Bug 75317
| Summary: | [busdma] [patch] ATA DMA broken on PCalpha | ||
|---|---|---|---|
| Product: | Base System | Reporter: | Sven Petai <hadara> |
| Component: | alpha | Assignee: | freebsd-alpha (Nobody) <alpha> |
| Status: | Closed FIXED | ||
| Severity: | Affects Only Me | CC: | sos |
| Priority: | Normal | ||
| Version: | 6.0-CURRENT | ||
| Hardware: | Any | ||
| OS: | Any | ||
|
Description
Sven Petai
2004-12-20 16:40:12 UTC
I managed to put a wrong link in the original bugreport. The correct dmesg -v output is located at http://bsd.ee/~hadara/debug/pcalpha/pcalpha_panic_08.11.2004.txt I think I now understand what causes this bug, First of all there seems to be a slight bug in Alphas bus_dmamap_load(), which could be fixed with following patch: --- sys/alpha/alpha/busdma_machdep.c.orig Wed Dec 22 17:11:27 2004 +++ sys/alpha/alpha/busdma_machdep.c Wed Dec 22 17:13:57 2004 @@ -581,7 +581,8 @@ if (sg->ds_len == 0) { sg->ds_addr = paddr + alpha_XXX_dmamap_or; sg->ds_len = size; - } else if (paddr == nextpaddr) { + } else if (paddr == nextpaddr && + (sg->ds_len + size) <= dmat->maxsegsz) { sg->ds_len += size; } else { /* Go to the next segment */ without that we really could return larger chunks than specified by maxsegsz to bus_dma_tag_create function. anyway this still doesn't make things work correctly for me, because the real problem seems to be the Pyxis page crossing bug. Basically it comes down to corrupting DMA transfers larger than 8k. It didn't cause problems before, since we never did larger than PAGE_SIZE transfers before the ATA dma change mentioned in the original report. There's a detection code for the buggy chip @ src/sys/alpha/pci/cia.c but it's little too naive, since it assumes only DEC_ST550 can have it, in reality it seems to be used in some very early revisions of 164LX(SX too?). But there doesn't seem to be a reliable way to detect if we have the faulty chip since it was worked around in later revisions by doing some changes elsewhere. One of the possible easy solutions would be to hack ata_dmaalloc() to use PAGE_SIZE as max segment size argument to the bus_dma_tag_create function if machine has Pyxis chip at all, no matter if it's faulty or not. Would that be acceptable and if so what is the best way to propagate knowledge about existence of this chip from cia driver to ATA ? This occurs on my 164SX running 6.1-PRERELEASE and can be reproduced on 6.0-RELEASE and 5.4-STABLE as well. None of the above patches fix the problem on 6.0. With ATA drives in the system, and DMA on, machine checks occur every few hours at best, every few minutes at worst. With the same drives attached by firewire, I haven't had a single machine check ever. State Changed From-To: open->closed alpha is no longer supported and time will not be invested to get things fixed. Our apologies for our lack of attention for this issue. State Changed From-To: closed->open People are actively working on this, reopen the ticket (Thanks John, again my apologies) jhb 2007-11-27 17:43:50 UTC
FreeBSD src repository
Modified files: (Branch: RELENG_6)
sys/alpha/alpha busdma_machdep.c
Log:
Cleanup the alpha bus dma code a bit and sync it up with i386. Changes
include:
- Honor alignment and boundary restrictions on DMA tags by using bounce
pages for misaligned buffers and not coalescing pages if the resulting
segment would cross a boundary.
- Teach the _bus_dmamap_load_buffer() helper function to use bounce pages
when needed and change bus_dmamap_load() to use the helper function
instead of largely duplicating it. As a side effect, this enables bounce
page support for the other load routines (load_mbuf(), load_mbuf_sg(),
and load_uio()).
Honoring the boundary restrictions partially helps with the Alpha ATA DMA
problem. More work is needed for that however (and forthcoming).
PR: alpha/75317
Tested by: wilko
Approved by: re (kensmith)
Revision Changes Path
1.51.2.2 +155 -158 src/sys/alpha/alpha/busdma_machdep.c
_______________________________________________
cvs-all@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/cvs-all
To unsubscribe, send any mail to "cvs-all-unsubscribe@freebsd.org"
jhb 2007-12-10 20:14:16 UTC
FreeBSD src repository
Modified files: (Branch: RELENG_6)
sys/alpha/alpha busdma_machdep.c
sys/alpha/include md_var.h
sys/alpha/pci cia.c
Log:
- Add a workaround for the DMA bugs on some alpha chipsets that ATA DMA
trips over often. Specifically, in these chipsets DMA transfers that
cross a page boundary result in data corruption. The workaround is to
not allow any DMA transfers for non-static DMA maps (i.e. "real"
transfers as opposed to work areas allocated with bus_dmamem_alloc()) to
cross a page in a single S/G element. This behavior is enabled by
setting 'busdma_pyxis_bug' to 1.
- Add a new tunable 'machdep.busdma_pyxis_bug' that can be used to enable
the workaround from the loader. This can be used to enable it on
chipsets where we don't automatically enable it.
- Auto-enable the workaround for buggy PYXIS 1 chipsets supported via
cia(4).
PR: alpha/75317
Revision Changes Path
1.51.2.3 +23 -6 src/sys/alpha/alpha/busdma_machdep.c
1.23.10.1 +1 -0 src/sys/alpha/include/md_var.h
1.44.2.1 +1 -0 src/sys/alpha/pci/cia.c
_______________________________________________
cvs-all@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/cvs-all
To unsubscribe, send any mail to "cvs-all-unsubscribe@freebsd.org"
State Changed From-To: open->closed This is believed to be fixed in 6.3 based on at least one positive user report. |
