attaching a trace

Deleted: SxsTrace.etl 128 KB

SxsTrace.etl 128 KB Download

maybe this is needed too/instead?

Deleted: sxstrace.txt 12.2 KB

sxstrace.txt 12.2 KB View Download

Relevant bit:
INFO: Parsing Manifest File C:\Program Files (x86)\Google\Chrome\Application\57.0.2950.0\57.0.2950.0.MANIFEST.
	INFO: Manifest Definition Identity is 57.0.2950.0,type="win32",version="57.0.2950.0".
	ERROR: The file osmesa.dll is included by multiple assemblies.
ERROR: Activation Context generation failed.

This is happening for me, too, but with the non-clang build. I'd say this is pretty important since it makes development a pain. Bumping to P1.

Okay, so osmesa.dll, libGLESv2.dll, and libEGL.dll are each listed twice in the manifest. Delete the duplicates and chrome.exe will launch.

aaaaand that's because .../gen/chrome_component.runtime_deps contains duplicates:

./libEGL.dll
./libGLESv2.dll
./osmesa.dll
Packages/ui/libEGL.dll
Packages/ui/libGLESv2.dll
Packages/ui/osmesa.dll

I think this is a result of //services/ui. I have no idea what's going on there.

Ken: What is this "Packages" directory? Is it expected that chrome.dll has a dependency on both copies of these dlls?

Here's a snippet of the dependency tree leading to references to the two instances of the DLLs:
  //chrome/installer/mini_installer:mini_installer_archive
    //chrome:chrome
      //chrome:chrome_initial
        //chrome:chrome_dll
          //chrome:main_dll
            //chrome:browser_dependencies
              //chrome/browser:browser
                //apps:apps
                  //chrome/browser/extensions:extensions
                    //chrome/browser/devtools:devtools
                      //chrome/common:common
                        //chrome/common/extensions/api:extensions_features
                          //chrome/common/extensions/api:api_features
                            //chrome/common/extensions/api:api_features_json_features
                              //extensions/common:common
                                //content/public/common:common
                                  //content:content
                                    //content/gpu:gpu_sources
                                      //content/child:child
                                        //components/mime_util:mime_util
                                          //media:media
                                            //gpu/command_buffer/common:common
                                              //gpu:gpu
                                                //gpu/command_buffer/service:service_sources
                                                  //gpu/config:config_sources
                                                    //ui/gl:gl
                                                      //third_party/mesa:osmesa
                                        //content/common:common
                                          //ui/shell_dialogs:shell_dialogs
                                            //ui/aura:aura
                                              //services/ui:ui
                                                //services/ui:copy_gl_libraries
                                                  //third_party/angle:libEGL...
                                                  //third_party/angle:libGLESv2...
                                                  //third_party/mesa:osmesa...

+gab and jbauman who know more about how create_installer_archive.py generates the manifest than I do. tl;dr: for you two: component chrome no longer runs after install because the manifest has duplicate entries for three dlls. i think this started happening as a result of what's going on in //services/ui. maybe we just need to filter duplicates when generating the manifest, but maybe the problem is that these duplicate dependencies are there in the first place. i'm not sure.

We should not be archiving anything in Packages.

Should chrome.dll really have a dependency on DLLs in Packages?

The dependencies are determined by GN: https://cs.chromium.org/chromium/src/chrome/installer/mini_installer/BUILD.gn?rcl=0&l=82

If we need any other solution than addressing the DEPS, the right solution is to address issue 581133 (better manifest method post-XP).

I've confirmed that r437776 ("Build services as standalone executables") is the one that caused the breakage by introducing the new dependency. Ken: can you confirm that this is expected and correct?

The data dependency is technically correct, yes, though it's only needed
under certain runtime conditions -- namely certain flags which are never
used in production by default.

Any progress with this issue?
I can't run locally built Chromium on a test laptop anymore.
Any workarounds other than not doing a component build?

I've just tried the workaround mentioned in comment #7 and it works.

r445870 reversed the change that caused this breakage, so I'm closing this issue. Fixing issue 581133 is still important to avoid future regressions.

I've just hit this issue again.
The duplicating entries are libGLESv2.dll and libEGL.dll.

FYI SwiftShader team. Possibly related to the inclusion of the SwiftShader dlls?

I'll take a swing.

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/d3f0779344128cbeea778bbc6ee2172f88aab25b

commit d3f0779344128cbeea778bbc6ee2172f88aab25b
Author: Greg Thompson <grt@chromium.org>
Date: Thu Aug 03 15:57:13 2017

Fix duplicate entries in component builds' version manifest.

Specifically, staged files are now ignored regardless of case and
containing directory.

BUG= 673900 

Change-Id: I72ded358fbde1c525eaec538ebdb184a79b58c6e
Reviewed-on: https://chromium-review.googlesource.com/598091
Reviewed-by: Gabriel Charette <gab@chromium.org>
Reviewed-by: Sébastien Marchand <sebmarchand@chromium.org>
Commit-Queue: Greg Thompson <grt@chromium.org>
Cr-Commit-Position: refs/heads/master@{#491752}
[modify] https://crrev.com/d3f0779344128cbeea778bbc6ee2172f88aab25b/chrome/tools/build/win/create_installer_archive.py

We should be back to normal now. Please re-open if you notice the dreaded side-by-side configuration error again. Thanks.