Add missing read memory barrier to Intel Ethernet device drivers
Subsystems (2)
PATCH 1/1 Add missing read memory barrier to Intel Ethernet device drivers
From: Milton Miller <[email protected]>
The PowerPC architecture does not require loads to independent bytes to be ordered without adding an explicit barrier.
In ixgbe_clean_rx_irq we load the status bit then load the packet data. With packet split disabled if these loads go out of order we get a stale packet, but we will notice the bad sequence numbers and drop it.
The problem occurs with packet split enabled where the TCP/IP header and data are in different descriptors. If the reads go out of order we may have data that doesn't match the TCP/IP header. Since we use hardware checksumming this bad data is never verified and it makes it all the way to the application.
This bug was found during stress testing and adding this barrier has been shown to fix it. The bug can manifest as a data integrity issue (bad payload data) or as a BUG in skb_pull().
This was a nasty bug to hunt down, if people agree with the fix I think it's a candidate for stable.
Previously Submitted to e1000-devel only for ixgbe
http://marc.info/?l=e1000-devel&m=126593062701537&w=3
We've now seen this problem hit with other device drivers (e1000e mostly) So I'm resubmitting with fixes for other Intel Device Drivers with similar issues: ixgb, e100, e1000, and e1000e
---
drivers/net/e1000/e1000_main.c +1 -0
@@ -3448,6 +3448,7 @@ static bool e1000_clean_tx_irq(struct e1while ((eop_desc->upper.data &cpu_to_le32 (E1000_TXD_STAT_DD )) && (count < tx_ring->count)) {bool cleaned = false; +rmb ();/* read buffer_info after eop_desc */ for ( ; !cleaned; count++) { tx_desc =E1000_TX_DESC (*tx_ring, i); buffer_info = &tx_ring->buffer_info[i];
drivers/net/e1000e/netdev.c +1 -0
@@ -984,6 +984,7 @@ static bool e1000_clean_tx_irq(struct e1while ((eop_desc->upper.data &cpu_to_le32 (E1000_TXD_STAT_DD )) && (count < tx_ring->count)) {bool cleaned = false; +rmb ();/* read buffer_info after eop_desc */ for (; !cleaned; count++) { tx_desc =E1000_TX_DESC (*tx_ring, i); buffer_info = &tx_ring->buffer_info[i];
drivers/net/ixgb/ixgb_main.c +1 -0
@@ -1816,6 +1816,7 @@ ixgb_clean_tx_irq(struct ixgb_adapter *awhile (eop_desc->status &IXGB_TX_DESC_STATUS_DD ) { +rmb ();/* read buffer_info after eop_desc */ for (cleaned = false; !cleaned; ) { tx_desc =IXGB_TX_DESC (*tx_ring, i); buffer_info = &tx_ring->buffer_info[i];
drivers/net/e100.c +1 -0
@@ -1779,6 +1779,7 @@ static int e100_tx_clean(struct nic *nicfor (cb = nic->cb_to_clean; cb->status &cpu_to_le16 (cb_complete); cb = nic->cb_to_clean = cb->next) { +rmb ();/* read skb after status */ netif_printk(nic, tx_done, KERN_DEBUG, nic->netdev, "cb[%d]->status = 0x%04X\n", (int)(((void*)cb - (void*)nic->cbs)/sizeof(struct cb)),
drivers/net/ixgbe/ixgbe_main.c +1 -0
@@ -748,6 +748,7 @@ static bool ixgbe_clean_tx_irq(struct ixwhile ((eop_desc->wb.status &cpu_to_le32 (IXGBE_TXD_STAT_DD )) && (count < tx_ring->work_limit)) {bool cleaned = false; +rmb ();/* read buffer_info after eop_desc */ for ( ; !cleaned; count++) {struct sk_buff *skb; tx_desc =IXGBE_TX_DESC_ADV (*tx_ring, i);
------------------------------------------------------------------------------
This SF.net email is sponsored by Sprint
What will you do first with EVO, the first 4G phone?
Visit sprint.com/first -- http://p.sf.net/sfu/sprint-com-first
_______________________________________________
E1000-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/e1000-devel
To learn more about Intel® Ethernet, visit http://communities.intel.com/community/wired
Quoted 25 lines · Sonny Rao
From: Milton Miller <[email protected]>
The PowerPC architecture does not require loads to independent bytes to be ordered without adding an explicit barrier.
In
ixgbe_clean_rx_irqwe load the status bit then load the packet data. With packet split disabled if these loads go out of order we get a stale packet, but we will notice the bad sequence numbers and drop it.The problem occurs with packet split enabled where the TCP/IP header and data are in different descriptors. If the reads go out of order we may have data that doesn't match the TCP/IP header. Since we use hardware checksumming this bad data is never verified and it makes it all the way to the application.
This bug was found during stress testing and adding this barrier has been shown to fix it. The bug can manifest as a data integrity issue (bad payload data) or as a BUG in
skb_pull().This was a nasty bug to hunt down, if people agree with the fix I think it's a candidate for stable.
Previously Submitted to e1000-devel only for ixgbe
Quoted 9 lines
We've now seen this problem hit with other device drivers (e1000e mostly) So I'm resubmitting with fixes for other Intel Device Drivers with similar issues: ixgb, e100, e1000, and e1000e
Signed-off-by Milton Miller <[email protected]>Signed-off-by Anton Blanchard <[email protected]>Signed-off-by Sonny Rao <[email protected]>cc stable <[email protected]>
Thanks Milton, I have added the patch to my queue.
Signature · 2 lines
Cheers, Jeff