Nvidia stable 10.1 ankita bugfixes 1216 #9

ankita-nv · 2025-12-17T03:07:20Z

Address the following bugs:

Correct setting of numa distances
GPU memory VMA alignment adjustment for hugepfnmap

nvmochs · 2025-12-17T03:51:25Z

General question: What is the upstream plan for these?

NVIDIA: SAUCE: hw/vfio: adjust alignment for hugepfnmap

Can you add an example in the commit message where this was causing an issue?

NVIDIA: SAUCE: acpi: generic initiator in sorted order

Does this need a fixes tag?

Can you add an example in the commit message of the mismatch and the how it looks when fixed up?

On the iterator, in other parts of the QEMU source I see g_slist_next() being used to traverse ‘next’. I don’t care but upstream might.

ankita-nv · 2025-12-17T04:32:34Z

General question: What is the upstream plan for these?
I will start a formal internal review in the next few days but want to unblock users at this point. I see several folks doing verification.

NVIDIA: SAUCE: hw/vfio: adjust alignment for hugepfnmap

Can you add an example in the commit message where this was causing an issue?

Sure, will do.

NVIDIA: SAUCE: acpi: generic initiator in sorted order

Does this need a fixes tag?

Can you add an example in the commit message of the mismatch and the how it looks when fixed up?

Sure, will address them.

On the iterator, in other parts of the QEMU source I see g_slist_next() being used to traverse ‘next’. I don’t care but upstream might.

Thanks for the headsup. I'll fix this before posting for upstream.

During creation of the VM's SRAT table, the generic intiator entries are added. Currently, the code queries the object, which may not be in the sorted order. This results in the mismatch in the VMs view of the PXM and the numa node ids. As a fix, the patch builds a list of generic intiator objects, sorts them and then put it in the VM's SRAT table. Original (unsorted) PXM in the VM SRAT table [152h 0338 004h] Proximity Domain : 00000000 [17Ah 0378 004h] Proximity Domain : 00000001 [1A4h 0420 004h] Proximity Domain : 00000007 [1C4h 0452 004h] Proximity Domain : 00000006 [1E4h 0484 004h] Proximity Domain : 00000005 [204h 0516 004h] Proximity Domain : 00000004 [224h 0548 004h] Proximity Domain : 00000003 [244h 0580 004h] Proximity Domain : 00000009 [264h 0612 004h] Proximity Domain : 00000002 [284h 0644 004h] Proximity Domain : 00000008 [2A2h 0674 004h] Proximity Domain : 00000009 After the patch (sorted) [152h 0338 004h] Proximity Domain : 00000000 [17Ah 0378 004h] Proximity Domain : 00000001 [1A4h 0420 004h] Proximity Domain : 00000002 [1C4h 0452 004h] Proximity Domain : 00000003 [1E4h 0484 004h] Proximity Domain : 00000004 [204h 0516 004h] Proximity Domain : 00000005 [224h 0548 004h] Proximity Domain : 00000006 [244h 0580 004h] Proximity Domain : 00000007 [264h 0612 004h] Proximity Domain : 00000008 [284h 0644 004h] Proximity Domain : 00000009 Fixes: 0a5b5ac ("hw/acpi: Implement the SRAT GI affinity structure") Signed-off-by: Ankit Agrawal <[email protected]>

Qemu's determination of the VMA address for a region needs an update to handle regions that may be a BAR, but with the actual size of the mapping to not be at a power-of-2 alignment. This happens for the case of Grace based systems, where the device memory is exposed as a BAR. The mapping however is only of the size of the actual physical memory, which may not be a power-of-2 aligned. This affects hugepfnmap mappings on such regions. The current algorithm determines the VMA address alignment based on the mapping alignment. This needs change so as to be based on the next power-of-2 of the mapping size. This patch updates the algorithm to achieve the alignment. Original VMA mapping to the device memory of size 0x2F00F00000 on a GB200 ff88ff000000-ffb7fff00000 rw-s 400000000000 00:06 727 /dev/vfio/devices/vfio1 After the patch application (aligned at order 13 PMD) ff8ac0000000-ffb9c0f00000 rw-s 400000000000 00:06 727 /dev/vfio/devices/vfio1 Signed-off-by: Ankit Agrawal <[email protected]>

ankita-nv · 2025-12-17T20:45:45Z

Hi Matt, the branch is ready for review after addressing your suggestions. Thanks!

nvmochs

No further issues from me.

Acked-by: Matthew R. Ochs <[email protected]>

MitchellAugustin · 2025-12-17T22:42:19Z

I don't see links to any upstream discussions included in the commits. Could you add those so we can track the upstreaming progress?

ankita-nv · 2025-12-17T23:16:11Z

I don't see links to any upstream discussions included in the commits. Could you add those so we can track the upstreaming progress?

Hi Mitchell, it has not been posted upstream just yet.

MitchellAugustin

Functionally, all of these changes look good to me. Please add the upstream discussion links to these commit messages once they are available (preferably in advance of us initiating the next build.)

MitchellAugustin · 2025-12-17T23:53:14Z

hw/vfio/region.c


    for (i = 0; i < region->nr_mmaps; i++) {
-        size_t align = MIN(1ULL << ctz64(region->mmaps[i].size), 1 * GiB);
+        size_t align = MIN(pow2ceil(region->mmaps[i].size), 1 * GiB);


MitchellAugustin · 2025-12-17T23:58:01Z

hw/acpi/pci.c

 }

-static int build_acpi_generic_initiator(Object *obj, void *opaque)
+static gint memory_device_addr_sort(gconstpointer a, gconstpointer b)


MitchellAugustin · 2025-12-18T00:01:15Z

hw/vfio/region.c


    for (i = 0; i < region->nr_mmaps; i++) {
-        size_t align = MIN(1ULL << ctz64(region->mmaps[i].size), 1 * GiB);
+        size_t align = MIN(pow2ceil(region->mmaps[i].size), 1 * GiB);


MitchellAugustin · 2025-12-18T00:01:22Z

hw/acpi/pci.c

 }

-static int build_acpi_generic_initiator(Object *obj, void *opaque)
+static gint memory_device_addr_sort(gconstpointer a, gconstpointer b)


MitchellAugustin · 2025-12-18T00:13:01Z

hw/acpi/pci.c

+    return 0;
+}
+
+static int acpi_generic_initiator_list(Object *obj, void *opaque)


MitchellAugustin · 2025-12-18T00:24:29Z

hw/acpi/pci.c

    return 0;
 }

+static void build_all_acpi_generic_initiators(GArray *table_data)


MitchellAugustin · 2025-12-18T00:25:53Z

hw/acpi/pci.c

+    return list;
+}
+
+static int build_acpi_generic_initiator(AcpiGenericInitiator *gi,


MitchellAugustin · 2025-12-18T00:26:00Z

hw/acpi/pci.c


 void build_srat_generic_affinity_structures(GArray *table_data)
 {
-    object_child_foreach_recursive(object_get_root(),


ankita-nv · 2025-12-18T00:42:00Z

Thanks Matt and Mitchell for the review. Yes, I'll update with the links once I post the patches upstream.

shamiali2008

Is this going to be the upstream proposal as well? Based on the earlier discussion, there was a suggestion to map the entire BAR region and then unmap the non-mmap regions. If this is a temp fix, may be good to mention that in commit log.

ankita-nv added 2 commits December 17, 2025 20:40

ankita-nv force-pushed the nvidia_stable-10.1-ankita-bugfixes-1216 branch from f869f1b to c12786a Compare December 17, 2025 20:44

nvmochs self-requested a review December 17, 2025 21:30

nvmochs approved these changes Dec 17, 2025

View reviewed changes

MitchellAugustin approved these changes Dec 18, 2025

View reviewed changes

shamiali2008 reviewed Dec 18, 2025

View reviewed changes

Nvidia stable 10.1 ankita bugfixes 1216 #9

Are you sure you want to change the base?

Nvidia stable 10.1 ankita bugfixes 1216 #9

Uh oh!

Conversation

ankita-nv commented Dec 17, 2025

Uh oh!

nvmochs commented Dec 17, 2025

Uh oh!

ankita-nv commented Dec 17, 2025

Uh oh!

ankita-nv commented Dec 17, 2025

Uh oh!

nvmochs left a comment

Choose a reason for hiding this comment

Uh oh!

MitchellAugustin commented Dec 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ankita-nv commented Dec 17, 2025

Uh oh!

MitchellAugustin left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ankita-nv commented Dec 18, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

shamiali2008 left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

MitchellAugustin commented Dec 17, 2025 •

edited

Loading

ankita-nv commented Dec 18, 2025 •

edited

Loading