On Wed, 17 Feb 2016 11:09:49 +0000 "Tian, Kevin" kevin.tian@intel.com wrote:
From: Alex Williamson Sent: Wednesday, February 17, 2016 5:39 AM
QEMU provides two fw_cfg files to support IGD. The first holds the OpRegion data which holds the Video BIOS Table (VBT). This needs to be copied into reserved memory and the address stored in the ASL Storage register of the device at 0xFC offset in PCI config space. The OpRegion is generally 8KB. This file is named "etc/igd-opregion".
The second file tells us the required size of the stolen memory space for the device. This is a dummy file, it has no backing so we only allocate the space without copying anything into it. This space requires 1MB alignment and is generally either 1MB or 2MB, depending on the hardware config. If the user has opted in QEMU to expose additional stolen memory beyond the GTT (GGMS), the GMS may add an additional 32MB to 512MB. The base address of the reserved memory allocated for this is written back to the Base Data of Stolen Memory register (BDSM) at PCI config offset 0x5C on the device. This file is named "etc/igd-bdsm".
What would happen if guest tries to access this range while there is no actual memory behind? Isn't it more clear to hide stolen memory at all instead of reporting a dummy range?
It's a fw_cfg file, it's not exposed to the guest, the purpose is to convey the size of stolen memory, which the BIOS then allocates as reserved memory and writes back to the BDSM register. It would be more clean to ignore stolen memory, but in cases where we need the vBIOS, such as laptops, where my LCD panel won't turn on without it, we don't have that luxury. The vBIOS programs the device to use stolen memory, at least 1MB, I assume for GTT purposes, and makes use of that for VESA modes. So, we need the vBIOS to support laptop panels, the vBIOS needs stolen memory for GTT space, therefore we need to provide some stolen memory.
This support is enabled in QEMU by using a vBIOS, I assume vGPUs won't expose a ROM BAR and therefore won't enable this. Additionally for BDW/SKL+ (gen8+) GPUs, if the user specifies rombar=0 then all of this is disabled, including the hostbridge and ISA bridge manipulation in order to support UPT mode. Thanks,
Alex
From: Alex Williamson [mailto:alex.williamson@redhat.com] Sent: Wednesday, February 17, 2016 9:50 PM
On Wed, 17 Feb 2016 11:09:49 +0000 "Tian, Kevin" kevin.tian@intel.com wrote:
From: Alex Williamson Sent: Wednesday, February 17, 2016 5:39 AM
QEMU provides two fw_cfg files to support IGD. The first holds the OpRegion data which holds the Video BIOS Table (VBT). This needs to be copied into reserved memory and the address stored in the ASL Storage register of the device at 0xFC offset in PCI config space. The OpRegion is generally 8KB. This file is named "etc/igd-opregion".
The second file tells us the required size of the stolen memory space for the device. This is a dummy file, it has no backing so we only allocate the space without copying anything into it. This space requires 1MB alignment and is generally either 1MB or 2MB, depending on the hardware config. If the user has opted in QEMU to expose additional stolen memory beyond the GTT (GGMS), the GMS may add an additional 32MB to 512MB. The base address of the reserved memory allocated for this is written back to the Base Data of Stolen Memory register (BDSM) at PCI config offset 0x5C on the device. This file is named "etc/igd-bdsm".
What would happen if guest tries to access this range while there is no actual memory behind? Isn't it more clear to hide stolen memory at all instead of reporting a dummy range?
It's a fw_cfg file, it's not exposed to the guest, the purpose is to convey the size of stolen memory, which the BIOS then allocates as reserved memory and writes back to the BDSM register. It would be more clean to ignore stolen memory, but in cases where we need the vBIOS, such as laptops, where my LCD panel won't turn on without it, we don't have that luxury. The vBIOS programs the device to use stolen memory, at least 1MB, I assume for GTT purposes, and makes use of that for VESA modes. So, we need the vBIOS to support laptop panels, the vBIOS needs stolen memory for GTT space, therefore we need to provide some stolen memory.
Thanks for explaining the rationale. It's a bit surprise to me but anyway since you do observe the effect. Allen might help confirm whether your assumption makes sense. :-)
This support is enabled in QEMU by using a vBIOS, I assume vGPUs won't expose a ROM BAR and therefore won't enable this. Additionally for BDW/SKL+ (gen8+) GPUs, if the user specifies rombar=0 then all of this is disabled, including the hostbridge and ISA bridge manipulation in order to support UPT mode. Thanks,
Correct. vGPU doesn't use a host vBIOS. We just leverage existing Seabios for early booting.
Thanks Kevin
-
Alex Williamson -
Tian, Kevin