tpearson at raptorengineeringinc.com wrote:
I have patched src/arch/i386/smp/mpspec.c to write a correct, multi-core MP table under amdfam10.
I think this is very desirable and a great functionality improvement!
Thanks!
But this code is not nice at all.
You're right, it is quite complex for what it does. It was hacked together over a period of about 30 minutes just to get my board working. Also, I wasn't sure if there would be any interest in such a feature when I submitted the patch.
Could you shift it around so that it uses continue aggressively, and has shorter variable names? Sure thing.
It looks like that could reduce indentation two or three levels, and then the code might actually be visible in my terminal... Heh. I have the same problem in my Webmail client!
Is this romcc code? If not, maybe it could even be recursive.. Not sure, but I know it is very sensitive. Running printk here reliably crashed coreboot, so the stack may be limited; I didn't want to risk recursion causing some problem later on for anyone else.
//Peter
I probably won't be able to get to this for a couple weeks, as the board I was working with has already entered production. I should have an identical board in a week or two that I can experiment with. Timothy Pearson Raptor Engineering
Is this romcc code? If not, maybe it could even be recursive.. Not sure, but I know it is very sensitive. Running printk here reliably crashed coreboot, so the stack may be limited;
Have you tried increasing the stack size? It's too bad not to be able to use printk. Thanks, Myles
On Fri, February 12, 2010 10:52 am, Myles Watson wrote:
Is this romcc code? If not, maybe it could even be recursive.. Not sure, but I know it is very sensitive. Running printk here reliably crashed coreboot, so the stack may be limited;
Have you tried increasing the stack size? It's too bad not to be able to use printk.
Thanks, Myles
Yes I did. Actually, at one point I quadrupled it--coreboot crashed in a different spot, but still crashed. Removing the printk statement from that function fixed the crash, even with the quadrupled stack still in place. Timothy Pearson Raptor Engineering
Here is a cleaned up and tested version of the SMP APIC autodetect patch. Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com> --- Let me know what you think! Timothy Pearson Raptor Engineering
Here is a cleaned up and tested version of the SMP APIC autodetect patch. Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com> --- It would of course be helpful to attach the patch. My Webmail client keeps eating it... Timothy Pearson Raptor Engineering
On 2/16/10 5:11 AM, Timothy Pearson wrote:
Here is a cleaned up and tested version of the SMP APIC autodetect patch.
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
---
It would of course be helpful to attach the patch. My Webmail client keeps eating it...
Timothy Pearson Raptor Engineering
Do you have a comparison mptable as generated by the old and the new version on your system? And, can you describe high level, what the patch changes? It looks to me as if you are recursing through the tree instead of just walking the "all_devices" list. So this implies that you don't catch all devices when running through all_devices. This sounds like a problem in the resource allocator and maybe it should be fixed there instead? The code as it is runs in roughly O(n^3) and on top of that calls itself which is a little bit scary. But, looking at it a bit closer I see that the outer loop with p_it does nothing. I dropped the outer loop and moved the lapic stuff to the inner loop where it is used, except passing it around on the stack. It will probably mean 3 more cpuid calls and 3 more lapic reads as opposed to heavy stack usage, but I think the stack usage has been more troublesome in the past.. Please let me know what you think... Stefan
On Tue, Feb 16, 2010 at 12:02 PM, Stefan Reinauer <stepan@coresystems.de>wrote:
On 2/16/10 5:11 AM, Timothy Pearson wrote:
Here is a cleaned up and tested version of the SMP APIC autodetect patch.
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
And, can you describe high level, what the patch changes? It looks to me as if you are recursing through the tree instead of just walking the "all_devices" list. So this implies that you don't catch all devices when running through all_devices. This sounds like a problem in the resource allocator and maybe it should be fixed there instead?
I don't understand why this would be a resource allocator problem. Aren't we talking about the device tree? Maybe the real problem is related to the memory corruption seen with printk? Thanks, Myles
On 2/16/10 8:42 PM, Myles Watson wrote:
>
>
> On Tue, Feb 16, 2010 at 12:02 PM, Stefan Reinauer
> <stepan@coresystems.de <mailto:stepan@coresystems.de>> wrote:
>
> On 2/16/10 5:11 AM, Timothy Pearson wrote:
> > Here is a cleaned up and tested version of the SMP APIC
> autodetect patch.
> >
> > Signed-off-by: Timothy Pearson
> <tpearson@raptorengineeringinc.com
> <mailto:tpearson@raptorengineeringinc.com>>
>
>
>
> And, can you describe high level, what the patch changes? It looks
> to me
> as if you are recursing through the tree instead of just walking the
> "all_devices" list.
> So this implies that you don't catch all devices when running through
> all_devices. This sounds like a problem in the resource allocator and
> maybe it should be fixed there instead?
>
> I don't understand why this would be a resource allocator problem.
> Aren't we talking about the device tree?
Oh, sorry I didn't say more... What the old code did was the following:
device_t cpu;
for(cpu = all_devices; cpu; cpu = cpu->next) {
....
<- check if the device is an APIC and if it is, dump an APIC entry
in the mptable ->
....
}
But it seems that the code does not see all devices -- anymore -- not
sure if it ever did, but I think to remember that's what happened at
some point, or was at least intended.
So now the code looks l
scan_for_apic_devices(parent)
{
for(c_it=0; c_it < parent->links; c_it++) {
for (child = parent->link[c_it].children; child; child =
child->sibling) {
....
if (child->path.type == DEVICE_PATH_APIC_CLUSTER) {
// Found an APIC cluster, scan it for APICs
smp_scan_for_apic_devices(child);
}
}
}
}
So two more steps are necessary:
- check all the downwards links of a device instead of just walking
devices and checking their type.
- run recursively in a special case on APIC clusters.
This sounds a whole lot like something changed in the way "all_devices"
works. And if "all_devices" does not mean "all devices" I am sure there
are more places in our code that need similar fixes.
> Maybe the real problem is related to the memory corruption seen with
> printk?
Not completely impossible, but I figured it's easier to ask the guy who
rewrote the resource allocator if he knows something about how the
intended behavior of "all_devices" ;-)
- Maybe the behavior is intended, then we just need to check in
Timothy's patch.
- Maybe the behavior is not intended, but that's how the code works
right now. Then we'd rather have to look at the resource allocator and
decide if we want "all_devices" to mean "all devices", or whether we
rename the variable, or redefine its meaning.
- Maybe none of the above applies, then we need to do nasty stack
corruption debugging. In this case it would be fundamentally wrong to
touch either the device tree code or the mp table creation code until we
fix the corruption in order to make sure we don't create funny special
cases.
So, if you have hints enlightening any of the maybes, please share!
Stefan
On Tue, Feb 16, 2010 at 4:40 PM, Stefan Reinauer <stepan@coresystems.de>wrote:
On 2/16/10 8:42 PM, Myles Watson wrote:
On Tue, Feb 16, 2010 at 12:02 PM, Stefan Reinauer <stepan@coresystems.de>wrote:
On 2/16/10 5:11 AM, Timothy Pearson wrote:
Here is a cleaned up and tested version of the SMP APIC autodetect patch.
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
And, can you describe high level, what the patch changes? It looks to me as if you are recursing through the tree instead of just walking the "all_devices" list. So this implies that you don't catch all devices when running through all_devices. This sounds like a problem in the resource allocator and maybe it should be fixed there instead?
I don't understand why this would be a resource allocator problem. Aren't we talking about the device tree?
Oh, sorry I didn't say more... What the old code did was the following:
device_t cpu; for(cpu = all_devices; cpu; cpu = cpu->next) { .... <- check if the device is an APIC and if it is, dump an APIC entry in the mptable -> .... }
But it seems that the code does not see all devices -- anymore -- not sure if it ever did, but I think to remember that's what happened at some point, or was at least intended.
I'm really surprised that it didn't see all devices.
So now the code looks l
scan_for_apic_devices(parent) { for(c_it=0; c_it < parent->links; c_it++) { for (child = parent->link[c_it].children; child; child = child->sibling) { ....
if (child->path.type == DEVICE_PATH_APIC_CLUSTER) { // Found an APIC cluster, scan it for APICs smp_scan_for_apic_devices(child); } } } }
So two more steps are necessary: - check all the downwards links of a device instead of just walking devices and checking their type. - run recursively in a special case on APIC clusters.
This sounds a whole lot like something changed in the way "all_devices" works. And if "all_devices" does not mean "all devices" I am sure there are more places in our code that need similar fixes.
I agree. I've never seen the problem where I couldn't run through all the devs with all_devices.
Maybe the real problem is related to the memory corruption seen with printk?
Not completely impossible, but I figured it's easier to ask the guy who rewrote the resource allocator if he knows something about how the intended behavior of "all_devices" ;-)
As far as I know I didn't change that behavior. Does anyone have a boot log where they can see that the tree of devices has more devices than the list (at the same stage of enumeration)?
- Maybe the behavior is intended, then we just need to check in Timothy's patch. - Maybe the behavior is not intended, but that's how the code works right now. Then we'd rather have to look at the resource allocator and decide if we want "all_devices" to mean "all devices", or whether we rename the variable, or redefine its meaning.
I think it should mean all devices. I changed resource allocation to go through the tree because children of disabled devices shouldn't have allocated resources. - Maybe none of the above applies, then we need to do nasty stack corruption
debugging. In this case it would be fundamentally wrong to touch either the device tree code or the mp table creation code until we fix the corruption in order to make sure we don't create funny special cases.
This is my vote, but I'm happy to be proven wrong. I don't see anywhere in the code where devices get added to the tree but not to the list of devices. I also don't remember a place where devices get removed from the all_devices list.
So, if you have hints enlightening any of the maybes, please share!
My suggestion would be to traverse the tree and the list and compare them. The problem is that printk was causing a hang for him. Thanks, Myles
This patch allows the user to select a maximum HyperTransport link frequency. Speed capping is required for maximum stability when attempting to use new HT3 CPUs on older HT1 mainboards, as the physical inter-CPU traces on the mainboard may not be able to reliably handle the faster HT3 link frequency. --- Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com> Timothy Pearson Raptor Engineering
Timothy Pearson wrote:
This patch allows the user to select a maximum HyperTransport link frequency.
I like this a lot.
+++ src/northbridge/amd/amdht/h3finit.c (working copy) @@ -1327,7 +1327,25 @@
for (i = 0; i < pDat->TotalLinks*2; i += 2) { - cbPCBFreqLimit = 0xFFFF; +#if CONFIG_LIMIT_HT_SPEED_200 + cbPCBFreqLimit = 0x0001; // 200MHz
But shouldn't this also be tied in with Kconfig files, with the correct maximum selected for each board? (An expert mode could allow overclocking though.) Oh, and the comments seem redundant to me, especially if Kconfig also gets involved in this. Maybe it already is, even? //Peter
So two more steps are necessary: - check all the downwards links of a device instead of just walking devices and checking their type. - run recursively in a special case on APIC clusters.
This sounds a whole lot like something changed in the way "all_devices" works. And if "all_devices" does not mean "all devices" I am sure there are more places in our code that need similar fixes.
This is the crux of the issue. all_devices does NOT mean "all devices", it means "all devices attached to the root node, which is all_devices". As the root node, usually the PCI bus and the APICs are visible. On my board, the APICs are all under an APIC cluster, so they are not immediately visible from the root node. Incidentally, there is code already in Coreboot (to generate the PCI device lists) that takes all_devices and simply probes the PCI downward link. It (sanely, IMHO) does not expect to see PCI devices on the all_devices root node. I don't have access to a board that I can generate before/after tables with at this time. This behavior is very simple to see though; if you turn on printk spew and look closely at the printed detected device tables on any amdfam10 board, you will see the root node only connects to the APIC cluster and the PCI root bus(es). If you have any more questions feel free to ask! Hopefully the root cause of the problem can be located and solved. Timothy Pearson Raptor Engineering
On Wed, Feb 24, 2010 at 12:26 AM, Timothy Pearson < tpearson@raptorengineeringinc.com> wrote:
So two more steps are necessary: - check all the downwards links of a device instead of just walking devices and checking their type. - run recursively in a special case on APIC clusters.
This sounds a whole lot like something changed in the way "all_devices" works. And if "all_devices" does not mean "all devices" I am sure there are more places in our code that need similar fixes.
This is the crux of the issue. all_devices does NOT mean "all devices", it means "all devices attached to the root node, which is all_devices".
You're correct; this is the crux. Here's a snippet of the fam10 boot log for SimNOW: Show all devs...Before Device Enumeration. Root Device: enabled 1, 0 resources APIC_CLUSTER: 0: enabled 1, 0 resources APIC: 00: enabled 1, 0 resources PCI_DOMAIN: 0000: enabled 1, 0 resources PCI: 00:18.0: enabled 1, 0 resources PCI: 00:00.0: enabled 1, 0 resources PCI: 00:00.1: enabled 1, 0 resources PCI: 00:01.0: enabled 1, 0 resources PCI: 00:01.1: enabled 1, 0 resources PCI: 00:00.0: enabled 1, 0 resources PCI: 00:00.0: enabled 1, 0 resources PCI: 00:00.1: enabled 1, 0 resources PCI: 00:00.2: enabled 0, 0 resources PCI: 00:01.0: enabled 0, 0 resources PCI: 00:01.0: enabled 1, 0 resources PNP: 002e.0: enabled 0, 3 resources PNP: 002e.1: enabled 0, 2 resources PNP: 002e.2: enabled 1, 2 resources PNP: 002e.3: enabled 0, 2 resources PNP: 002e.5: enabled 1, 4 resources PNP: 002e.6: enabled 0, 1 resources PNP: 002e.7: enabled 0, 3 resources PNP: 002e.8: enabled 0, 0 resources PNP: 002e.9: enabled 0, 0 resources PNP: 002e.a: enabled 0, 0 resources PNP: 002e.b: enabled 1, 2 resources PCI: 00:01.1: enabled 1, 0 resources PCI: 00:01.2: enabled 1, 0 resources PCI: 00:01.3: enabled 1, 0 resources I2C: 00:18: enabled 1, 0 resources I2C: 00:50: enabled 1, 0 resources I2C: 00:51: enabled 1, 0 resources I2C: 00:52: enabled 1, 0 resources I2C: 00:53: enabled 1, 0 resources I2C: 00:50: enabled 1, 0 resources I2C: 00:51: enabled 1, 0 resources I2C: 00:52: enabled 1, 0 resources I2C: 00:53: enabled 1, 0 resources PCI: 00:01.5: enabled 0, 0 resources PCI: 00:01.6: enabled 0, 0 resources PCI: 00:18.1: enabled 1, 0 resources PCI: 00:18.2: enabled 1, 0 resources PCI: 00:18.3: enabled 1, 0 resources PCI: 00:18.4: enabled 1, 0 resources Compare with tree... Root Device: enabled 1, 0 resources APIC_CLUSTER: 0: enabled 1, 0 resources APIC: 00: enabled 1, 0 resources PCI_DOMAIN: 0000: enabled 1, 0 resources PCI: 00:18.0: enabled 1, 0 resources PCI: 00:00.0: enabled 1, 0 resources PCI: 00:00.1: enabled 1, 0 resources PCI: 00:01.0: enabled 1, 0 resources PCI: 00:01.1: enabled 1, 0 resources PCI: 00:00.0: enabled 1, 0 resources PCI: 00:00.0: enabled 1, 0 resources PCI: 00:00.1: enabled 1, 0 resources PCI: 00:00.2: enabled 0, 0 resources PCI: 00:01.0: enabled 0, 0 resources PCI: 00:01.0: enabled 1, 0 resources PNP: 002e.0: enabled 0, 3 resources PNP: 002e.1: enabled 0, 2 resources PNP: 002e.2: enabled 1, 2 resources PNP: 002e.3: enabled 0, 2 resources PNP: 002e.5: enabled 1, 4 resources PNP: 002e.6: enabled 0, 1 resources PNP: 002e.7: enabled 0, 3 resources PNP: 002e.8: enabled 0, 0 resources PNP: 002e.9: enabled 0, 0 resources PNP: 002e.a: enabled 0, 0 resources PNP: 002e.b: enabled 1, 2 resources PCI: 00:01.1: enabled 1, 0 resources PCI: 00:01.2: enabled 1, 0 resources PCI: 00:01.3: enabled 1, 0 resources I2C: 00:18: enabled 1, 0 resources I2C: 00:50: enabled 1, 0 resources I2C: 00:51: enabled 1, 0 resources I2C: 00:52: enabled 1, 0 resources I2C: 00:53: enabled 1, 0 resources I2C: 00:50: enabled 1, 0 resources I2C: 00:51: enabled 1, 0 resources I2C: 00:52: enabled 1, 0 resources I2C: 00:53: enabled 1, 0 resources PCI: 00:01.5: enabled 0, 0 resources PCI: 00:01.6: enabled 0, 0 resources PCI: 00:18.1: enabled 1, 0 resources PCI: 00:18.2: enabled 1, 0 resources PCI: 00:18.3: enabled 1, 0 resources PCI: 00:18.4: enabled 1, 0 resources There are 44 devices in the tree, and 44 in the all_devices list. If there is some place that the list is getting broken, we should fix it. As the root node, usually the PCI bus and the APICs are visible. On my
board, the APICs are all under an APIC cluster, so they are not immediately visible from the root node.
They're not children of the root node, but they are accessible through the next pointer: for (dev = all_devices; dev; dev = dev->next) { do_printk(debug_level, "%s: enabled %d, %d resources\n", dev_path(dev), dev->enabled, dev->resources); }
I don't have access to a board that I can generate before/after tables with at this time. This behavior is very simple to see though; if you turn on printk spew and look closely at the printed detected device tables
What are printed detected device tables? If you have any more questions feel free to ask! Hopefully the root cause
of the problem can be located and solved.
Sure. Let's make sure it gets taken care of.
Thanks, Myles
participants (4)
-
Myles Watson -
Peter Stuge -
Stefan Reinauer -
Timothy Pearson