src: make --use-largepages a runtime option

[gabrielschulhof]

Moves the option that instructs Node.js to-remap its static code to
large pages from a configure-time option to a runtime option. This
should make it easy to assess the performance impact of such a change
without having to custom-build.

Checklist

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

[addaleax]

Can you add a basic test that e.g. just spawns node --use-largepages -p 42 and makes sure that the output is correct (ideally verify both stderr and stdout)? That might give us a better idea under which circumstances this is or is not working

[bnoordhuis]

LGTM but can you also update doc/node.1 and can I suggest applying this patch so the build won't break on older Linux systems?

diff --git a/src/large_pages/node_large_page.cc b/src/large_pages/node_large_page.cc index 68fa178b40..d13af4c11e 100644 --- a/src/large_pages/node_large_page.cc+++ b/src/large_pages/node_large_page.cc@@ -352,7 +352,7 @@ MoveTextRegionToLargePages(const text_region& r){return -1} - ret = madvise(tmem, size, MADV_HUGEPAGE);+ ret = madvise(tmem, size, 14 /* MADV_HUGEPAGE */); if (ret == -1){PrintSystemError(errno); ret = munmap(tmem, size);

(MADV_HUGEPAGE was added in Linux 2.6.38 and our baseline for running node is 3.x but I don't know what exactly we require for building node.)

[gabrielschulhof]

@addaleax I added the test and I fixed the doc.
@bnoordhuis I replaced the constant with its value and added the man page entry.

[gabrielschulhof]

@bnoordhuis@addaleax@cjihrig@devnexen@gireeshpunathil I changed the implementation to one that accepts three values: 0 (default) to not map, 1 to map but to not report failure, and 2 to map and to report failure on stderr.

[lundibundi]

I think at least IsLargePagesEnabled and MapStaticCodeToLargePages in node_large_page.cc have to be changed as they won't compile/won't be correct on unsupported platforms due to conditional compilation in there. I think just adding conditional #else with correct error ret will be fine. They both won't have a return statement on the unsupported platforms.

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/27752/

[gabrielschulhof]

@addaleax@joyeecheung I updated the possible values to strings: "off", "silent", "verbose".
@lundibundi I applied your nits, and I added reporting for the case where mapping is not supported. I also placed the whole chunk into an ifdef to prevent compiling the code on unsupported platforms. On those platforms and in the verbose case, a message will be written to stderr that they're not supported.

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/27755/

[gabrielschulhof]

@devnexen thanks, updated.

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/27763/

[gabrielschulhof]

Dang! I made a copy/paste error while landing. Sorry I omitted you as a reviewer @jasnell!

[bcoe]

@gabrielschulhof we're seeing a bit of a hiccup in our nightly linux tests:

20:01:53 /bin/sh: 1: realpath: not found 20:01:53 gyp: Call to 'realpath src/large_pages/ld.implicit.script' returned exit status 127 whilein /home/iojs/build/workspace/node-test-commit-linux-coverage-daily/nodes/benchmark/node.gyp.

Which seems to crop up somewhere between December 21st and December 22nd when this landed. I'm guessing there might be a workaround in changing our compiler flags? Wondering if this is expected however (would guess moving this to a flag was meant to be a noop?).