Skip to contentSkip to conversation

Is it possible to implement interrupt time printk's reliably?

DISCUSSION Jon Smirl 3 participants 10 messages May 7, 2004 – May 11, 2004

lore
Share & tools
lore

Jon Smirl

author
#0 lore raw

Problem:
1) Some operations on graphics cards cannot be stopped once they are started. It's not reasonable to turn interrupts off around these operations.
2) Kernel developers want console printk's to work from interrupt routines.

How do you fix this situation?

1) Grpahics driver has started non-restartable operation. For example transferring a bitmap. This is not an automatic DMA operation, CPU involvement is needed.
2) Interrupt happens
3) Printk happens from interrupt.

Now we're stuck. The graphics chip is in a non-interruptible state and printk wants to use it.

We need some mechanism to get back to the driver code and finish the non-restartable operation before the printk can be allowed to proceed.

Another solution also comes to mind. Mark the appropriate sections in the video driver with BEGIN/END_INT_PRINTK. Then add a kernel build option to convert these macros to en/disable interrupts if interrupt time printk's are allowed. Would it be acceptable to disable interrupts for signifcant time on a development kernel where the developer is printk'ing from interrupts?

=====
Jon Smirl
[email protected]

__________________________________
Do you Yahoo!?
Win a $20,000 Career Makeover at Yahoo! HotJobs
http://hotjobs.sweepstakes.yahoo.com/careermakeover

#1 lore raw
Quoted 7 lines · Jon Smirl

Problem:
1) Some operations on graphics cards cannot be stopped once they are started. It's not reasonable to turn interrupts off around these operations.
2) Kernel developers want console printk's to work from interrupt routines.

How do you fix this situation?

Really you should use spin_lock_irqsave() on some driver-private lock around the operation. Why is it not reasonable to disable irq's? Duration, presumably?

If you're in process context you can use acquire_console_sem(), which will serialise against printk.

author
#2 lore raw
Quoted 5 lines · --- Andrew Morton
Quoted 3 lines · Jon Smirl

Problem:
1) Some operations on graphics cards cannot be stopped once they are

started.

Quoted 8 lines
Quoted 4 lines

It's not reasonable to turn interrupts off around these operations.
2) Kernel developers want console printk's to work from interrupt routines.

How do you fix this situation?

Really you should use spin_lock_irqsave() on some driver-private lock around the operation. Why is it not reasonable to disable irq's? Duration, presumably?

The operations take a while and would ruin latency. You might be copying 8MB of data.

Quoted 3 lines

If you're in process context you can use acquire_console_sem(), which will serialise against printk.

Won't I deadlock if I have acquire_console_sem(), take an interrupt, and then a printk is issued from the interrupt handelr?

=====
Jon Smirl
[email protected]

__________________________________
Do you Yahoo!?
Win a $20,000 Career Makeover at Yahoo! HotJobs
http://hotjobs.sweepstakes.yahoo.com/careermakeover

#3 lore raw
Quoted 8 lines · Jon Smirl
Quoted 4 lines

If you're in process context you can use acquire_console_sem(), which will serialise against printk.

Won't I deadlock if I have acquire_console_sem(), take an interrupt, and then a printk is issued from the interrupt handelr?

Nope. If printk finds the semaphore to be held it queues up the characters and returns without printing them. The console_sem-holding process will print the newly buffered characters before releasing the semaphore.

author
#4 lore raw
Quoted 6 lines · --- Andrew Morton
Quoted 5 lines · Jon Smirl
Quoted 4 lines

If you're in process context you can use acquire_console_sem(), which will serialise against printk.

Quoted 8 lines
Quoted 1 line

Won't I deadlock if I have acquire_console_sem(), take an interrupt, and

then a

Quoted 2 lines

printk is issued from the interrupt handelr?

Nope. If printk finds the semaphore to be held it queues up the characters and returns without printing them. The console_sem-holding process will print the newly buffered characters before releasing the semaphore.

Is this solution sufficient for kernel developers wanting to use printk from interrupt handlers? I've gotten negative feedback from Linus when I suggested queuing them before.

=====
Jon Smirl
[email protected]

__________________________________
Do you Yahoo!?
Win a $20,000 Career Makeover at Yahoo! HotJobs
http://hotjobs.sweepstakes.yahoo.com/careermakeover

#5 lore raw
Quoted 14 lines · Jon Smirl
Quoted 11 lines · --- Andrew Morton
Quoted 6 lines · Jon Smirl
Quoted 4 lines

If you're in process context you can use acquire_console_sem(), which will serialise against printk.

Won't I deadlock if I have acquire_console_sem(), take an interrupt, and

then a

Quoted 2 lines

printk is issued from the interrupt handelr?

Quoted 8 lines
Quoted 3 lines

Nope. If printk finds the semaphore to be held it queues up the characters and returns without printing them. The console_sem-holding process will print the newly buffered characters before releasing the semaphore.

Is this solution sufficient for kernel developers wanting to use printk from interrupt handlers? I've gotten negative feedback from Linus when I suggested queuing them before.

It has been this way for several years.

author
#6 lore raw

So how do printk's work in the very early boot? Is the video card active before the kernel probes it's module, or are these very early printk's being queued until the video driver is probed?

Quoted 21 lines · --- Andrew Morton
Quoted 6 lines · Jon Smirl
Quoted 3 lines · --- Andrew Morton
Quoted 2 lines · Jon Smirl
Quoted 2 lines

If you're in process context you can use acquire_console_sem(), which

will

Quoted 9 lines
Quoted 9 lines
Quoted 4 lines
Quoted 2 lines

serialise against printk.

Won't I deadlock if I have acquire_console_sem(), take an interrupt, and

then a

Quoted 2 lines

printk is issued from the interrupt handelr?

Nope. If printk finds the semaphore to be held it queues up the

characters

Quoted 3 lines
Quoted 2 lines

and returns without printing them. The console_sem-holding process will print the newly buffered characters before releasing the semaphore.

Quoted 7 lines
Quoted 2 lines

Is this solution sufficient for kernel developers wanting to use printk from interrupt handlers? I've gotten negative feedback from Linus when I

suggested

Quoted 2 lines

queuing them before.

It has been this way for several years.

=====
Jon Smirl
[email protected]

__________________________________
Do you Yahoo!?
Win a $20,000 Career Makeover at Yahoo! HotJobs
http://hotjobs.sweepstakes.yahoo.com/careermakeover

#7 lore raw
Quoted 4 lines

So how do printk's work in the very early boot? Is the video card active before the kernel probes it's module, or are these very early printk's being queued until the video driver is probed?

printk messages are stored in log_buf in printk.c. The console driver just reads the buffer and displays what is in the buffer. Look at printk.c carefully.

author
#8 lore raw

So how long is the delay between PCI probe time (when the framebuffer goes active) and when early user space is up with initrd? Or is initrd up first? If initrd is up first then early user space mode setting will occur at the same time that it does currently.

Quoted 12 lines · --- James Simmons
Quoted 2 lines

So how do printk's work in the very early boot? Is the video card active

before

Quoted 2 lines

the kernel probes it's module, or are these very early printk's being queued until the video driver is probed?

printk messages are stored in log_buf in printk.c. The console driver just reads the buffer and displays what is in the buffer. Look at printk.c carefully.

=====
Jon Smirl
[email protected]

__________________________________
Do you Yahoo!?
Win a $20,000 Career Makeover at Yahoo! HotJobs
http://hotjobs.sweepstakes.yahoo.com/careermakeover

#9 lore raw
Quoted 5 lines

So how long is the delay between PCI probe time (when the framebuffer goes active) and when early user space is up with initrd? Or is initrd up first? If initrd is up first then early user space mode setting will occur at the same time that it does currently.

Why do we even need a fbdev layer then!!! We might as well remove it completly!!!!

Its totally destroy's my dream of a real multidesktop OS :-( This is why I'm so upset. Using userland libraries to do multidesktop OS is a really limited retarded way of doing it. I worked on this for 3 years. Testing different systems with different configurations. Now I realize it will die a painful death.