deps: upgrade to libuv 1.20.2 (formerly 1.20.1)

[cjihrig]

Fixes: #20112
Fixes: #19903

Checklist

• make -j4 test (UNIX), or vcbuild test (Windows) passes
• commit message follows commit guidelines

[MylesBorins]

rubber stamp LGTM
We should fast track to get this in 10.x asap

[vsemozhetbyt]

CI: https://ci.nodejs.org/job/node-test-pull-request/14362/

[Trott]

Not sure why macOS CI failed, but here's a re-run: https://ci.nodejs.org/job/node-test-commit-osx/17931/

AIX re-run: https://ci.nodejs.org/job/node-test-commit-aix/14325/

[richardlau]

AIX compilation failure is in new code introduced by this PR, but the libuv CI for the PR that introduced it passed on AIX, libuv/libuv#1795.

../deps/uv/src/unix/thread.c:542:28: error: size of array 'static_assert_failed' is negative
 STATIC_ASSERT(sizeof(uv_sem_t) >= sizeof(uv_semaphore_t*));
                            ^~~~~~~~~~~~~~~~~~~~

[addaleax]

I guess we could just move that STATIC_ASSERT into the #ifdef __GLIBC__ part of the new code … it would mask the fact that the uv__custom_sem_* methods don’t work if the condition isn’t met, but at least right now they wouldn’t be used then anyway.

I would expect for Node.js CI and libuv CI to behave the same too, though.

[richardlau]

I would expect for Node.js CI and libuv CI to behave the same too, though.

Perhaps a different level of compiler is being used. There are lines like this for AIX, e.g., https://ci.nodejs.org/job/node-test-commit-aix/nodes=aix61-ppc64/14325/console

ccache /home/iojs/gcc-6.3.0-1/opt/freeware/bin/gcc ...

which would indicate gcc 6.3.0 bring used -- @mhdawson is this expected? The most recent updates in nodejs/build#925 suggest the upgrade to gcc 6.3.0 is still in progress.

[jasnell]

We really need to try to get this landed today before I do the first RC build. I will be kicking off that build this afternoon between 1 and 2 pm pacific.

[richardlau]

I think I've worked out why the libuv AIX CI passed and the Node.js AIX CI is failing -- I think libuv is building 32-bit code while Node.js is building 64-bit. The following is a reduced test showing the difference (I believe uv_sem_t ends up as a sem_t from eyeballing the code):

-bash-4.4$ cat test.c
#include <stdio.h>
#include <semaphore.h>

int main(int argc, void* argv)
{
  printf("sizeof sem_t %i\n", sizeof(sem_t));
  printf("sizeof void* %i\n", sizeof(void*));
  printf("static assert %i\n", 1-2*!(sizeof(sem_t) >= sizeof(void*)));
  return 0;
}
-bash-4.4$ gcc -o test test.c # 32-bit
-bash-4.4$ ./test
sizeof sem_t 4
sizeof void* 4
static assert 1
-bash-4.4$ gcc -o test -maix64 test.c # 64-bit
-bash-4.4$ ./test
sizeof sem_t 4
sizeof void* 8
static assert -1
-bash-4.4$

This seems to suggest that the following comment won't work on 64-bit AIX as the uv_sem_t type isn't large enough to hold a pointer:

/* To preserve ABI compatibility, we treat the uv_sem_t as storage for
 * a pointer to the actual struct we're using underneath. */

cc @addaleax

[mhdawson]

We just moved up the compiler levels for 10.x (not just for AIX) so it is expected that mater is using a newer compiler. We'll have to see if the requirements for libuv have been moved up as well and we should use the newer compiler as well. That may or may not be the case will need to get input from libuv maintainers, but I think that should be in a different thread.

[mhdawson]

@gireeshpunathil can you take a look as well if this is not resolved by the time you get in.

[richardlau]

@mhdawson I believe the compilers being different was a red herring and the main issue is that libuv doesn't actually appear to have support for compiling 64-bit binaries on AIX (I can't find any -maix64 flags in the libuv repository) and one of the new libuv commits doesn't compile on 64-bit AIX.

[mhdawson]

@richardlau ok, so we should open an issue to have the job in the libuv CI to be 64 bit instead of 32 bit. Can you do that. There still needs to be a solution to the compile failure though.

[richardlau]

@mhdawson I've raised nodejs/build#1245 but it needs libuv/libuv#1807 to land first so that libuv can pass through the right compile and link flags.

[addaleax]

I mean, leaving all the CI stuff aside, is there anything wrong with the suggestion from #20129 (comment)? If yes, then we can still just disable the STATIC_ASSERT completely (and temporarily) as a stop-gap measure.

[richardlau]

I mean, leaving all the CI stuff aside, is there anything wrong with the suggestion from #20129 (comment)?

Sorry about being distracted with the CI. The suggestion looks okay to me.

[richardlau]

Raised libuv/libuv#1808 as per #20129 (comment).

[jasnell]

Is this on track to land by monday afternoon?

[cjihrig]

Landed the relevant PRs in libuv. Here is a CI run of current Node master and libuv 1.x. macOS hasn't finished yet, but so far everything else is green. I can cut a 1.20.2 release to replace this PR before Monday afternoon (cc: @bnoordhuis @santigimeno).

UPDATE: macOS finished. It doesn't look like there were failures, but test/parallel/test-child-process-fork-net2.js left a child process behind. This appears unrelated, per #20185 (comment) from a couple hours ago.

[cjihrig]

This has been updated to libuv 1.20.2. CI: https://ci.nodejs.org/job/node-test-pull-request/14426/

[cjihrig]

CI is green except for #18435. This should be good to go.

[jasnell]

Landed in 63d991b