Ah, that sweet power savings! :) ----C1E% C10% CPU%c7 PkgTmp GFX%rc6 GFXMHz GFXAMHz CPUGFX% Pkg%pc8 Pk%pc10 CPU%LPI SYS%LPI PkgWatt CorWatt GFXWatt 0.65 93.77 36.12 49 0.00 1033 1050 2.93 7.53 21.66 21.70 0.00 3.28 1.34 0.02 0.09 99.23 41.68 48 0.00 1033 1050 0.91 11.98 10.22 10.24 0.00 1.95 0.22 0.01 0.09 99.17 41.41 49 0.00 1033 1050 0.25 29.83 39.18 39.26 11.38 1.10 0.34 0.00 0.02 99.78 42.40 48 0.00 1033 1050 0.20 31.63 48.31 48.40 42.58 0.55 0.11 0.00 0.06 99.35 41.67 46 0.00 1033 1050 0.21 23.86 51.95 52.04 46.83 0.80 0.30 0.00 C1E% C10% CPU%c7 PkgTmp GFX%rc6 GFXMHz GFXAMHz CPUGFX% Pkg%pc8 Pk%pc10 CPU%LPI SYS%LPI PkgWatt CorWatt GFXWatt 0.04 99.71 42.38 46 0.00 1033 1050 0.20 24.74 58.05 58.15 51.85 0.44 0.11 0.00 0.07 99.41 41.77 46 0.00 1033 1050 0.17 25.20 52.57 52.65 46.99 0.73 0.27 0.00 0.03 99.76 42.39 45 0.00 1033 1050 0.18 26.67 56.59 56.68 50.16 0.46 0.11 0.00 0.06 99.24 41.54 45 0.00 1033 1050 0.24 23.89 46.83 46.90 41.40 0.91 0.32 0.00 0.03 99.75 42.42 45 0.00 1033 1050 0.20 22.65 59.64 59.73 52.87 0.46 0.10 0.00
---- -K On 12/13/24 14:26, Kenneth Crudup wrote:
OK, it looks like the effective change (that's not already contained in the LTR SNOOP patches already in Linus' master (et al.)) comes from this line from the Ubuntu commit 1a0102a0 ("UBUNTU: SAUCE: PCI/ASPM: Enable ASPM for links under VMD domain"):---- diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c index 00143f5fb83a..d2ff44e7fbb1 100644 --- a/drivers/pci/pcie/aspm.c +++ b/drivers/pci/pcie/aspm.c@@ -688,7 +688,8 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)aspm_l1ss_init(link); /* Save default state */ - link->aspm_default = link->aspm_enabled;+ link->aspm_default = parent->dev_flags & PCI_DEV_FLAGS_ENABLE_ASPM ?+ ASPM_STATE_ALL : link->aspm_enabled; ----Where "PCI_DEV_FLAGS_ENABLE_ASPM" (a new flag) is set if the parent bridge is a VMD (determined via Ubuntu commit 154d48da2 ("UBUNTU: SAUCE: vmd: fixup bridge ASPM by driver name instead").So, as we've both noticed, most of the three patches I'd pointed out aren't required as the LTR SNOOP values are either fixed up, or previously set in the BIOS.So I've folded the new commits into the attached patch (against Linus' master), which now gives me full power-saving.-K On 12/13/24 12:27, Kenneth Crudup wrote:Turns out that my current VMD is already listed as "VMD_FEATS_CLIENT" after all:---- {PCI_VDEVICE(INTEL, 0x467f), .driver_data = VMD_FEATS_CLIENT,}, ---- 0000:00:0e.0 0104: 8086:467f Subsystem: 1028:0af3 Flags: bus master, fast devsel, latency 0, IOMMU group 9 Memory at 603c000000 (64-bit, non-prefetchable) [size=32M] Memory at 72000000 (32-bit, non-prefetchable) [size=32M] Memory at 6040100000 (64-bit, non-prefetchable) [size=1M] Capabilities: [80] MSI-X: Enable+ Count=19 Masked-Capabilities: [90] Express Root Complex Integrated Endpoint, MSI 00Capabilities: [e0] Power Management version 3 Kernel driver in use: vmd ----I'm going to compare the commits you'd pointed out in master to what the working (Ubuntu) patches do and figure out where the disconnect is.-K On 12/13/24 11:48, Kenneth Crudup wrote:I was really hopeful this would have handled it, but no joy.I also tried dropping "PCI/ASPM: Enable LTR for endpoints behind VMD" and "PCI/ASPM: Enable ASPM for links under VMD domain" each separately on top of the below quirk patch to no avail.The only thing that works is the aggregate patch I've added. > However, IDs 0x9a09, 0xa0b0, and 0xa0bc are NOT tagged with > VMD_FEATS_CLIENT. In fact, they're not included in vmd_ids[] at all, > so I'd be surprised if VMD worked for those devices unless BIOS set > up the VMD itself.Yeah, my BIOS does- not sure if you'd missed it, but I'd rewritten "Enable LTR for endpoints behind VMD" to print if the BIOS already does that for you, and sent the patch here (since I wasn't seeing the message printed when the fixup was being done and wanted to know why).I'd REALLY like to get this into mainline, so if there's anything I can do to help, LMK.Thanks, -Kenny On 12/13/24 08:43, Bjorn Helgaas wrote:[+cc David, Nirmal, linux-pci] On Thu, Dec 12, 2024 at 03:04:53PM -0800, Kenneth Crudup wrote:On 12/12/24 12:56, Bjorn Helgaas wrote:On Wed, Dec 11, 2024 at 03:26:37PM -0800, Kenneth Crudup wrote:So we're on 6.13-rc2 and the patches are getting closer and closer, but they still need to be manually added. The good news is now only (variants of) "PCI/ASPM: Enable LTR for endpoints behind VMD" and "PCI/ASPM: Enable ASPM for links under VMD domain" are needed....https://git.launchpad.net/~ubuntu-kernel/ubuntu/+source/linux/+git/ lunar/commit/?id=47c7bfd31514e7b54a1f830f7707297aebbb8679This is "UBUNTU: SAUCE: PCI/ASPM: Enable LTR for endpoints behind VMD", which writes PCI_LTR_MAX_SNOOP_LAT and PCI_LTR_MAX_NOSNOOP_LAT for PCI_CLASS_STORAGE_EXPRESS devices directly below VMDs with Device IDs 0x9a09, 0xa0b0, or 0xa0bc. This looks equivalent in spirit to upstream https://git.kernel.org/linus/f492edb40b54 ("PCI: vmd: Add quirk to configure PCIe ASPM and LTR"), which writes PCI_LTR_MAX_SNOOP_LAT and PCI_LTR_MAX_NOSNOOP_LAT for any kind of device below VMDs tagged with VMD_FEATS_CLIENT, which includes 0x467f, 0x4c3d, 0x7d0b, 0x9a0b, 0xa77f, 0xad0b, 0xb06f, 0xb60b. However, IDs 0x9a09, 0xa0b0, and 0xa0bc are NOT tagged with VMD_FEATS_CLIENT. In fact, they're not included in vmd_ids[] at all, so I'd be surprised if VMD worked for those devices unless BIOS set up the VMD itself. Maybe David or Nirmal can comment on this?https://git.launchpad.net/~ubuntu-kernel/ubuntu/+source/linux/+git/ lunar/commit/?id=1a0102a08f206149d9abd56c2b28877c878b5526This is "UBUNTU: SAUCE: PCI/ASPM: Enable ASPM for links under VMD domain", which adds "link->aspm_default = ASPM_STATE_ALL" for device IDs 0x9a09 and 0xa0b0. This looks like it should also be handled by upstream f492edb40b54 ("PCI: vmd: Add quirk to configure PCIe ASPM and LTR") [1], which adds "pci_enable_link_state(pdev, PCIE_LINK_STATE_ALL)". But again, the Device IDs mentioned in the Ubuntu commit are NOT included in the upstream VMD_FEATS_CLIENT list.https://git.launchpad.net/~ubuntu-kernel/ubuntu/+source/linux/+git/ lunar/commit/?id=154d48da2c57514e4b5dadc7b8c70a4edb550981This is "UBUNTU: SAUCE: vmd: fixup bridge ASPM by driver name instead", which applies the quirk that writes PCI_LTR_MAX_SNOOP_LAT and PCI_LTR_MAX_NOSNOOP_LAT for PCI_CLASS_STORAGE_EXPRESS devices below any VMD claimed by the "vmd" driver, not just VMDs with Device IDs 0x9a09, 0xa0b0, or 0xa0bc. I think the only thing that's missing is that the upstream vmd_ids[] needs to be updated with some new VMD Device IDs that are tagged with VMD_FEATS_CLIENT. I don't know what the vmd_ids[] strategy is, but Kenneth, you might try an upstream patch like the one below. If that resolves the standby/low-power issues, maybe David or Nirmal can figure out the "right" way to do this. Bjorndiff --git a/drivers/pci/controller/vmd.c b/drivers/pci/controller/ vmd.cindex 9d9596947350..4de7ff3bbf23 100644 --- a/drivers/pci/controller/vmd.c +++ b/drivers/pci/controller/vmd.c @@ -1109,6 +1109,12 @@ static const struct pci_device_id vmd_ids[] = { .driver_data = VMD_FEATS_CLIENT,}, {PCI_VDEVICE(INTEL, 0xb06f), .driver_data = VMD_FEATS_CLIENT,}, + {PCI_VDEVICE(INTEL, 0x9a09), + .driver_data = VMD_FEATS_CLIENT,}, + {PCI_VDEVICE(INTEL, 0xa0b0), + .driver_data = VMD_FEATS_CLIENT,}, + {PCI_VDEVICE(INTEL, 0xa0bc), + .driver_data = VMD_FEATS_CLIENT,}, {0,} }; MODULE_DEVICE_TABLE(pci, vmd_ids);
--Kenneth R. Crudup / Sr. SW Engineer, Scott County Consulting, Orange County CA
