Intel AGP support attaching to wrong PCI IDs
I have an i875 chipset with these two devices:
8086:2578 - 00:00.0 Host bridge: Intel Corp. 82875P/E7210 Memory Controller Hub (rev 02)
8086:2579 - 00:01.0 PCI bridge: Intel Corp. 82875P Processor to AGP Controller (rev 02)
In the legacy io space thread we are talking about making a device driver for host bridges. The Intel AGP drivers (in my case agpgart-intel-mch) are attaching to the PCI IDs of the host bus device instead of the AGP bridge. This blocks us from making a host bridge driver.PCI_DEVICE_ID_INTEL_82875_HB 0x2578
Shouldn't they be attaching to device 0x2579? It looks like all of the drivers have this problem and are attaching to the host bus PCI IDs instead of the AGP bridge ID.
[root@jonsmirl pci]# cat devices
0000 80862578 0 ec000008 00000000 00000000 00000000 00000000 00000000 00000000 04000000 00000000 00000000 00000000 00000000 00000000 00000000 agpgart-intel-mch 0008 80862579 0 00000000 00000000 00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000 00000000
00000000 00000000 00000000
Am I correct in thinking that in all cases there has to be two PCI IDs, one for the host bridge and one for the APG bridge?
Signature · 2 lines
Jon Smirl [email protected]
Quoted 14 lines · Jon Smirl
I have an i875 chipset with these two devices:
8086:2578 - 00:00.0 Host bridge: Intel Corp. 82875P/E7210 Memory Controller Hub (rev 02)
8086:2579 - 00:01.0 PCI bridge: Intel Corp. 82875P Processor to AGP Controller (rev 02)In the legacy io space thread we are talking about making a device driver for host bridges. The Intel AGP drivers (in my case agpgart-intel-mch) are attaching to the PCI IDs of the host bus device instead of the AGP bridge. This blocks us from making a host bridge driver.
PCI_DEVICE_ID_INTEL_82875_HB0x2578
Quoted 3 lines
Shouldn't they be attaching to device
0x2579? It looks like all of the drivers have this problem and are attaching to the host bus PCI IDs instead of the AGP bridge ID.
Take a peek at 'lspci -vv' output. You'll notice that the AGP capabilities are attached to the host bridge.
Dave
Quoted 2 lines · Dave Jones
Take a peek at 'lspci -vv' output. You'll notice that the AGP capabilities are attached to the host bridge.
I see that now, why is it on the host bridge instead of the AGP bridge? So that means if we add drivers for the host bridges we have to add the code to the AGP drivers. It also implies that we have to load them.
Signature · 2 lines
Jon Smirl [email protected]
Quoted 6 lines · Jon Smirl
Quoted 2 lines · Dave Jones
Take a peek at 'lspci -vv' output. You'll notice that the AGP capabilities are attached to the host bridge.
I see that now, why is it on the host bridge instead of the AGP bridge?
Not sure. Maybe its partly due to the host bridge having all the smarts to deal with the memory controller.
Quoted 3 lines
So that means if we add drivers for the host bridges we have to add the code to the AGP drivers. It also implies that we have to load them.
Why exactly are you trying to write host bridge drivers anyway ? Confused.
If there's a sensible reason for such drivers, we could at some stage have the bridge drivers check for AGP capabilities, and if found, start up the initialisation of the relevant AGP chipset driver. (and then rip out the whole PCI detection stuff in agpgart). This is quite a lot of work though, so unless there's a really compelling reason, I don't think its worth doing.
Another way forward (somewhat hacky in one sense, but a lot cleaner in another) would be to change the PCI code so that it'll load and init multiple drivers that claim to support the same PCI ID.
This may cause issues for some other drivers however where
we have an old and a new driver with ID overlap.
So,.. what are you up to?
Dave
Quoted 2 lines · Dave Jones
Why exactly are you trying to write host bridge drivers anyway ? Confused.
We're trying to add sysfs attributes for identifying and controlling legacy address spaces. Desktop x86 has just a single legacy IO/mem address space but more advanced systems like the IA64 support multiple ones. See Documentation/filesystems/sysfs-pci.txt. All architectures need to provide a consistent API to make it easier to write the user space app (like video reset). If IA64 is the example then x86 needs to add legacy_io/mem attributes to the host bus bridge which just passes the accesses on without change.
Another part of this is VGA control. When there are multiple VGA devices the bridges have to be set to route VGA appropriately. This is a different feature than multiple legacy address spaces on the IA64.
Quoted 5 lines
Another way forward (somewhat hacky in one sense, but a lot cleaner in another) would be to change the PCI code so that it'll load and init multiple drivers that claim to support the same PCI ID.
This may cause issues for some other drivers however where
we have an old and a new driver with ID overlap.
This problem already exits in DRM/fbdev. DRM loads like a normal driver and binds to the PCI IDs. But if it loads and finds fbdev already bound to the PCI IDs it goes into stealth mode and runs without binding to the ID.
I would prefer that we stick with the one driver per ID rule and instead sort out DRM/fbdev to coordinate more. The legacy_io/mem support can probably be added some place in common AGP code since the attributes need to be created on all x86 boxes.
Quoted 1 line
What are you up to?
Putting together a foundation for X on GL. That involves reworking video support in Linux so that X can remove all the horrible code that plays with bridge chips and PCI config space from a user space app. Right now the foundation is not there to allow X to remove the code.
Several things are needed:
1) Secondary card reset at boot
2) Mode setting on secondary heads - fbdev does not have this
3) A way to set modes without being root
4) Memory management coordination between fbdev/DRM when multiple scan buffers are active
5) Mouse cursor support in fbdev
6) Fix video reset when resuming from suspend
Something that isn't required but would be nice it to fix things so that an OOPs will be visible even if X is running.
Once DRM/fbdev is fixed mesa-solo will run on it with full features (it already runs on fbdev/DRM with limited features). This will bring up a standalone OpenGL stack.
X on GL is already written and is part of the xserver project. This will run on the standalone OpenGL stack. Combine this with Cairo/Glitz and we have a graphics system that can compete with Windows Longhorn.
Signature · 2 lines
Jon Smirl [email protected]
Quoted 3 lines · Jon Smirl
X on GL is already written and is part of the xserver project. This will run on the standalone OpenGL stack. Combine this with Cairo/Glitz and we have a graphics system that can compete with Windows Longhorn.
Why compete with vaporware (mostly)? If you really want to see a complete modern graphics system, check out OS X. See the below screenshots for one example.
http://tjhsst.edu/~kmoffett/2k4-1.png
http://tjhsst.edu/~kmoffett/2k4-2.png
The rendered window from 2k4 has been distorted by the windowing system post-processing using a mesh transform. It runs just as quickly this way as it does full screen normally, at least as far as I can tell while playing. :-D A truely well-performing graphics system should be able to handle multiple applications processing to generate a single output stream without effort. I also ran some informal and simple end-user-style tests a while back where I rapidly switched between and moved around the following windows (each distorted in the same way):
1) A UT2k4 window playing a demo
2) A couple translucent terminal windows continuously scrolling
text
3) A DVD player window playing The Matrix
4) A couple Quicktime player windows with movies running
The best part was: ~60 FPS on everything, despite the distortions, translucency, rapid movement, DVD playing, etc. I hope that when the new linux graphics and Soft-RT stuff is done it will be able to achieve similar performance. If my coding skills were a little more up to snuff, I would try to help out, but as it is, I fear I'd just muddy the waters :-\.
Cheers,
Kyle Moffett
-----BEGIN GEEK CODE BLOCK-----
Version: 3.12
GCM/CS/IT/U d- s++: a18 C++++>$ UB/L/X/*++++(+)>$ P+++(++++)>$ L++++(+++) E W++(+) N+++(++) o? K? w--- O? M++ V? PS+() PE+(-) Y+ PGP+++ t+(+++) 5 X R? tv-(--) b++++(++) DI+ D+ G e->++++$ h!*()>++$ r !y?(-) ------END GEEK CODE BLOCK------
Quoted 3 lines · Dave Jones
Another way forward (somewhat hacky in one sense, but a lot cleaner in another) would be to change the PCI code so that it'll load and init multiple drivers that claim to support the same PCI ID.
That might actually be useful to support some weird hardware. E.g. I
have a single-function PCI device which reports as a graphics adapter (either VGA or other, depending on the configuration), but besides the actual graphics adapter the chip includes a frame grabber, an USB host controller, an Ethernet interface, an IEEE 1284 parallel port, an UART, an I2C controller and a PS/2 keybard/mouse controller. ;-) All of these mapped with a single PCI BAR; the chip can also be configured for legacy mapping of some devices, like the PS/2 stuff.
Maciej
Quoted 5 lines · Dave Jones
Another way forward (somewhat hacky in one sense, but a lot cleaner in another) would be to change the PCI code so that it'll load and init multiple drivers that claim to support the same PCI ID.
This may cause issues for some other drivers however where
we have an old and a new driver with ID overlap.
This sounds like it would allow the use of the parallel ATA interfaces on SATA controllers, which would make a lot of people happy.
Signature · 3 lines
======================================================================== Ian Pilcher [email protected] ========================================================================