Re: Kernel/driver hacking: panic: Assertion vm_object_busied((m->object)) failed at /usr/src/sys/vm/vm_page.c:5455

From: Neel Chauhan <nc_at_FreeBSD.org>
Date: Sun, 20 Jun 2021 21:32:37 -0700
Hi Hans,

Sorry if I have to bother you again.

On 2021-06-20 01:36, Hans Petter Selasky wrote:
> sg_dma_address() is zero, because the memory hasn't been loaded.

Makes sense!

> You need to handle two cases there:
> 
> When r->iobase is -1 and when it is not.
> 
> I suspect you should add r->iobase to the sg_dma_address() only and
> only when it is non -1.

I have addressed it in my code.

Hans, while your approach is correct (that's also done in Linux), for 
some reason it still crashes on vm_page_insert(). I must be doing "pa = 
" wrong.

Stacktrace: https://misc.neelc.org/drm-kmod/stacktrace2.txt

I have a separate function for getting the physical address: 
https://github.com/neelchauhan/drm-kmod/blob/5.7-wip/drivers/gpu/drm/i915/intel_freebsd.c#L219

This crash happens when "iobase == -1", so we execute Lines 228-231:
-
	} else {
		struct sgt_iter sgt = __sgt_iter(sgl, 0);
		pa = (sgt.pfn + (sgt.curr >> PAGE_SHIFT)) << PAGE_SHIFT;
	}

sgt.curr is zero, since (I believe) it's the first iteration.

I have copied the Linux approach and ported what I felt was necessary, I 
am worried I am doing this wrong.

> Also, there is a superfluous "pa = " in the beginning of the function.

(Hopefully) I have cleaned it up.

> --HPS

-Neel (nc_at_)
Received on Mon Jun 21 2021 - 04:32:37 UTC

Original text of this message