Skip to contentSkip to conversation

Problems with dma_mmap_writecombine on mach-pxa

DISCUSSION Frank Buss 6 participants 16 messages Feb 17, 2005 – Feb 18, 2005 Arm (32-bit)

lore
Share & tools
lore

Frank Buss

author
#0 lore raw

I'm trying to use the pxafb driver on mach-pxa, but I can't mmap the framebuffer memory. I can access it from the driver, filling the entire screen, but when I access the pointer returned from mmap from a user space program, the following two things happens:

- the vm_pgoff is ignored and I get the start of the internal buffer, which caused writing to the palette and DMA descriptors

- when I change the location of the framebuffer to the start of the internal buffer, I can write to the screen, but only to the first 4 k; any write after this address is ignored, but no segfault is generated.

Any ideas what I can do to find the reason? I don't think that it is a kernel bug, but perhaps a wrong configuration for my platform.

Signature · 2 lines
Frank Buß, [email protected]
http://www.frank-buss.de, http://www.it4-systems.de
#1 lore raw
Quoted 8 lines · Frank Buss

I'm trying to use the pxafb driver on mach-pxa, but I can't mmap the framebuffer memory. I can access it from the driver, filling the entire screen, but when I access the pointer returned from mmap from a user space program, the following two things happens:

- the vm_pgoff is ignored and I get the start of the internal buffer, which caused writing to the palette and DMA descriptors

Quoted 6 lines

- when I change the location of the framebuffer to the start of the internal buffer, I can write to the screen, but only to the first 4 k; any write after this address is ignored, but no segfault is generated.

Any ideas what I can do to find the reason? I don't think that it is a kernel bug, but perhaps a wrong configuration for my platform.

Please try this (and revert your changes):

===== arch/arm/mm/consistent.c 1.27 vs edited =====

arch/arm/mm/consistent.c +5 -3
@@ -284,13 +284,15 @@
 	spin_unlock_irqrestore(&consistent_lock, flags);
 
 	if (c) {
+		unsigned long off = vma->vm_pgoff;
+
 		kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
 
-		if (vma->vm_pgoff < kern_size &&
-		    user_size <= (kern_size - vma->vm_pgoff)) {
+		if (off < kern_size &&
+		    user_size <= (kern_size - off)) {
 			vma->vm_flags |= VM_RESERVED;
 			ret = remap_pfn_range(vma, vma->vm_start,
-					      page_to_pfn(c->vm_pages),
+					      page_to_pfn(c->vm_pages) + off,
 					      user_size, vma->vm_page_prot);
 		}
 	}
Signature · 4 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:  2.6 PCMCIA      - http://pcmcia.arm.linux.org.uk/
                 2.6 Serial core
author
#2 lore raw
Quoted 1 line

Please try this (and revert your changes):

thanks, this could fix the bug with the vm_pgoff, but I don't think this will fix the problem with the ignored memory access after the first PAGE_SIZE from the mapped memory. I'll try it tommorow, when I'm again at my customers site, where I have access to the board.

Signature · 2 lines
Frank Buß, [email protected]
http://www.frank-buss.de, http://www.it4-systems.de
#3 lore raw
Quoted 6 lines · Frank Buss
Quoted 1 line

Please try this (and revert your changes):

thanks, this could fix the bug with the vm_pgoff, but I don't think this will fix the problem with the ignored memory access after the first PAGE_SIZE from the mapped memory. I'll try it tommorow, when I'm again at my customers site, where I have access to the board.

Since we map the whole lot in one go, if you get one page, there's no reason why you shouldn't get the lot. This is why I'm wondering if it has something to do with your other modifications.

Signature · 4 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:  2.6 PCMCIA      - http://pcmcia.arm.linux.org.uk/
                 2.6 Serial core
#4 lore raw

[TTY] 2 points seems strange to me.

In reply to Russell King

Hi,

Looking at TTY code, I noticed a weird test done in "opost_bock" located in n_tty.c file. I don't understand why the following test is done at the start of the function:
if (nr > sizeof(buf))

       nr = sizeof(buf);

Actually it limits the size of processing blocks to 4 bytes and I can't find a reason why.

Second point, a lot of serial drivers call in their interrupt handler "tty_flip_buffer_push" function. This function must no be called in interrupt context. Why is it done anyway ?

Thanks for your answers.

Franck

#5 lore raw

Re: [TTY] 2 points seems strange to me.

Quoted 7 lines · Franck Bui-Huu

Looking at TTY code, I noticed a weird test done in "opost_bock" located in n_tty.c file. I don't understand why the following test is done at the start of the function:
if (nr > sizeof(buf))

       nr = sizeof(buf);

Actually it limits the size of processing blocks to 4 bytes and I can't find a reason why.

No, it limits the size to 80 bytes,
which is the size of buf.

sizeof returns the size of the char array buf[80]
(standard C)

Quoted 3 lines

Second point, a lot of serial drivers call in their interrupt handler "tty_flip_buffer_push" function. This function must no be called in interrupt context. Why is it done anyway ?

Calling tty_flip_buffer_push() is fine from interrupt as long as tty->low_latency is not set. It just queues work for later.

Signature · 2 lines
Paul Fulghum
Microgate Systems, Ltd.
#6 lore raw

Re: [TTY] 2 points seems strange to me.

Quoted 11 lines · Paul Fulghum
Quoted 7 lines · Franck Bui-Huu

Looking at TTY code, I noticed a weird test done in "opost_bock" located in n_tty.c file. I don't understand why the following test is done at the start of the function:
if (nr > sizeof(buf))

       nr = sizeof(buf);

Actually it limits the size of processing blocks to 4 bytes and I can't find a reason why.

Quoted 5 lines

No, it limits the size to 80 bytes,
which is the size of buf.

sizeof returns the size of the char array buf[80]
(standard C)

Looking at the code, I think Franck is right. buf is a "const unsigned char *" for which sizeof(buf) is the size of a pointer.

This certainly looks like a bug.

Since the function doesn't guarantee that nr bytes are written, and the caller must handle the case of fewer bytes, this probably went unnoticed.

Signature · 4 lines
Paulo Marques - www.grupopie.com

All that is necessary for the triumph of evil is that good men do nothing.
Edmund Burke (1729 - 1797)
#7 lore raw

Re: [TTY] 2 points seems strange to me.

Quoted 8 lines · Paulo Marques
Quoted 5 lines · Paul Fulghum

No, it limits the size to 80 bytes,
which is the size of buf.

sizeof returns the size of the char array buf[80]
(standard C)

Looking at the code, I think Franck is right. buf is a "const unsigned char *" for which sizeof(buf) is the size of a pointer.

What kernel version are you looking at?
I'm looking at 2.4.20 n_tty.c opost_block() and
buf is a char array.

Signature · 2 lines
Paul Fulghum
Microgate Systems, Ltd.
#8 lore raw

Re: [TTY] 2 points seems strange to me.

Quoted 11 lines · Paul Fulghum
Quoted 9 lines · Paulo Marques
Quoted 5 lines · Paul Fulghum

No, it limits the size to 80 bytes,
which is the size of buf.

sizeof returns the size of the char array buf[80]
(standard C)

Quoted 6 lines
Quoted 1 line

Looking at the code, I think Franck is right. buf is a "const unsigned char *" for which sizeof(buf) is the size of a pointer.

What kernel version are you looking at?
I'm looking at 2.4.20 n_tty.c opost_block() and
buf is a char array.

Oops, this should have been clarified sooner...

I was looking at a 2.6.11-rc2 tree I have hanging around here. So maybe the "buf" type was changed and the condition remained, and that produced the bug.

Signature · 4 lines
Paulo Marques - www.grupopie.com

All that is necessary for the triumph of evil is that good men do nothing.
Edmund Burke (1729 - 1797)
#9 lore raw

Re: [TTY] 2 points seems strange to me.

In reply to Paul Fulghum
Quoted 5 lines · Paul Fulghum
Quoted 4 lines · Paulo Marques
Quoted 3 lines · Paul Fulghum

No, it limits the size to 80 bytes,
which is the size of buf.

Quoted 12 lines
Quoted 4 lines
Quoted 2 lines

sizeof returns the size of the char array buf[80]
(standard C)

Looking at the code, I think Franck is right. buf is a "const unsigned char *" for which sizeof(buf) is the size of a pointer.

What kernel version are you looking at?
I'm looking at 2.4.20 n_tty.c opost_block() and
buf is a char array.

Signature · 2 lines
Paul Fulghum
Microgate Systems, Ltd.

Ahaa! That's how the bug got introduced. It used to be an array and then it got changed to a pointer! linux-2.4.26 also shows a local array.

Cheers,
Dick Johnson
Penguin : Linux version 2.6.10 on an i686 machine (5537.79 BogoMips).
Notice : All mail here is now cached for review by Dictator Bush.
98.36% of all statistics are fiction.

#10 lore raw

Re: [TTY] 2 points seems strange to me.

Quoted 20 lines · linux-os
Quoted 18 lines · Paul Fulghum
Quoted 10 lines · Paulo Marques
Quoted 5 lines · Paul Fulghum

No, it limits the size to 80 bytes,
which is the size of buf.

sizeof returns the size of the char array buf[80]
(standard C)

Looking at the code, I think Franck is right. buf is a "const unsigned char *" for which sizeof(buf) is the size of a pointer.

What kernel version are you looking at?
I'm looking at 2.4.20 n_tty.c opost_block() and
buf is a char array.

Quoted 8 lines
Quoted 3 lines
Signature · 2 lines
Paul Fulghum
Microgate Systems, Ltd.

Ahaa! That's how the bug got introduced. It used to be an array and then it got changed to a pointer! linux-2.4.26 also shows a local array.

Yes, just looked at the revision history in linux.bkbits.net and Linus just fixed this 67 hours ago... So we're too late :)

Signature · 4 lines
Paulo Marques - www.grupopie.com

All that is necessary for the triumph of evil is that good men do nothing.
Edmund Burke (1729 - 1797)
#11 lore raw

Re: [TTY] 2 points seems strange to me.

Quoted 9 lines
Quoted 5 lines

Ahaa! That's how the bug got introduced. It used to be an array and then it got changed to a pointer! linux-2.4.26 also shows a local array.

Yes, just looked at the revision history in linux.bkbits.net and Linus just fixed this 67 hours ago... So we're too late :)

ok, maybe next time :)

Franck

#12 lore raw

Re: [TTY] 2 points seems strange to me.

In reply to Paulo Marques
Quoted 10 lines · Paulo Marques
Quoted 7 lines · Franck Bui-Huu

Looking at TTY code, I noticed a weird test done in "opost_bock" located in n_tty.c file. I don't understand why the following test is done at the start of the function:
if (nr > sizeof(buf))

       nr = sizeof(buf);

Actually it limits the size of processing blocks to 4 bytes and I can't find a reason why.

Quoted 6 lines

No, it limits the size to 80 bytes,
which is the size of buf.

sizeof returns the size of the char array buf[80]
(standard C)

Huh?? "buf" is a pointer. On this system pointers are 4-bytes in length. What "standard C" are you using? Frank found a bug.

Cheers,
Dick Johnson
Penguin : Linux version 2.6.10 on an i686 machine (5537.79 BogoMips).
Notice : All mail here is now cached for review by Dictator Bush.
98.36% of all statistics are fiction.

#13 lore raw

Re: [TTY] 2 points seems strange to me.

In reply to Paul Fulghum
Quoted 10 lines
Quoted 4 lines

Second point, a lot of serial drivers call in their interrupt handler "tty_flip_buffer_push" function. This function must no be called in interrupt context. Why is it done anyway ?

Calling tty_flip_buffer_push() is fine from interrupt as long as tty->low_latency is not set. It just queues work for later.

I was looking at driver for 8250 in 8250.c file and at the end of "receive_chars" interrupt handler, it calls "tty_flip_buffer_push" even if "tty->low_latency" is set since no such test is done before the call...
I was also wondering why not always calling "schedule_delayed_work" whatever the state of "tty->latency"?

Franck

#14 lore raw

Re: [TTY] 2 points seems strange to me.

Quoted 4 lines · Franck Bui-Huu

I was looking at driver for 8250 in 8250.c file and at the end of "receive_chars" interrupt handler, it calls "tty_flip_buffer_push" even if "tty->low_latency" is set since no such test is done before the call...

Yes this is a known problem. In the case of SMP kernel and low_latency, and 8250.c, this causes a dead lock. This has resurfaced again and again for many months from different sources.

There is disagreement on what code is
in error and what should be fixed. So it stays broken.

Quoted 2 lines

I was also wondering why not always calling "schedule_delayed_work" whatever the state of "tty->latency"?

If low_latency is set and you are not calling
from interrupt, then calling work directly speeds processing.

I submitted a patch that forced schedule_delayed_work if in interrupt context to avoid this problem. It was rejected.

Signature · 2 lines
Paul Fulghum
Microgate Systems, Ltd.
author
#15 lore raw
In reply to Russell King
Quoted 3 lines · Russell King

Since we map the whole lot in one go, if you get one page, there's no reason why you shouldn't get the lot. This is why I'm wondering if it has something to do with your other modifications.

your patch works, thanks, but only for the problem with the ignored offset, as expected. Now I can use the original pxafb driver, but with the same
problem: All writes from user space in the mmap'ed region after the first 4096 bytes are ignored.

Perhaps it is not a kernel bug, but a configuration problem with the platform, because it is a in-house developed platform. Now a colleague is working on this problem and reading a book about the Linux 2.6 kernel details over weekend and learning how to use the kernel debugger (until now we have used printk and flashing new zImages every time, which is very time consuming). I'll post the solution next week, if we'll found one.

Signature · 2 lines
Frank Buß, [email protected]
http://www.frank-buss.de, http://www.it4-systems.de