Spurious 3-hour timeout for dist-* jobs due to 20 minutes spent on compiling stage0-libstd

[kennytm]

Symptom: One of the dist-*-linux job timed out after 3 hours. The stage0-std phase needs over 1000 seconds to compile.

Typically all 3 compilers (host stage0, host stage1, target stage1) are completely built, but the RLS build or final deployment doesn't finish in time.

The log should be inspected to ensure this is not caused by network error (e.g. cloning a submodule taking over 1 hour to complete).

Examples:

• PowerPC (Build completed at 2:57:22, but 3 minutes isn't enough for deployment)
• MIPS (Timed out before completing RLS)
• PowerPC 64 (Timed out before completing RLS) (not 20-minute stage0-libstd)
• i686 (Timed out before completing RLS) (not 20-minute stage0-libstd)

[Mark-Simulacrum]

I'm going to try and investigate at some point this week so assigning myself.

[kennytm]

Timing breakdown for these 4 logs

	i686	ppc64	mips	ppc
make-prepare	69.19	74.98	60.87	52.38
pytest/bootstrap	0.00	0.00	0.00	0.00
stage0-std	91.80	104.70	1260.49	1250.99
stage0-test	20.93	24.50	22.87	17.52
stage0-rustc	1478.91	1616.18	1511.29	1234.51
llvm	223.83	253.40	379.83	320.81
stage0-trans-llvm	205.13	244.81	206.33	168.54
llvm-emscripten	136.07
stage0-trans-emscripten	116.82
stage1-std	200.31	101.66	86.61	68.47
stage1-test	23.13	27.84	23.53	18.37
stage1-rustc	1727.49	1871.25	1667.80	1325.40
stage1-trans-llvm	217.83	255.12	221.41	177.87
stage1-trans-emscripten	126.75
stage1-std	89.00	99.63	85.85	65.50
stage0-unstable-book-gen	13.26	14.43	14.54	11.37
stage0-rustbook	240.16	251.63	208.86	196.78
stage2-rustdoc	287.74	306.87	289.94	239.42
doc/std	15.35	15.90	14.23	12.92
doc/compiler	22.85	26.30	24.32	18.23
stage1-rustc	1701.37	1885.18	1566.31	1342.24
doc/proc_macro	1.35	1.38	1.43	1.19
stage2-error_index_generator	8.24	8.74	8.56	6.48
stage0-fabricate	129.00	137.81	139.86	109.77
llvm	223.84	228.95	368.32	287.82
stage1-trans-llvm	222.48	281.78	218.72	196.22
llvm-emscripten	133.85
stage1-trans-emscripten	131.67
stage2-rustdoc	304.44	327.96	275.49	236.46
dist/rustsrc	171.99
stage2-cargo	589.43	629.74	498.94	406.81
stage2-rustfmt	338.24	361.24	314.88	264.20
stage2-cargo-fmt	0.00	0.00	0.00	0.00
stage2-rls		665.78	523.11	440.94

The MIPS and PowerPC spent >20 minutes on stage0-std which may be a different bug (maybe related to rust-lang/cargo#5030 as suggested by @alexcrichton)

PowerPC64 looks most suspicious, as everything looks so uniform.

Graphed

[alexcrichton]

Looking at some of those with the travis output:

• PowerPC
  • The travis container was rebuilt here which took ~20m, this would be an anomaly
  • This also took 20m to build libstd in stage0 (!)
• MIPS
  • Took 20m to build libstd in stage0 (!)

Not sure what happened with PowerPC 64 and i686... I'm looking into the 20m libstd stage0 bug

[kennytm]

I've narrowed down this bug to focus on the 20-minute libstd builds.

[alexcrichton]

So some information about why libstd is taking 20 minutes to build. I randomly ran across this the other day and was very surprised to see what was happening. The behavior that I was seeing is that building libstd hung in stage0 and when looking at top Cargo was just sitting there eating tons of CPU.

Looking into why Cargo was eating CPU I discovered that Cargo's read2 function wasn't quite right, namely that when one of the fds it was interested in was closed it would just spin infinitely without blocking until the other fd was closed. The fix (rust-lang/cargo#5030) should fix the spin behavior but I don't believe it will block the "cargo hangs for 20 minutes" bug.

I have no idea why cago is waiting 20 minutes for child file descriptors to get closed. The fix in rust-lang/cargo#5030, if I understand it correctly, will basically just cause cargo to block correctly rather than spin, but it won't help actually get the condition Cargo is waiting on to come about any sooner.

I've got no idea why one fd of a child process would get closed and the other wouldn't for 20 minutes. I wonder, though if this is related to the sccache and build scripts change as sccache could in theory hold a descriptor open for a long time by accident. I haven't been able to confirm this locally though.

[alexcrichton]

Ok I'm like 99% sure it's the sccache change now. The sccache server is starting in the libstd/libcore build scripts and is somehow inheriting too many fds. My current assumption for what's happening is:

• Cargo spawns a build script with stdout/stderr fds
• The build script then spawns a subprocess like the configure script
• The configure script now does stuff that doesn't use CLOEXEC. For example the configure script may dup its own stdout/stderr onto separate fds
• The configure script eventually invokes sccache to test it
• Sccache realizes no server is running so it daemonizes
• Sccache's daemon inherits the original stdout/stderr fds, dup'd by the configure script as not CLOEXEC, and keeps them open
• Everything exits except for sccache's daemon which stays alive for a 10m idle time
• Cargo now attempts to finish reading the original stdout/sdterr fds. One is closed but one leaked into sccache's server, so Cargo thinks it never gets closed.

I'm not entirely sure why a configure script (aka sh) would dup a stdio fd into a separate location without CLOEXEC, but it's currently the only way that I can explain this. I think we've got two fixes: either revert the sccache patch or try to start the sccache server earlier on in the compilation process. I vote we revert the sccache patch.

@kennytm I'm gonna be tight on time soon, mind doing the revert for me?

[alexcrichton]

#48209 includes a revert of the sccache commit, presumably if things stabilize after that lands we can close this.