svn commit: r192027 - head/sys/arm/at91
M. Warner Losh
imp at bsdimp.com
Fri May 15 19:31:13 UTC 2009
In message: <20090515231922.fb760af4.stas at FreeBSD.org>
Stanislav Sedov <stas at FreeBSD.org> writes:
: On Fri, 15 May 2009 09:05:31 -0600 (MDT)
: "M. Warner Losh" <imp at bsdimp.com> mentioned:
: > In message: <20090515141642.ebc06b59.stas at FreeBSD.org>
: > Stanislav Sedov <stas at FreeBSD.org> writes:
: > : On Thu, 14 May 2009 23:35:36 -0600 (MDT)
: > : "M. Warner Losh" <imp at bsdimp.com> mentioned:
: > :
: > : > In message: <20090515092205.6f6d06fa.stas at FreeBSD.org>
: > : > Stanislav Sedov <stas at FreeBSD.org> writes:
: > : > : On Thu, 14 May 2009 21:37:12 -0600 (MDT)
: > : > : "M. Warner Losh" <imp at bsdimp.com> mentioned:
: > : > :
: > : > : > In message: <200905122114.n4CLEag9033208 at svn.freebsd.org>
: > : > : > Stanislav Sedov <stas at FreeBSD.org> writes:
: > : > : > : @@ -926,6 +937,7 @@ atestart_locked(struct ifnet *ifp)
: > : > : > : * tell the hardware to xmit the packet.
: > : > : > : */
: > : > : > : WR4(sc, ETH_TAR, segs.ds_addr);
: > : > : > : + BARRIER(sc, ETH_TAR, 8, BUS_SPACE_BARRIER_WRITE);
: > : > : > : WR4(sc, ETH_TCR, segs.ds_len);
: > : > : >
: > : > : > Why is a barrier needed here?
: > : > : >
: > : > : Writing the TCR register triggers the transmit, so it had to be written
: > : > : strongly after the TAR register. That's why I added the barrier here.
: > : >
: > : > Then shouldn't the barrier be after TCR write? Or does this ensure
: > : > that the write is before TCR?
: > : >
: > :
: > : Yeah, this barrier is to ensure that the TCR register gets written after the
: > : TAR register has been written, not before. I don't think an additional barrier
: > : is needed after the TCR write.
: > Did this fix an observed bug, or is it theoretical? None of Atmel's
: > code does this, but maybe we turn on some flag that reorders writes.
: > On the other hand, I've seen some minor flakiness from time to time
: > that could be explained by reordering....
: > There's likely a bunch of other places where something like this may
: > be needed. The PDC has size/address information, followed by an
: > enable bit. The MCI device has some similar weirdness as well...
: I don't think there're any reordering possible on at91 platform,
: though I need to check first. The bus_space_barrier call is currently
: a no-op on arm platforms, so this modifications were mostly to make
: the code more correct theoretically then fixing any possible real-world
: PDC is the entirely another thing, so it need to be checked separately.
: EMAC doesn't use PDC but a real DMA implementation.
Yes. Understood. Just thinking of other places this might matter.
Any idea if this matters on the AVR32?
Then again, the built-in devices are mapped into uncached memory, so
maybe it just doesn't matter :).
More information about the svn-src-head