Collect GPU information (vendor id and device id) in Linux....

chrome/gpu/gpu_info_collector_linux.cc

 // Copyright (c) 2006-2010 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "chrome/gpu/gpu_info_collector.h"
 
+#include <dlfcn.h>
+#include <vector>
+
+#include "app/gfx/gl/gl_bindings.h"
+#include "app/gfx/gl/gl_context.h"
+#include "app/gfx/gl/gl_implementation.h"
+#include "base/logging.h"
+#include "base/scoped_ptr.h"
+#include "base/string_util.h"
+
+namespace {
+
+// PciDevice and PciAccess are defined to access libpci functions.  Their
+// members match the corresponding structures defined by libpci in size up to
+// fields we may access.  For those members we don't use, their names are
+// defined as "fieldX", etc., or, left out if they are declared after the
+// members we care about in libpci.
+
+struct PciDevice {
+  PciDevice* next;
+
+  uint16 field0;
+  uint8 field1;
+  uint8 field2;
+  uint8 field3;
+  int field4;
+
+  uint16 vendor_id;
+  uint16 device_id;
+  uint16 device_class;
+};
+
+struct PciAccess {
+  unsigned int field0;
+  int field1;
+  int field2;
+  char* field3;
+  int field4;
+  int field5;
+  unsigned int field6;
+  int field7;
+
+  void (*function0)();
+  void (*function1)();
+  void (*function2)();
+
+  PciDevice* device_list;
+};
+
+// Define function types.
+typedef PciAccess* (*FT_pci_alloc)();
+typedef void (*FT_pci_init)(PciAccess*);
+typedef void (*FT_pci_cleanup)(PciAccess*);
+typedef void (*FT_pci_scan_bus)(PciAccess*);
+typedef void (*FT_pci_scan_bus)(PciAccess*);
+typedef int (*FT_pci_fill_info)(PciDevice*, int);
+typedef char* (*FT_pci_lookup_name)(PciAccess*, char*, int, int, ...);
+
+// This includes dynamically linked library handle and functions pointers from
+// libpci.
+struct PciInterface {
+  void* lib_handle;
+
+  FT_pci_alloc pci_alloc;
+  FT_pci_init pci_init;
+  FT_pci_cleanup pci_cleanup;
+  FT_pci_scan_bus pci_scan_bus;
+  FT_pci_fill_info pci_fill_info;
+  FT_pci_lookup_name pci_lookup_name;
+};
+
+// This dynamically opens libpci and get function pointers we need.  Return
+// NULL if library fails to open or any functions can not be located.
+// Returned interface (if not NULL) should be deleted in FinalizeLibPci.
+PciInterface* InitializeLibPci(const char* lib_name) {
+  void* handle = dlopen(lib_name, RTLD_LAZY);
+  if (handle == NULL) {
+    LOG(ERROR) << "Fail to dlopen libpci";
+    return NULL;
+  }
+  PciInterface* interface = new struct PciInterface;
+  interface->lib_handle = handle;
+  interface->pci_alloc = reinterpret_cast<FT_pci_alloc>(
+      dlsym(handle, "pci_alloc"));
+  interface->pci_init = reinterpret_cast<FT_pci_init>(
+      dlsym(handle, "pci_init"));
+  interface->pci_cleanup = reinterpret_cast<FT_pci_cleanup>(
+      dlsym(handle, "pci_cleanup"));
+  interface->pci_scan_bus = reinterpret_cast<FT_pci_scan_bus>(
+      dlsym(handle, "pci_scan_bus"));
+  interface->pci_fill_info = reinterpret_cast<FT_pci_fill_info>(
+      dlsym(handle, "pci_fill_info"));
+  interface->pci_lookup_name = reinterpret_cast<FT_pci_lookup_name>(
+      dlsym(handle, "pci_lookup_name"));
+  if (interface->pci_alloc == NULL ||
+      interface->pci_init == NULL ||
+      interface->pci_cleanup == NULL ||
+      interface->pci_scan_bus == NULL ||
+      interface->pci_fill_info == NULL ||
+      interface->pci_lookup_name == NULL) {
+    LOG(ERROR) << "Missing required function(s) from libpci";
+    dlclose(handle);
+    delete interface;
+    return NULL;
+  }
+  return interface;
+}
+
+// This close the dynamically opened libpci and delete the interface.
+void FinalizeLibPci(PciInterface*& interface) {
+  DCHECK(interface != NULL && interface->lib_handle != NULL);
+  dlclose(interface->lib_handle);
+  delete interface;
+  interface = NULL;
+}
+
+// This creates an offscreen GL context fir gl queries.  Returned GLContext

[Ken Russell (switch to Gerrit), 2010/12/18 03:03:27]

fir -> for

[Zhenyao Mo, 2010/12/20 17:36:47]

Done.

+// should be deleted in FinalizeGLContext.
+gfx::GLContext* InitializeGLContext() {
+  if (!gfx::GLContext::InitializeOneOff()) {
+    LOG(ERROR) << "gfx::GLContext::InitializeOneOff() failed";
+    return NULL;
+  }
+  gfx::GLContext* context = gfx::GLContext::CreateOffscreenGLContext(NULL);
+  if (context == NULL) {
+    LOG(ERROR) << "gfx::GLContext::CreateOffscreenGLContext(NULL) failed";
+    return NULL;
+  }
+  if (!context->MakeCurrent()) {
+    LOG(ERROR) << "gfx::GLContext::MakeCurrent() failed";
+    context->Destroy();
+    delete context;
+    return NULL;
+  }
+  return context;
+}
+
+// This destroy and delete the GL context.
+void FinalizeGLContext(gfx::GLContext*& context) {
+  DCHECK(context != NULL);
+  context->Destroy();
+  delete context;
+  context = NULL;
+}
+
+}  // namespace anonymous
+
 namespace gpu_info_collector {
 
 bool CollectGraphicsInfo(GPUInfo* gpu_info) {
-  // TODO(rlp): complete this function
-  // TODO(apatrick): this illustrates how can_lose_context will be implemented
-  // on this platform in the future.
-  // bool can_lose_context =
-  //     gfx::GetGLImplementation() == gfx::kGLImplementationEGLGLES2;
-
-  gpu_info->SetProgress(GPUInfo::kComplete);
-
-  return true;
+  gfx::GLContext* context = InitializeGLContext();
+  if (context == NULL)
+    return false;
+
+  // TODO(zmo): collect driver version, pixel shader version, vertex shader
+  // version, and gl version.
+  std::wstring driver_version = L"";
+  uint32 pixel_shader_version = 0;
+  uint32 vertex_shader_version = 0;
+  uint32 gl_version = 0;
+  bool can_lose_context =
+      (gfx::GetGLImplementation() == gfx::kGLImplementationEGLGLES2);
+
+  // TODO(zmo): be more flexible about library name.
+  PciInterface* interface = InitializeLibPci("libpci.so.3");
+  if (interface == NULL) {
+    FinalizeGLContext(context);
+    return false;
+  }
+
+  PciAccess* access = (interface->pci_alloc)();
+  DCHECK(access != NULL);
+  (interface->pci_init)(access);
+  (interface->pci_scan_bus)(access);
+  std::vector<PciDevice*> gpu_list;
+  PciDevice* gpu_active = NULL;
+  for (PciDevice* device = access->device_list;
+       device != NULL; device = device->next) {
+    (interface->pci_fill_info)(device, 33);  // Fill the IDs and class fields.
+    if (device->device_class == 0x0300) {  // Device class is DISPLAY_VGA.
+      gpu_list.push_back(device);
+    }
+  }
+  if (gpu_list.size() == 1) {
+    gpu_active = gpu_list[0];
+  } else {
+    // If more than one graphics card are identified, find the one that matches
+    // gl VENDOR and RENDERER info.

[Ken Russell (switch to Gerrit), 2010/12/18 03:03:27]

Out of curiosity did you test this code path on a machine with more than one
graphics card?

[Zhenyao Mo, 2010/12/20 17:36:47]

I've tested this path by commenting out the condition temporarily and it worked
fine (able to identify the exact card through libpci and glGetString name
matching).

On 2010/12/18 03:03:27, kbr wrote:

+    std::string gl_vendor_string =
+        reinterpret_cast<const char*>(glGetString(GL_VENDOR));
+    std::string gl_renderer_string =
+        reinterpret_cast<const char*>(glGetString(GL_RENDERER));
+    const int buffer_size = 255;
+    scoped_array<char> buffer(new char[buffer_size]);

[Ken Russell (switch to Gerrit), 2010/12/18 03:03:27]

The contents of this array are uninitialized at this point.

[Zhenyao Mo, 2010/12/20 17:36:47]

I looked into pci_lookup_name source code.  It returns buffer.get() if it
successfully found the name; returns NULL if no matching; returns an error
message (not using the input buffer.get()) if buffer size is not large enough. 
So I'll check the returned value of pci_lookup_name, and no need to initialize
the buffer.

On 2010/12/18 03:03:27, kbr wrote:

+    std::vector<PciDevice*> candidates;
+    for (size_t i = 0; i < gpu_list.size(); ++i) {
+      PciDevice* gpu = gpu_list[i];
+      (interface->pci_lookup_name)(
+          access, buffer.get(), buffer_size, 1, gpu->vendor_id);

[Ken Russell (switch to Gerrit), 2010/12/18 03:03:27]

Is this guaranteed to succeed? If not, how does it indicate failure? If it fails
and doesn't touch the contents of "buffer" then "buffer" will contain garbage.

[Zhenyao Mo, 2010/12/20 17:36:47]

See comment above.

On 2010/12/18 03:03:27, kbr wrote:

+      std::string vendor_string = buffer.get();

[Ken Russell (switch to Gerrit), 2010/12/18 03:03:27]

Must ensure that we don't try to construct a std::string out of garbage.

[Zhenyao Mo, 2010/12/20 17:36:47]

See comment above.

On 2010/12/18 03:03:27, kbr wrote:

+      if (!StartsWithASCII(gl_vendor_string, vendor_string, false))

[Ken Russell (switch to Gerrit), 2010/12/18 03:03:27]

I think it's good you're using a case-insensitive search here. I'd suggest
creating a

const bool case_sensitive = false;

and passing that value down to make the call sites more self-descriptive.

[Zhenyao Mo, 2010/12/20 17:36:47]

Done.

+        continue;
+      (interface->pci_lookup_name)(
+          access, buffer.get(), buffer_size, 2, gpu->vendor_id, gpu->device_id);
+      std::string device_string = buffer.get();

[Ken Russell (switch to Gerrit), 2010/12/18 03:03:27]

Same caveat about constructing a string out of potential garbage.

[Zhenyao Mo, 2010/12/20 17:36:47]

See comment above.

On 2010/12/18 03:03:27, kbr wrote:

+      size_t begin = device_string.find_first_of('[');
+      size_t end = device_string.find_last_of(']');
+      if (begin != std::string::npos && end != std::string::npos &&
+          begin < end) {
+        device_string = device_string.substr(begin + 1, end - begin - 1);
+      }
+      if (StartsWithASCII(gl_renderer_string, device_string, false)) {

[Ken Russell (switch to Gerrit), 2010/12/18 03:03:27]

Use case_sensitive here too.

[Zhenyao Mo, 2010/12/20 17:36:47]

Done.

+        gpu_active = gpu;
+        break;
+      }
+      // If a device's vendor matches gl VENDOR string, we want to consider the
+      // possibility that libpci may not return the exact same name as gl
+      // RENDERER string.
+      candidates.push_back(gpu);
+    }
+    if (gpu_active == NULL && candidates.size() == 1)
+      gpu_active = candidates[0];
+  }
+  if (gpu_active != NULL) {
+    gpu_info->SetGraphicsInfo(gpu_active->vendor_id,
+                              gpu_active->device_id,
+                              driver_version,
+                              pixel_shader_version,
+                              vertex_shader_version,
+                              gl_version,
+                              can_lose_context);
+    gpu_info->SetProgress(GPUInfo::kComplete);
+  }
+  (interface->pci_cleanup)(access);
+  FinalizeLibPci(interface);
+  FinalizeGLContext(context);
+  return (gpu_active != NULL);
 }
 
 }  // namespace gpu_info_collector