Skip to contentSkip to conversation

Missing critical phys_to_virt in lib/swiotlb.c

PATCH In review David Moore 5 participants 7 messages Feb 4, 2007 – Feb 6, 2007 Library code

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Missing critical phys_to_virt in lib/swiotlb.c

From: David Moore <[email protected]>

Adds missing call to phys_to_virt() in the lib/swiotlb.c:swiotlb_sync_sg() function. Without this change, a kernel panic will always occur whenever a SWIOTLB bounce buffer from a scatter-gather list gets synced.

---

This change was originally part of a larger patch by Jan Beulich, which was more extensive and doesn't look destined to make it into 2.6.20:
http://kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.20-rc6/2.6.20-rc6-mm3/broken-out/ia64-swiotlb-bug-fixes.patch

However, considering the severity of this one-liner bug, I would like to request that this simplified patch make it into 2.6.20, despite how close we are to the final cut. It fixes real crashes:
http://lists.opensuse.org/opensuse-bugs/2006-12/msg02943.html
http://qa.mandriva.com/show_bug.cgi?id=28224
http://www.pchdtv.com/forum/viewtopic.php?t=2063&sid=a959a14a4c2db0eebaab7b0df56103ce

lib/swiotlb.c renamed from lib/swiotlb.c.orig +1 -1
@@ -750,7 +750,7 @@ swiotlb_sync_sg(struct device *hwdev, st
 
 	for (i = 0; i < nelems; i++, sg++)
 		if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
-			sync_single(hwdev, (void *) sg->dma_address,
+			sync_single(hwdev, phys_to_virt(sg->dma_address),
 				    sg->dma_length, dir, target);
 }
 
#1 lore raw

David Moore wrote:
[...]

Quoted 3 lines

considering the severity of this one-liner bug, I would like to request that this simplified patch make it into 2.6.20, despite how close we are to the final cut.

So we were too close. Maybe the -stable team likes to have it in 2.6.20.1.

Quoted 4 lines

It fixes real crashes:
http://lists.opensuse.org/opensuse-bugs/2006-12/msg02943.html
http://qa.mandriva.com/show_bug.cgi?id=28224
http://www.pchdtv.com/forum/viewtopic.php?t=2063&sid=a959a14a4c2db0eebaab7b0df56103ce

and FireWire crashes too.

Quoted 11 lines
lib/swiotlb.c renamed from lib/swiotlb.c.orig +1 -1
@@ -750,7 +750,7 @@ swiotlb_sync_sg(struct device *hwdev, st
 
 	for (i = 0; i < nelems; i++, sg++)
 		if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
-			sync_single(hwdev, (void *) sg->dma_address,
+			sync_single(hwdev, phys_to_virt(sg->dma_address),
 				    sg->dma_length, dir, target);
 }
 
Signature · 3 lines
Stefan Richter
-=====-=-=== --=- --=--
http://arcgraph.de/sr/
maintainer
#2 lore raw
In reply to David Moore
Quoted 24 lines · On Sun, 04 Feb 2007 13:39:40 -0500 David Moore <[email protected]> wrote:

From: David Moore <[email protected]>

Adds missing call to phys_to_virt() in the lib/swiotlb.c:swiotlb_sync_sg() function. Without this change, a kernel panic will always occur whenever a SWIOTLB bounce buffer from a scatter-gather list gets synced.

---

This change was originally part of a larger patch by Jan Beulich, which was more extensive and doesn't look destined to make it into 2.6.20:
http://kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.20-rc6/2.6.20-rc6-mm3/broken-out/ia64-swiotlb-bug-fixes.patch

However, considering the severity of this one-liner bug, I would like to request that this simplified patch make it into 2.6.20, despite how close we are to the final cut. It fixes real crashes:
http://lists.opensuse.org/opensuse-bugs/2006-12/msg02943.html
http://qa.mandriva.com/show_bug.cgi?id=28224
http://www.pchdtv.com/forum/viewtopic.php?t=2063&sid=a959a14a4c2db0eebaab7b0df56103ce

lib/swiotlb.c swiotlb_sync_sg(struct device *hwdev, st · line 756

Quoted 8 lines
lib/swiotlb.c renamed from lib/swiotlb.c.orig +1 -1
@@ -750,7 +750,7 @@ swiotlb_sync_sg(struct device *hwdev, st
 
 	for (i = 0; i < nelems; i++, sg++)
 		if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
-			sync_single(hwdev, (void *) sg->dma_address,
+			sync_single(hwdev, phys_to_virt(sg->dma_address),
 				    sg->dma_length, dir, target);
 }
 

argh. I didn't know that Jan's patches fixed crashes. I thought they were ia64-only things.

Who maintains the swiotlb code?

I shall queue this up, tag it for 2.6.20.1, then I'll fix up Jan's alleged ia64 patches to account for that. I'll still push it through Tony, but the x86_64/ia64 linkage here seems to be a source of problems.

#3 lore raw
Quoted 28 lines · Andrew Morton
Quoted 25 lines · On Sun, 04 Feb 2007 13:39:40 -0500 David Moore <[email protected]> wrote:

From: David Moore <[email protected]>

Adds missing call to phys_to_virt() in the lib/swiotlb.c:swiotlb_sync_sg() function. Without this change, a kernel panic will always occur whenever a SWIOTLB bounce buffer from a scatter-gather list gets synced.

---

This change was originally part of a larger patch by Jan Beulich, which was more extensive and doesn't look destined to make it into 2.6.20:
http://kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.20-rc6/2.6.20-rc6-mm3/broken-out/ia64-swiotlb-bug-fixes.patch

However, considering the severity of this one-liner bug, I would like to request that this simplified patch make it into 2.6.20, despite how close we are to the final cut. It fixes real crashes:
http://lists.opensuse.org/opensuse-bugs/2006-12/msg02943.html
http://qa.mandriva.com/show_bug.cgi?id=28224
http://www.pchdtv.com/forum/viewtopic.php?t=2063&sid=a959a14a4c2db0eebaab7b0df56103ce

lib/swiotlb.c renamed from lib/swiotlb.c.orig +0 -0
@@ -750,7 +750,7 @@ swiotlb_sync_sg(struct device *hwdev, st
 
Quoted 10 lines

lib/swiotlb.c swiotlb_sync_sg(struct device *hwdev, st · line 756

Quoted 7 lines
lib/swiotlb.c renamed from lib/swiotlb.c.orig +1 -1
 	for (i = 0; i < nelems; i++, sg++)
 		if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
-			sync_single(hwdev, (void *) sg->dma_address,
+			sync_single(hwdev, phys_to_virt(sg->dma_address),
 				    sg->dma_length, dir, target);
 }
 

argh. I didn't know that Jan's patches fixed crashes. I thought they were ia64-only things.

Sounds weird. If this really didn't work much more should be broken (e.g. no cdroms/sound on Intel x86-64 boxes with >4GB)
I'm a little sceptical. Perhaps the TV driver is doing something bogus here?

Also I haven't heard of this problem before at all and I'm sure I would have if sounds/cdroms were broken.

Shouldn't be applied without further analysis.

Quoted 1 line

Who maintains the swiotlb code?

Nobody. But I hacked last on it.

-Andi

#4 lore raw
Quoted 1 line · Andi Kleen
Quoted 1 line
Quoted 1 line

From: David Moore <[email protected]>

...

Quoted 15 lines
Quoted 15 lines
Quoted 15 lines

It fixes real crashes:
http://lists.opensuse.org/opensuse-bugs/2006-12/msg02943.html
http://qa.mandriva.com/show_bug.cgi?id=28224
http://www.pchdtv.com/forum/viewtopic.php?t=2063&sid=a959a14a4c2db0eebaab7b0df56103ce

lib/swiotlb.c renamed from lib/swiotlb.c.orig +1 -1
@@ -750,7 +750,7 @@ swiotlb_sync_sg(struct device *hwdev, st
 
 	for (i = 0; i < nelems; i++, sg++)
 		if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
-			sync_single(hwdev, (void *) sg->dma_address,
+			sync_single(hwdev, phys_to_virt(sg->dma_address),
 				    sg->dma_length, dir, target);
 }

...

Quoted 2 lines

Sounds weird. If this really didn't work much more should be broken (e.g. no cdroms/sound on Intel x86-64 boxes with >4GB)

Yes, it's weird; e.g. it seems that reports of that on linux1394-user and -devel came in only recently. But I may have missed something. (OTOH, this bug was partially hidden for FireWire users because one of the relevant FireWire drivers was lacking respective _sync_sg calls.)

Quoted 2 lines

I'm a little sceptical. Perhaps the TV driver is doing something bogus here?

sync_single() definitely wants a virtual address there. sg->dma_address is a physical address. The bug and the fix are obvious.

Unfortunately an author of lib/swiotlb.c chose to call many variables holding *virtual* addresses "dma_addr". Note how that file at the same time contains variables like "dma_addr_t dma_handle". A recipe for disaster.

Quoted 4 lines

Also I haven't heard of this problem before at all and I'm sure I would have if sounds/cdroms were broken.

Shouldn't be applied without further analysis.

Maybe some recent changes elsewhere cause more frequent use of swiotlb-driven bounce buffers? That should be verified.

Or maybe people are deploying affected hardware more often now?

Signature · 3 lines
Stefan Richter
-=====-=-=== --=- --==-
http://arcgraph.de/sr/
#5 lore raw
Quoted 3 lines · I wrote:

Unfortunately an author of lib/swiotlb.c chose to call many variables holding *virtual* addresses "dma_addr". Note how that file at the same time contains variables like "dma_addr_t dma_handle".

And there is even one occurrence of a "dma_addr" holding a physical address/ bus address, unlike the other dma_addr's:
int
swiotlb_dma_mapping_error(dma_addr_t dma_addr)

{
	return (dma_addr == virt_to_phys(io_tlb_overflow_buffer));
}
Signature · 3 lines
Stefan Richter
-=====-=-=== --=- --==-
http://arcgraph.de/sr/
#6 lore raw
In reply to Andi Kleen
Quoted 1 line

Shouldn't be applied without further analysis.

I don't think further analysis is required before this change be done, the missing conversion is rather obvious when comparing to all other functions in that file.

Whether the observed crashes really origin from the missing bits here is a different question.

Jan