declance: Restore tx descriptor ring locking
Subsystems (3)
PATCH 1/1 declance: Restore tx descriptor ring locking
A driver overhaul on 29 Feb 2000 (!) broke locking around fiddling with
the tx descriptor ring in start_xmit(); a follow-on "fix" removed the broken remnants altogether. Here's a patch to restore proper locking in the function -- the complement in the interrupt handler has been correct all the time.
This *may* have been the reason for the occasional confusion of the chip
-- triggering a tx timeout followed by a chip reset sequence -- seen on R4k-based DECstations with the onboard Ethernet interface. Another theory is the confusion is due to an unindentified problem -- perhaps a silicon erratum -- associated with the variation of the MT ASIC used to interface the R4k CPU to the rest of the system on these computers; with its aggressive write-back buffering the design is particularly weakly ordered when it comes to MMIO (in the absence of ordering barriers uncached reads are allowed to bypass earlier uncached writes, even if to the same location), which may trigger all kinds of corner cases in peripheral hardware as well as software.
Either way this piece of code is buggy.
---
Works for me. Dave, please apply.
Maciej
patch-mips-2.6.27-rc8-20081004-declance-lock-2
drivers/net/declance.c +5 -0
@@ -906,6 +906,7 @@ static int lance_start_xmit(struct sk_bustruct lance_private *lp =netdev_priv (dev);volatile struct lance_regs *ll = lp->ll;volatile u16 *ib = (volatile u16 *)dev->mem_start; +unsigned long flags;int entry, len; len = skb->len;
@@ -918,6 +919,8 @@ static int lance_start_xmit(struct sk_bu dev->stats.tx_bytes += len; +spin_lock_irqsave (&lp->lock, flags); + entry = lp->tx_new; *lib_ptr (ib, btx_ring[entry].length, lp->type) = (-len); *lib_ptr (ib, btx_ring[entry].misc, lp->type) =0 ;
@@ -936,6 +939,8 @@ static int lance_start_xmit(struct sk_bu/* Kick the lance: transmit now */ writereg (&ll->rdp,LE_C0_INEA |LE_C0_TDMD ); +spin_unlock_irqrestore (&lp->lock, flags); + dev->trans_start = jiffies;dev_kfree_skb (skb);
David Miller
maintainerFrom: "Maciej W. Rozycki" <[email protected]>
Date: Mon, 8 Jun 2009 02:47:05 +0100 (WEST)
Quoted 18 lines
A driver overhaul on 29 Feb 2000 (!) broke locking around fiddling with the tx descriptor ring in
start_xmit(); a follow-on "fix" removed the broken remnants altogether. Here's a patch to restore proper locking in the function -- the complement in the interrupt handler has been correct all the time.This *may* have been the reason for the occasional confusion of the chip
-- triggering a tx timeout followed by a chip reset sequence -- seen on R4k-based DECstations with the onboard Ethernet interface. Another theory is the confusion is due to an unindentified problem -- perhaps a silicon erratum -- associated with the variation of the MT ASIC used to interface the R4k CPU to the rest of the system on these computers; with its aggressive write-back buffering the design is particularly weakly ordered when it comes to MMIO (in the absence of ordering barriers uncached reads are allowed to bypass earlier uncached writes, even if to the same location), which may trigger all kinds of corner cases in peripheral hardware as well as software.
Quoted 3 lines
Either way this piece of code is buggy.
Signed-off-by Maciej W. Rozycki <[email protected]>
Looks good to me, applied.