Linux 2.5.69
[ Linux-kernel added to the cc, since I got several queries about what the
crashes were.. ]
Quoted 4 lines · David van Hoose
Can I get some details regarding the AGP problem? I had some really bad random crashes, panics, and hardlocks up through 2.5.68, and I'm wondering if this is the same issue. I first noticed them around 2.5.63.
They actually started in 2.5.60 if it's the same bug.
And yes, you'd get random crashes, panics, lockups and even reboots. The problem was that the pmd/pgd's were put in the slab cache in between 2.5.59 and 2.5.60, and that was simply wrong because the AGP code changes the cacheability of the kernel pages when it maps stuff into the AGP aperture. That in turn will change the page tables but it won't update the cached entries in the pmd slab caches.
So what happens is that once you exit X, and the page tables are put back together without the cacheability changes, and you start a new program, that program may get a page table with partly bogus kernel page table entries.
That, in turn, when it happens will cause _major_ memory corruption, and your machine is toast, often in very interesting ways because the internal kernel data structures got corrupted. It can also cause random SIGSEGV's etc.
But it only happens with AGP, and a lot of people either don't use it or run only one X session.
Linus
Quoted 9 lines · Linus Torvalds
They actually started in 2.5.60 if it's the same bug.
And yes, you'd get random crashes, panics, lockups and even reboots. The problem was that the pmd/pgd's were put in the slab cache in between 2.5.59 and 2.5.60, and that was simply wrong because the AGP code changes the cacheability of the kernel pages when it maps stuff into the AGP aperture. That in turn will change the page tables but it won't update the cached entries in the pmd slab caches.
So what happens is that once you exit X, and the page tables are put back together without the cacheability changes, and you start a new program,
Quoted 8 lines
that program may get a page table with partly bogus kernel page table entries.
That, in turn, when it happens will cause _major_ memory corruption, and your machine is toast, often in very interesting ways because the internal kernel data structures got corrupted. It can also cause random SIGSEGV's etc.
But it only happens with AGP, and a lot of people either don't use it or run only one X session.
Any chance one of you could try out the fixed slabification patches?
Message-ID: <[email protected]>
-- wli
Quoted 17 lines · Linus Torvalds
[ Linux-kernel added to the cc, since I got several queries about what the
crashes were.. ]Quoted 1 line · David van Hoose
Can I get some details regarding the AGP problem? I had some really bad random crashes, panics, and hardlocks up through 2.5.68, and I'm wondering if this is the same issue. I first noticed them around 2.5.63.
They actually started in 2.5.60 if it's the same bug.
And yes, you'd get random crashes, panics, lockups and even reboots. The problem was that the pmd/pgd's were put in the slab cache in between 2.5.59 and 2.5.60, and that was simply wrong because the AGP code changes the cacheability of the kernel pages when it maps stuff into the AGP aperture. That in turn will change the page tables but it won't update the cached entries in the pmd slab caches.
So what happens is that once you exit X, and the page tables are put back together without the cacheability changes, and you start a new program, that program may get a page table with partly bogus kernel page table entries.
Quoted 6 lines
That, in turn, when it happens will cause _major_ memory corruption, and your machine is toast, often in very interesting ways because the internal kernel data structures got corrupted. It can also cause random SIGSEGV's etc.
But it only happens with AGP, and a lot of people either don't use it or run only one X session.
Okay.. I kept having all of those problems. I guess the memory corruption could also explain the buffer overflow in the panic information I posted around the time of 2.5.65 then. I'll stick the beta kernel back on my system and play with it again. If I find anything else, I'll be sure to mention it. Don't want any nasty bugs like that hiding in 2.6.x.
Thanks!
David
Quoted 28 lines · Linus Torvalds
[ Linux-kernel added to the cc, since I got several queries about what the
crashes were.. ]Quoted 4 lines · David van Hoose
Can I get some details regarding the AGP problem? I had some really bad random crashes, panics, and hardlocks up through 2.5.68, and I'm wondering if this is the same issue. I first noticed them around 2.5.63.
They actually started in 2.5.60 if it's the same bug.
And yes, you'd get random crashes, panics, lockups and even reboots. The problem was that the pmd/pgd's were put in the slab cache in between 2.5.59 and 2.5.60, and that was simply wrong because the AGP code changes the cacheability of the kernel pages when it maps stuff into the AGP aperture. That in turn will change the page tables but it won't update the cached entries in the pmd slab caches.
So what happens is that once you exit X, and the page tables are put back together without the cacheability changes, and you start a new program, that program may get a page table with partly bogus kernel page table entries.
That, in turn, when it happens will cause _major_ memory corruption, and your machine is toast, often in very interesting ways because the internal kernel data structures got corrupted. It can also cause random SIGSEGV's etc.
Quoted 4 lines
But it only happens with AGP, and a lot of people either don't use it or run only one X session.
Linus
This fixes a problem I reported here:
http://marc.theaimsgroup.com/?l=linux-kernel&m=104990441203933&w=2
Thanks,
Steven
Quoted 2 lines · Linus Torvalds
But it only happens with AGP, and a lot of people either don't use it or run only one X session.
The integrated graphics seemed to be a key too. I test the agpgart changes I make on a half dozen boxes before asking you to pull them, and still couldn't reproduce this bug. The only difference was that I only have boxes without onboard graphics. All of them worked just fine, and all of them used agp & dri. Maybe I was hitting the bug, but the integrated chipsets thrash agpgart a little harder..
Dave
Quoted 2 lines · Dave Jones
The integrated graphics seemed to be a key too.
No, there were reports for radeon users etc, so it wasn't a requirement. I think the real requirement is that it uses a fair _amount_ of AGP.
Linus