It seems to me that it's impossible for 'pktlen' to every be '1' here as the DMA debug code is claiming. It must always be at least ETH_ZLEN. And that is what is passed in for the unmap length.
David is there something wonky in your build or do you have any local patches applied? Is it possible that for some reason your build is forcing VELOCITY_ZERO_COPY_SUPPORT to be defined for some reason?
Thanks.
Quoted 37 lines · David Miller
Quoted 4 lines
------------[ cut here ]------------ WARNING: at lib/dma-debug.c:505 check_unmap+0x1f8/0x4d4() Hardware name:via-velocity 0000:00:0e.0: DMA-API: device driver frees DMA memory with different size [device address=0x000000001a456242] [map size=90 bytes] [unmap size=1 bytes]
Ok, bad unmap size is "1".
Quoted 7 lines
[<
c05e1029>] ?check_unmap+0x1f8/0x4d4[<c0443533>] warn_slowpath_fmt+0x34/0x48 [<c05e1029>] check_unmap+0x1f8/0x4d4 [<c05e15a8>] debug_dma_unmap_page+0x71/0x8a [<dcf88829>] pci_unmap_single+0x74/0x90 [via_velocity] [<dcf88922>] velocity_tx_srv+0xdd/0x1a0 [via_velocity] [<dcf89e76>] velocity_intr+0x52f/0x5a1 [via_velocity]So since this is happening in
velocity_tx_srv()it has to bevelocity_free_tx_buf(). It has two cases, one for whenVELOCITY_ZERO_COPY_SUPPORTis defined and one for when that is not defined.There is no way to set that define that I can see in the tree, so we only need to consider the case where this macro is not defined:
static void
velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_info *tdinfo){
...if (tdinfo->skb_dma) { pktlen = (skb->len > ETH_ZLEN ? : ETH_ZLEN); for (i = 0; i < tdinfo->nskb_dma; i++) { pci_unmap_single(vptr->pdev, tdinfo->skb_dma[i], pktlen, PCI_DMA_TODEVICE); tdinfo->skb_dma[i] = 0; } }...
}
Quoted 7 lines
It seems to me that it's impossible for 'pktlen' to every be '1' here as the DMA debug code is claiming. It must always be at least
ETH_ZLEN. And that is what is passed in for the unmap length.David is there something wonky in your build or do you have any local patches applied?
Nothing obvious that could explain it.
(It's the Fedora 11 kernel, which has no patches to this driver,
nor to net/)
Quoted 3 lines
Is it possible that for some reason
your build is forcingVELOCITY_ZERO_COPY_SUPPORTto be defined for some reason?
Memory corruption maybe?
It's especially odd in that it only happens once during boot, and never happens again. This is my firewall/router, so there's more packets going through that box than any other I have.
btw, given the zerocopy stuff has been disabled for so long, is it worth keeping it ?
Dave
David Miller
maintainerFrom: Dave Jones <[email protected]>
Date: Sun, 21 Jun 2009 22:40:44 -0400
Quoted 10 lines · David Miller
Quoted 4 lines
Is it possible that for some reason
your build is forcingVELOCITY_ZERO_COPY_SUPPORTto be defined for some reason?Memory corruption maybe?
It's especially odd in that it only happens once during boot, and never happens again. This is my firewall/router, so there's more packets going through that box than any other I have.
I would not be inclined to this theory. Regardless of what garbage is read from skb->len, it's always going to be at least ETH_ZLEN due to the "max of skb->len and ETH_ZLEN" assignment there.
Quoted 2 lines
btw, given the zerocopy stuff has been disabled for so long, is it worth keeping it ?
I'm pretty sure it should be tossed. I wonder if anyone even tries to build that code these days.
Quoted 9 lines · David Miller
From: Dave Jones <[email protected]>
Date: Sun, 21 Jun 2009 22:40:44 -0400Quoted 6 lines · David Miller
Quoted 4 lines
Is it possible that for some reason
your build is forcingVELOCITY_ZERO_COPY_SUPPORTto be defined for some reason?
Quoted 8 lines
Quoted 4 lines
Memory corruption maybe?
It's especially odd in that it only happens once during boot, and never happens again. This is my firewall/router, so there's more packets going through that box than any other I have.I would not be inclined to this theory. Regardless of what garbage is read from
skb->len, it's always going to be at leastETH_ZLENdue to the "max ofskb->lenandETH_ZLEN" assignment there.
Yeah, seems unlikely. Compiler bug seems unlikely too, given it's a one-off. I'll see if I can come up with some other debug patches to figure out what's going on in there.
Quoted 5 lines
Quoted 2 lines
btw, given the zerocopy stuff has been disabled for so long, is it worth keeping it ?
I'm pretty sure it should be tossed. I wonder if anyone even tries to build that code these days.
Remove unused Velocity zero copy code.
This code hasn't been enabled in years.
drivers/net/via-velocity.c +11 -66
@@ -976,9 +976,6 @@ static int __devinit velocity_found1(struct pci_dev *pdev, const struct pci_devi dev->netdev_ops = &velocity_netdev_ops; dev->ethtool_ops = &velocity_ethtool_ops; -#ifdef VELOCITY_ZERO_COPY_SUPPORT - dev->features |=NETIF_F_SG ; -#endif dev->features |=NETIF_F_HW_VLAN_TX |NETIF_F_HW_VLAN_FILTER |NETIF_F_HW_VLAN_RX ;
@@ -1849,11 +1846,7 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_ 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], pktlen,PCI_DMA_TODEVICE ); -#endif tdinfo->skb_dma[i] =0 ; } }
@@ -2095,13 +2088,6 @@ static int velocity_xmit(struct sk_buff *skb, struct net_device *dev) len =cpu_to_le16 (pktlen); -#ifdef VELOCITY_ZERO_COPY_SUPPORT -if (skb_shinfo (skb)->nr_frags >6 &&__skb_linearize (skb)) { -kfree_skb (skb); -return 0 ; - } -#endif -spin_lock_irqsave (&vptr->lock, flags); index = vptr->tx.curr[qnum];
@@ -2111,59 +2097,18 @@ static int velocity_xmit(struct sk_buff *skb, struct net_device *dev) td_ptr->tdesc1.TCR =TCR0_TIC ; td_ptr->td_buf[0 ].size &= ~TD_QUEUE ; -#ifdef VELOCITY_ZERO_COPY_SUPPORT -if (skb_shinfo (skb)->nr_frags >0 ) { -int nfrags =skb_shinfo (skb)->nr_frags; - tdinfo->skb = skb; -if (nfrags >6 ) { -skb_copy_from_linear_data (skb, tdinfo->buf, skb->len); - tdinfo->skb_dma[0 ] = tdinfo->buf_dma; - td_ptr->tdesc0.len = len; - td_ptr->tx.buf[0 ].pa_low =cpu_to_le32 (tdinfo->skb_dma[0 ]); - td_ptr->tx.buf[0 ].pa_high =0 ; - td_ptr->tx.buf[0 ].size = len;/* queue is 0 anyway */ - tdinfo->nskb_dma =1 ; - }else { -int i =0 ; - tdinfo->nskb_dma =0 ; - tdinfo->skb_dma[i] =pci_map_single (vptr->pdev, skb->data, -skb_headlen (skb),PCI_DMA_TODEVICE ); - - td_ptr->tdesc0.len = len; - -/* FIXME: support 48bit DMA later */ - td_ptr->tx.buf[i].pa_low =cpu_to_le32 (tdinfo->skb_dma); - td_ptr->tx.buf[i].pa_high =0 ; - td_ptr->tx.buf[i].size =cpu_to_le16 (skb_headlen (skb)); - -for (i =0 ; i < nfrags; i++) { -skb_frag_t *frag = &skb_shinfo (skb)->frags[i]; -void *addr = (void *)page_address (frag->page) + frag->page_offset; - - tdinfo->skb_dma[i +1 ] =pci_map_single (vptr->pdev, addr, frag->size,PCI_DMA_TODEVICE ); - - td_ptr->tx.buf[i +1 ].pa_low =cpu_to_le32 (tdinfo->skb_dma[i +1 ]); - td_ptr->tx.buf[i +1 ].pa_high =0 ; - td_ptr->tx.buf[i +1 ].size =cpu_to_le16 (frag->size); - } - tdinfo->nskb_dma = i -1 ; - } +/* +* Map the linear network buffer into PCI space and +* add it to the transmit ring. +*/ + tdinfo->skb = skb; + tdinfo->skb_dma[0 ] =pci_map_single (vptr->pdev, skb->data, pktlen,PCI_DMA_TODEVICE ); + td_ptr->tdesc0.len = len; + td_ptr->td_buf[0 ].pa_low =cpu_to_le32 (tdinfo->skb_dma[0 ]); + td_ptr->td_buf[0 ].pa_high =0 ; + td_ptr->td_buf[0 ].size = len; + tdinfo->nskb_dma =1 ; - } else -#endif - { -/* -* Map the linear network buffer into PCI space and -* add it to the transmit ring. -*/ - tdinfo->skb = skb; - tdinfo->skb_dma[0 ] =pci_map_single (vptr->pdev, skb->data, pktlen,PCI_DMA_TODEVICE ); - td_ptr->tdesc0.len = len; - td_ptr->td_buf[0 ].pa_low =cpu_to_le32 (tdinfo->skb_dma[0 ]); - td_ptr->td_buf[0 ].pa_high =0 ; - td_ptr->td_buf[0 ].size = len; - tdinfo->nskb_dma =1 ; - } td_ptr->tdesc1.cmd =TCPLS_NORMAL + (tdinfo->nskb_dma +1 ) *16 ;if (vptr->vlgrp &&vlan_tx_tag_present (skb)) {
PATCH 1/1 Fix velocity driver unmapping incorrect size.
In reply to David MillerOh man, this was subtle. How did this never bite us before ?
Signature · 4 lines
When a packet is greater than ETH_ZLEN, we end up assigning the boolean result of a comparison to the size we unmap. Signed-off-by: Dave Jones <[email protected]>
drivers/net/via-velocity.c +1 -1
@@ -1847,7 +1847,7 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_*/ if (tdinfo->skb_dma) { - pktlen = (skb->len >ETH_ZLEN ? :ETH_ZLEN ); + pktlen = (skb->len >ETH_ZLEN ? skb->len :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 );
Re: [PATCH] Fix velocity driver unmapping incorrect size.
Quoted 1 line · Dave Jones
Oh man, this was subtle. How did this never bite us before ?
bah. I introduced that back in 59f8e169e25c5fce91826412c38359ecaf940b82
I suck at ternary operators.
Dave
PATCH v2 1/1 Fix velocity driver unmapping incorrect size.
In reply to Dave Jonesthis makes it even clearer what's going on.
Signature · 4 lines
When a packet is greater than ETH_ZLEN, we end up assigning the boolean result of a comparison to the size we unmap. Signed-off-by: Dave Jones <[email protected]>
drivers/net/via-velocity.c +1 -1
@@ -1847,7 +1847,7 @@ static void velocity_free_tx_buf(struct velocity_info *vptr, struct velocity_td_*/ if (tdinfo->skb_dma) { - pktlen = (skb->len >ETH_ZLEN ? :ETH_ZLEN ); + pktlen = max_t(unsigned int, skb->len, 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 );
David Miller
maintainerRe: [PATCH v2] Fix velocity driver unmapping incorrect size.
From: Dave Jones <[email protected]>
Date: Mon, 22 Jun 2009 01:35:30 -0400
Quoted 7 lines
this makes it even clearer what's going on.
Signature · 4 lines
When a packet is greater than ETH_ZLEN, we end up assigning the boolean result of a comparison to the size we unmap. Signed-off-by: Dave Jones <[email protected]>
Looks great Dave, thanks for figuring this out.
I'll queue this up for -stable too.