GPU (Linux): improves PCI GPU name resolution

CHANGELOG.md

@@ -1,5 +1,8 @@
 # Unreleased
 
+Changes:
+* Improved PCI GPU name resolution by checking udev's hardware database and the NVIDIA driver's procfs before falling back to pci.ids. (GPU, Linux)
+
 Bugfixes:
 * Fixed libc version detection on FreeBSD without binutils installed, and on DragonFly BSD. (Libc, FreeBSD / DragonFly)
 

src/detection/gpu/gpu_linux.c

@@ -1,12 +1,10 @@
 #include "detection/gpu/gpu.h"
 #include "common/debug.h"
 #include "common/io.h"
-#include "common/FFstrbuf.h"
+#include "common/properties.h"
 #include "common/strutil.h"
-#include "modules/gpu/option.h"
 
 #include <inttypes.h>
-#include <stdint.h>
 
 static bool detectDriverFromSysfs(FFstrbuf* result, FFstrbuf* pciDir, FFstrbuf* buffer, [[maybe_unused]] const char* drmKey) {
     uint32_t pciDirLength = pciDir->length;
@@ -337,6 +335,64 @@ static inline uint16_t pcieWidthToLanes(const FFstrbuf* buffer) {
     return FF_GPU_PCIE_SPEED_UNSET;
 }
 
+static bool parseModelName(const char* data, const char* key, bool stripVendorPrefix, FFstrbuf* name) {
+    ffStrbufClear(name);
+    if (!ffParsePropLines(data, key, name) || !name->length) {
+        return false;
+    }
+
+    ffStrbufDecodeHexEscapeSequences(name); // udev hex-escapes characters that are not allowed in property values
+    ffStrbufTrimSpace(name);
+
+    // hwdb names have the form `Codename [Marketing Name]`. Reduce them exactly like
+    // ffGPUParsePciIdsFile() does, so displaying the name does not change with the source.
+    if (name->length > 1 && name->chars[name->length - 1] == ']') {
+        const char* openingBracket = memrchr(name->chars, '[', name->length - 1);
+        if (openingBracket) {
+            ffStrbufSubstrAfter(name, (uint32_t) (openingBracket - name->chars));
+            ffStrbufSubstrBefore(name, name->length - 1);
+        }
+    }
+
+    // The driver repeats the vendor in its own model line. Dropping it keeps the pci.ids
+    // shape and lets ffGPUDetectTypeByVendorAndName() see the product family prefix.
+    if (stripVendorPrefix && ffStrbufStartsWithIgnCaseS(name, "NVIDIA ")) {
+        ffStrbufSubstrAfter(name, (uint32_t) strlen("NVIDIA ") - 1); // excludes the index itself
+    }
+
+    ffStrbufTrimSpace(name);
+    return name->length > 0;
+}
+
+static bool nameFromFile(const FFstrbuf* path, const char* key, bool stripVendorPrefix, FFstrbuf* buffer, FFstrbuf* name) {
+    if (!ffReadFileBuffer(path->chars, buffer) || !parseModelName(buffer->chars, key, stripVendorPrefix, name)) {
+        return false;
+    }
+    FF_DEBUG("GPU name resolved from %s: %s", path->chars, name->chars);
+    return true;
+}
+
+// The driver and udev already resolved the model name at probe / boot time. Only `Model:` is
+// consumed: the sibling `power` file goes through the RM and can wake a GPU suspended in runtime D3.
+static bool resolveNameFromFiles(const char* pciPath, bool isNvidia, FFstrbuf* buffer, FFstrbuf* name) {
+    FF_STRBUF_AUTO_DESTROY path = ffStrbufCreateA(128);
+
+    if (isNvidia) {
+        ffStrbufSetF(&path, "/proc/driver/nvidia/gpus/%s/information", pciPath);
+        if (nameFromFile(&path, "Model:", true, buffer, name)) {
+            return true;
+        }
+    }
+
+    ffStrbufSetF(&path, "/run/udev/data/+pci:%s", pciPath);
+    if (nameFromFile(&path, "E:ID_MODEL_FROM_DATABASE=", false, buffer, name)) {
+        return true;
+    }
+
+    ffStrbufClear(name);
+    return false;
+}
+
 static const char* detectPci(const FFGPUOptions* options, FFlist* gpus, FFstrbuf* buffer, FFstrbuf* deviceDir, const char* drmKey) {
     const uint32_t drmDirPathLength = deviceDir->length;
     uint32_t vendorId, deviceId, subVendorId, subDeviceId;
@@ -465,7 +521,12 @@ static const char* detectPci(const FFGPUOptions* options, FFlist* gpus, FFstrbuf
     }
 
     if (gpu->name.length == 0) {
-        ffGPUFillVendorAndName(subclassId, (uint16_t) vendorId, (uint16_t) deviceId, gpu);
+        // Stop at the first hit; unknown vendors are left to pci.ids, which also fills in `vendor`.
+        bool resolved = gpu->vendor.length > 0 &&
+            resolveNameFromFiles(pPciPath, gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA, buffer, &gpu->name);
+        if (!resolved) {
+            ffGPUFillVendorAndName(subclassId, (uint16_t) vendorId, (uint16_t) deviceId, gpu);
+        }
     }
 
     ffGPUDetectTypeByVendorAndName(gpu);