Skip to contentSkip to conversation

agpgart: when telling user you'll corrupt his data, at least do it at KERN_CRIT

PATCH Pavel Machek 4 participants 12 messages Mar 28, 2008 – Apr 4, 2008 x86

lore
Share & tools
lore
#0 lore raw

IOMMU off means very bad stuff may happen, like data corruption on your hard drives. At least tells users this is serious...

Plus, agp_amd64_init should go into .h file.

arch/x86/kernel/pci-gart_64.c +5 -7
index faf3229..19e4f74 100644
@@ -615,13 +615,11 @@ static __init int init_k8_gatt(struct ag
 
  nommu:
 	/* Should not happen anymore */
-	printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
-	       KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
+	printk(KERN_CRIT "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
+	       KERN_CRIT "PCI-DMA: 32bit PCI IO may malfunction.\n");
 	return -1;
 }
 
-extern int agp_amd64_init(void);
-
 static const struct dma_mapping_ops gart_dma_ops = {
 	.mapping_error			= NULL,
 	.map_single			= gart_map_single,
@@ -692,9 +690,9 @@ #endif
 	    !gart_iommu_aperture ||
 	    (no_agp && init_k8_gatt(&info) < 0)) {
 		if (end_pfn > MAX_DMA32_PFN) {
-			printk(KERN_ERR "WARNING more than 4GB of memory "
-					"but GART IOMMU not available.\n"
-			       KERN_ERR "WARNING 32bit PCI may malfunction.\n");
+			printk(KERN_CRIT "More than 4GB of memory "
+			       	          "but GART IOMMU not available.\n"
+			       KERN_CRIT "32bit PCI may malfunction.\n");
 		}
 		return;
 	}
include/asm-x86/agp.h +3 -0
index e4004a9..657c25b 100644
@@ -32,4 +32,7 @@ #define alloc_gatt_pages(order)		\
 #define free_gatt_pages(table, order)	\
 	free_pages((unsigned long)(table), (order))
 
+/* Only available in certain configs */
+extern int agp_amd64_init(void);
+
 #endif
Signature · 3 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
pomozte zachranit klanovicky les:  http://www.ujezdskystrom.info/
#1 lore raw
Quoted 2 lines · Pavel Machek

IOMMU off means very bad stuff may happen, like data corruption on

Actually swiotlb takes over then in this case. The printk really predates the introduction of swiotlb and could be actually removed. That is why swiotlb is always SELECTed for gart too.

Quoted 1 line

your hard drives. At least tells users this is serious...

No there shouldn't be any data corruption with the default options.

Also especially hard disk drivers are expected (and generally do) check the dma_map_sg() etc. return values and bail out if the IOMMU fails e.g. due to overflow.

-Andi

#2 lore raw

agpgart: scary messages are fortunately obsolete

Fix obsolete printks in aperture-64. We used not to handle missing agpgart, but we handle it okay now.

arch/x86/kernel/pci-gart_64.c +5 -5
index faf3229..3939c6d 100644
@@ -615,13 +615,13 @@ static __init int init_k8_gatt(struct ag
 
  nommu:
 	/* Should not happen anymore */
-	printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
-	       KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
+	printk(KERN_WARNING "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
+	       KERN_WARNING "falling back to iommu=soft.\n");
 	return -1;
 }
 
 extern int agp_amd64_init(void);
 
 static const struct dma_mapping_ops gart_dma_ops = {
 	.mapping_error			= NULL,
 	.map_single			= gart_map_single,
@@ -692,9 +690,9 @@ #endif
 	    !gart_iommu_aperture ||
 	    (no_agp && init_k8_gatt(&info) < 0)) {
 		if (end_pfn > MAX_DMA32_PFN) {
-			printk(KERN_ERR "WARNING more than 4GB of memory "
-					"but GART IOMMU not available.\n"
-			       KERN_ERR "WARNING 32bit PCI may malfunction.\n");
+			printk(KERN_WARNING "More than 4GB of memory "
+			       	          "but GART IOMMU not available.\n"
+			       KERN_WARNING "falling back to iommu=soft.\n");
 		}
 		return;
 	}
Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#3 lore raw

Re: agpgart: scary messages are fortunately obsolete

arch/x86/kernel/pci-gart_64.c static __init int init_k8_gatt(struct ag · line 619

Quoted 5 lines
arch/x86/kernel/pci-gart_64.c +2 -2
 	/* Should not happen anymore */
-	printk(KERN_ERR "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
-	       KERN_ERR "PCI-DMA: 32bit PCI IO may malfunction.\n");
+	printk(KERN_WARNING "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
+	       KERN_WARNING "falling back to iommu=soft.\n");

It might also fall back to another IOMMU (Calgary, Intel, ...) soft is just last resort.

You can probably just remove it.

-Andi

#4 lore raw

Re: agpgart: scary messages are fortunately obsolete

In reply to Pavel Machek
Quoted 2 lines · Pavel Machek

Fix obsolete printks in aperture-64. We used not to handle missing agpgart, but we handle it okay now.

thanks Pavel, applied.

Ingo

#5 lore raw
In reply to Pavel Machek
Quoted 2 lines · Pavel Machek

IOMMU off means very bad stuff may happen, like data corruption on your hard drives. At least tells users this is serious...

ouch. Please at minimum lets turn this into a panic(), but best would be to trim memory in this case, hm?

Ingo

#6 lore raw
Quoted 4 lines · Ingo Molnar
Quoted 2 lines

IOMMU off means very bad stuff may happen, like data corruption on your hard drives. At least tells users this is serious...

ouch. Please at minimum lets turn this into a panic(), but best would be to trim memory in this case, hm?

Are you suggesting to panic on all 64bit machines having more than 4G RAM without IOMMU turned on? Why, if swiotlb can still work nicely in such situations?

Signature · 2 lines
Jiri Kosina
SUSE Labs
#7 lore raw
Quoted 10 lines · Jiri Kosina
Quoted 4 lines · Ingo Molnar
Quoted 2 lines

IOMMU off means very bad stuff may happen, like data corruption on your hard drives. At least tells users this is serious...

ouch. Please at minimum lets turn this into a panic(), but best would be to trim memory in this case, hm?

Are you suggesting to panic on all 64bit machines having more than 4G RAM without IOMMU turned on? Why, if swiotlb can still work nicely in such situations?

no, but we should avoid Pavel's disk corruption scenario ;-)

Ingo

#8 lore raw
In reply to Ingo Molnar
Quoted 8 lines · Ingo Molnar
Quoted 2 lines · Pavel Machek

IOMMU off means very bad stuff may happen, like data corruption on your hard drives. At least tells users this is serious...

ouch. Please at minimum lets turn this into a panic(), but best would be to trim memory in this case, hm?

Andi tells me we already fallback to swiotlb, so messages are not on wrong loglevel; they contain confusing/obsolete text.

Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#9 lore raw
Quoted 4 lines · Pavel Machek
Quoted 3 lines · Ingo Molnar

* Pavel Machek <[email protected]> wrote:

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

IOMMU off means very bad stuff may happen, like data corruption on your hard drives. At least tells users this is serious...

ouch. Please at minimum lets turn this into a panic(), but best would be to trim memory in this case, hm?

Andi tells me we already fallback to swiotlb, so messages are not on wrong loglevel; they contain confusing/obsolete text.

yes that is the theory - but how did your disk get corrupted in practice? ;-)

Ingo

#10 lore raw
Quoted 2 lines

yes that is the theory - but how did your disk get corrupted in practice? ;-)

I suspect after wakeup due to the wrong ordering in resume (agp resume function called too late).

The IOMMU code unfortunately doesn't know that the GART has gone away.

The right fix for that is probably to move the GART resume code into some shared file and make it a sysdev.

-Andi

#11 lore raw
In reply to Ingo Molnar
Quoted 10 lines · Ingo Molnar
Quoted 7 lines · Pavel Machek
Quoted 6 lines · Ingo Molnar
Quoted 2 lines · Pavel Machek

IOMMU off means very bad stuff may happen, like data corruption on your hard drives. At least tells users this is serious...

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

ouch. Please at minimum lets turn this into a panic(), but best would be to trim memory in this case, hm?

Andi tells me we already fallback to swiotlb, so messages are not on wrong loglevel; they contain confusing/obsolete text.

yes that is the theory - but how did your disk get corrupted in practice? ;-)

Well, if you try suspend/resume, without iommu suspend/resume support (which is not there in some cases), you get nasty data corruption. ("PCI DMA will does not work").

Then, I saw [obsolete] printk(KERN_ERR) that told me that PCI DMA will not work, and happily continued to boot.

I'll fix the messages.
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html