Skip to contentSkip to conversation

6175ddf06b6172046a329e3abfd9c901a43efd2e breaks matroxfb console

DISCUSSION Ondrej Zary 2 participants 7 messages Aug 2, 2010 – Aug 3, 2010 Framebuffer layer

lore
Share & tools
lore
#0 lore raw

Hello,
matroxfb (at least with Mystique and Mystique 220) stopped working in 2.6.34 - the screen is completely corrupted. Bisection shows that 6175ddf06b6172046a329e3abfd9c901a43efd2e is first bad commit.

Reverting 6175ddf06b6172046a329e3abfd9c901a43efd2e in 2.6.34 fixes the problem (1c5b9069e12e20d2fe883076ae0bf73966492108 must be reverted first).

Signature · 1 line
Ondrej Zary
#1 lore raw
Quoted 8 lines · Ondrej Zary

Hello,
matroxfb (at least with Mystique and Mystique 220) stopped working in 2.6.34 - the screen is completely corrupted. Bisection shows that 6175ddf06b6172046a329e3abfd9c901a43efd2e is first bad commit.

Reverting 6175ddf06b6172046a329e3abfd9c901a43efd2e in 2.6.34 fixes the problem (1c5b9069e12e20d2fe883076ae0bf73966492108 must be reverted first).

Sounds like another driver which used memcpy_toio() when it should have used iowrite32_rep() or __iowrite32_copy().

Hmm... is __iowrite32_copy() and iowrite32_rep() redundant? If so, we should get rid of the former.

-hpa

#2 lore raw
Quoted 10 lines · On Monday 02 August 2010 23:04:28 H. Peter Anvin wrote:
Quoted 8 lines · Ondrej Zary

Hello,
matroxfb (at least with Mystique and Mystique 220) stopped working in 2.6.34 - the screen is completely corrupted. Bisection shows that 6175ddf06b6172046a329e3abfd9c901a43efd2e is first bad commit.

Reverting 6175ddf06b6172046a329e3abfd9c901a43efd2e in 2.6.34 fixes the problem (1c5b9069e12e20d2fe883076ae0bf73966492108 must be reverted first).

Quoted 5 lines

Sounds like another driver which used memcpy_toio() when it should have used iowrite32_rep() or __iowrite32_copy().

Hmm... is __iowrite32_copy() and iowrite32_rep() redundant? If so, we should get rid of the former.

There is a wrapper in drivers/video/matrox/matroxfb_base.h (see below) with some comment. So this commit changed one of the three points?

static inline void mga_memcpy_toio(vaddr_t va, const void* src, int len) { #if defined(__alpha__) || defined(__i386__) || defined(__x86_64__)
/*
* memcpy_toio works for us if:
* (1) Copies data as 32bit quantities, not byte after byte,
* (2) Performs LE ordered stores, and
* (3) It copes with unaligned source (destination is guaranteed to be page
* aligned and length is guaranteed to be multiple of 4).
*/
memcpy_toio(va.vaddr, src, len);

#else
        u_int32_t __iomem* addr = va.vaddr;

        if ((unsigned long)src & 3) {
                while (len >= 4) {
                        fb_writel(get_unaligned((u32 *)src), addr);
                        addr++;
                        len -= 4;
                        src += 4;
                }
        } else {
                while (len >= 4) {
                        fb_writel(*(u32 *)src, addr);
                        addr++;
                        len -= 4;
                        src += 4;
                }
        }

#endif
}

Signature · 1 line
Ondrej Zary
#3 lore raw
Quoted 20 lines · Ondrej Zary
Quoted 10 lines · On Monday 02 August 2010 23:04:28 H. Peter Anvin wrote:
Quoted 8 lines · Ondrej Zary

Hello,
matroxfb (at least with Mystique and Mystique 220) stopped working in 2.6.34 - the screen is completely corrupted. Bisection shows that 6175ddf06b6172046a329e3abfd9c901a43efd2e is first bad commit.

Reverting 6175ddf06b6172046a329e3abfd9c901a43efd2e in 2.6.34 fixes the problem (1c5b9069e12e20d2fe883076ae0bf73966492108 must be reverted first).

Quoted 5 lines

Sounds like another driver which used memcpy_toio() when it should have used iowrite32_rep() or __iowrite32_copy().

Hmm... is __iowrite32_copy() and iowrite32_rep() redundant? If so, we should get rid of the former.

There is a wrapper in drivers/video/matrox/matroxfb_base.h (see below) with some comment. So this commit changed one of the three points?

Quoted 11 lines

static inline void mga_memcpy_toio(vaddr_t va, const void* src, int len) { #if defined(__alpha__) || defined(__i386__) || defined(__x86_64__)
/*
* memcpy_toio works for us if:
* (1) Copies data as 32bit quantities, not byte after byte,
* (2) Performs LE ordered stores, and
* (3) It copes with unaligned source (destination is guaranteed to be page
* aligned and length is guaranteed to be multiple of 4).
*/
memcpy_toio(va.vaddr, src, len);
#else

Yes, point (1) is not guaranteed by memcpy_toio().

-hpa

#4 lore raw
Quoted 32 lines · On Monday 02 August 2010 23:15:33 H. Peter Anvin wrote:
Quoted 28 lines · Ondrej Zary
Quoted 10 lines · On Monday 02 August 2010 23:04:28 H. Peter Anvin wrote:
Quoted 8 lines · Ondrej Zary

Hello,
matroxfb (at least with Mystique and Mystique 220) stopped working in 2.6.34 - the screen is completely corrupted. Bisection shows that 6175ddf06b6172046a329e3abfd9c901a43efd2e is first bad commit.

Reverting 6175ddf06b6172046a329e3abfd9c901a43efd2e in 2.6.34 fixes the problem (1c5b9069e12e20d2fe883076ae0bf73966492108 must be reverted first).

Quoted 5 lines

Sounds like another driver which used memcpy_toio() when it should have used iowrite32_rep() or __iowrite32_copy().

Hmm... is __iowrite32_copy() and iowrite32_rep() redundant? If so, we should get rid of the former.

There is a wrapper in drivers/video/matrox/matroxfb_base.h (see below) with some comment. So this commit changed one of the three points?

static inline void mga_memcpy_toio(vaddr_t va, const void* src, int len) { #if defined(__alpha__) || defined(__i386__) || defined(__x86_64__) /*
* memcpy_toio works for us if:
* (1) Copies data as 32bit quantities, not byte after byte,
* (2) Performs LE ordered stores, and
* (3) It copes with unaligned source (destination is guaranteed
to be page * aligned and length is guaranteed to be multiple of 4).
*/

Quoted 2 lines
        memcpy_toio(va.vaddr, src, len);

#else

Quoted 1 line

Yes, point (1) is not guaranteed by memcpy_toio().

Now I wonder how many drivers will this break...

The patch below fixes it for me. Is it correct on all architectures?

drivers/video/matrox/matroxfb_base.h +1 -1
@@ -157,7 +157,7 @@ static inline void mga_memcpy_toio(vaddr
 	 *  (3) It copes with unaligned source (destination is guaranteed to be page
 	 *      aligned and length is guaranteed to be multiple of 4).
 	 */
-	memcpy_toio(va.vaddr, src, len);
+	iowrite32_rep(va.vaddr, src, len);
 #else
         u_int32_t __iomem* addr = va.vaddr;
 
Signature · 1 line
Ondrej Zary
#5 lore raw
Quoted 6 lines · Ondrej Zary

The patch below fixes it for me. Is it correct on all architectures?

drivers/video/matrox/matroxfb_base.h static inline void mga_memcpy_toio(vaddr · line 163

Quoted 8 lines
drivers/video/matrox/matroxfb_base.h +1 -1
@@ -157,7 +157,7 @@ static inline void mga_memcpy_toio(vaddr
 	 *  (3) It copes with unaligned source (destination is guaranteed to be page
 	 *      aligned and length is guaranteed to be multiple of 4).
 	 */
-	memcpy_toio(va.vaddr, src, len);
+	iowrite32_rep(va.vaddr, src, len);
 #else
         u_int32_t __iomem* addr = va.vaddr;
 

I don't think so; in particular I don't *think* non-x86 architectures will deal with the requirement that it handles an unaligned source. As such, the #if would still be necessary; the #else clause could be replaced with a get_unaligned() ... iowrite32() loop.

The other thing to watch out for is that "len" passed to iowrite32_rep() is a count of 32-bit words, whereas memcpy_toio() takes a byte count... you need to >> 2 there.

-hpa

#6 lore raw
Quoted 19 lines · On Monday 02 August 2010 23:42:29 H. Peter Anvin wrote:

Review

Quoted 13 lines · Ondrej Zary

The patch below fixes it for me. Is it correct on all architectures?

--- linux-2.6.35-rc2/drivers/video/matrox/matroxfb_base.h 2010-06-06
05:43:24.000000000 +0200 +++
linux-2.6.35-rc3/drivers/video/matrox/matroxfb_base.h 2010-08-02
23:31:34.000000000 +0200 @@ -157,7 +157,7 @@ static inline void mga_memcpy_toio(vaddr
* (3) It copes with unaligned source (destination is guaranteed to be
page * aligned and length is guaranteed to be multiple of 4). */

(quoted diff) +1 -1
-	memcpy_toio(va.vaddr, src, len);
+	iowrite32_rep(va.vaddr, src, len);
 #else
         u_int32_t __iomem* addr = va.vaddr;

I don't think so; in particular I don't *think* non-x86 architectures will deal with the requirement that it handles an unaligned source. As such, the #if would still be necessary; the #else clause could be replaced with a get_unaligned() ... iowrite32() loop.

Just tested the #else part on sparc and it seems to work fine - so I'm not going to touch (break) it.

Quoted 3 lines

The other thing to watch out for is that "len" passed to iowrite32_rep() is a count of 32-bit words, whereas memcpy_toio() takes a byte count... you need to >> 2 there.

Thanks.

Here's v2 patch:

Fix incorrect use of memcpy_toio() in matroxfb that broke in 2.6.34.

drivers/video/matrox/matroxfb_base.h +2 -2
@@ -151,13 +151,13 @@ static inline void mga_writel(vaddr_t va
 static inline void mga_memcpy_toio(vaddr_t va, const void* src, int len) {
 #if defined(__alpha__) || defined(__i386__) || defined(__x86_64__)
 	/*
-	 * memcpy_toio works for us if:
+	 * iowrite32_rep works for us if:
 	 *  (1) Copies data as 32bit quantities, not byte after byte,
 	 *  (2) Performs LE ordered stores, and
 	 *  (3) It copes with unaligned source (destination is guaranteed to be page
 	 *      aligned and length is guaranteed to be multiple of 4).
 	 */
-	memcpy_toio(va.vaddr, src, len);
+	iowrite32_rep(va.vaddr, src, len >> 2);
 #else
         u_int32_t __iomem* addr = va.vaddr;
 
Signature · 1 line
Ondrej Zary