uvm_pmm_gpu_device_p2p_init() doesn't correctly refcount p2pdma pages

[tim-day-387]

NVIDIA Open GPU Kernel Modules Version

590.48.01

Please confirm this issue does not happen with the proprietary driver (of the same version). This issue tracker is only for bugs specific to the open kernel driver.

• I confirm that this does not happen with the proprietary driver package.

Operating System and Version

Amazon Linux 2023

Kernel Release

I've tested multiple custom built > 6.16 kernels

Please confirm you are running a stable release kernel (e.g. not a -rc). We do not accept bug reports for unreleased kernels.

• I am running on a stable kernel release.

Hardware: GPU

NVIDIA H100 80GB HBM3

Describe the bug

After:

commit b7e2823787735ca009e63f35f164b46df0ef096c
Author: Alistair Popple <[apopple@nvidia.com](mailto:apopple@nvidia.com)>
Date: Fri Feb 28 14:31:05 2025 +1100

mm/mm_init: move p2pdma page refcount initialisation to p2pdma

p2pdma pages are not being refcounted correctly. This causes CUDA to incorrectly conclude that p2pdma is not supported. Applying a change like:

diff --git a/kernel-open/nvidia-uvm/uvm_pmm_gpu.c b/kernel-open/nvidia-uvm/uvm_pmm_gpu.c
index 97ff13dc..9585ad0d 100644
--- a/kernel-open/nvidia-uvm/uvm_pmm_gpu.c
+++ b/kernel-open/nvidia-uvm/uvm_pmm_gpu.c
@@ -3352,8 +3352,10 @@ void uvm_pmm_gpu_device_p2p_init(uvm_parent_gpu_t *parent_gpu)
     // allocate PCI P2PDMA pages directly
     p2p_page = pfn_to_page(pci_start_pfn);
     page_pgmap(p2p_page)->ops = &uvm_device_p2p_pgmap_ops;
-    for (; page_to_pfn(p2p_page) < pci_end_pfn; p2p_page++)
+    for (; page_to_pfn(p2p_page) < pci_end_pfn; p2p_page++) {
         p2p_page->zone_device_data = NULL;
+        set_page_count(p2p_page, 1);
+    }
 
     parent_gpu->device_p2p_initialised = true;
 }

appears to fix the issue.

To Reproduce

CUFILE_USE_PCIP2PDMA=1 /usr/local/cuda/gds/tools/gdscheck -p will fail to return p2pdma as supported.

Bug Incidence

Always

nvidia-bug-report.log.gz

nvidia-bug-report.log.gz

More Info

No response

[apopple-nvidia]

Thanks @tim-day-387 for the report, we are aware of this and that a change is required to UVM to support newer kernels. Your proposed fix will probably appear to fix the issue on a freshly booted system, however it will fail to reset refcounts after p2pdma pages have been used and freed. I expect an official driver release with a fix applied to be released soon, but in the meantime please feel free to use the below fix for development purposes only, noting that we can't support patched drivers.

diff --git a/kernel-open/nvidia-uvm/uvm_va_range_device_p2p.c b/kernel-open/nvidia-uvm/uvm_va_range_device_p2p.c
index f326a95..20d1711 100644
--- a/kernel-open/nvidia-uvm/uvm_va_range_device_p2p.c
+++ b/kernel-open/nvidia-uvm/uvm_va_range_device_p2p.c
@@ -343,6 +343,12 @@ static NV_STATUS alloc_device_p2p_mem(uvm_gpu_t *gpu,
     for (i = 0; i < p2p_mem_page_count(p2p_mem); i++) {
         struct page *page = p2p_mem_get_page(p2p_mem, i);
 
+        // Older kernel versions set page_ref_count(page) to 1, but newer
+        // kernels with commit b7e282378773 ("mm/mm_init: move p2pdma page
+        // refcount initialisation to p2pdma") set the ref count to 0. The
+        // ref count is checked in pci_p2pdma_page_free(). For compatbility
+        // between the two just set the page count to 1.
+        set_page_count(page, 1);
         if (!gpu->parent->cdmm_enabled && !pci_p2pdma_page_free(page)) {
             UVM_ASSERT(0);
 
@@ -356,12 +362,10 @@ static NV_STATUS alloc_device_p2p_mem(uvm_gpu_t *gpu,
         page_set_zone_device_data(page, p2p_mem);
 
 #if UVM_CDMM_PAGES_SUPPORTED()
-        // RM doesn't use DEVICE_COHERENT pages and therefore won't already hold
-        // a reference to them, so take one now if using DEVICE_COHERENT pages.
-        if (gpu->parent->cdmm_enabled) {
-            get_page(page);
+        // The page reference count was taken unconditionally above by setting
+        // page_count to 1. Take a page devmap reference.
+        if (gpu->parent->cdmm_enabled)
             NV_GET_DEV_PAGEMAP(page_to_pfn(page));
-        }
 #else
         // CDMM P2PDMA will never be enabled for this case
         if (gpu->parent->cdmm_enabled) {
@@ -391,10 +395,10 @@ out_free:
     return status;
 }
 
-static NV_STATUS alloc_pci_device_p2p(uvm_gpu_t *gpu,
-                                      NvHandle client,
-                                      NvHandle memory,
-                                      uvm_device_p2p_mem_t **p2p_mem_out)
+static NV_STATUS alloc_zone_device_p2p(uvm_gpu_t *gpu,
+                                       NvHandle client,
+                                       NvHandle memory,
+                                       uvm_device_p2p_mem_t **p2p_mem_out)
 {
     NvU64 pfn;
     NV_STATUS status;
@@ -548,7 +552,7 @@ NV_STATUS uvm_api_alloc_device_p2p(UVM_ALLOC_DEVICE_P2P_PARAMS *params, struct f
             goto out_release;
     }
     else {
-        status = alloc_pci_device_p2p(gpu, params->hClient, params->hMemory, &p2p_mem);
+        status = alloc_zone_device_p2p(gpu, params->hClient, params->hMemory, &p2p_mem);
         if (status != NV_OK)
             goto out_release;
     }