Rustup uses backtrace regardless of RUST_BACKTRACE

[pravic]

I noticed that rustup update scans my HDD every time it been called.

I found that it always initializes a symbol helper (dbghelp!SymInitialize) at begin.

Setting RUST_BACKTRACE=0 or removing RUST_BACKTRACE does not help.

Any explanation/recomendations about it? I tested cargo also, it does not load .pdb every time like rustup does.

[Diggsey]

I suspect this is due to the use of the error_chain crate, which itself depends on the backtrace crate, and is used to capture stack traces within Error objects.

[pravic]

@Diggsey thanks regardless of I am not so lucky to hear about that.

Any way to disable it?

I've set _NT_SYMBOL_PATH to NULL before calling rustup and it helps partially: rustup scans only root directory of the current drive about PDB files.

[brson]

@pravic Does this only happen when rustup ends with an error or does it also happen when rustup exits successfully? (I don't know why any backtrace code would be running on success)

I'd like to know more about what it means to 'scan your HDD'. Why is this a problem for you?

[pravic]

@brson There are no errors in regular rustup update, so thats why I complaining about it.

The problem is in unnesessary HDD access which is quiet noisy and slowly. And thats despite of incorrect behavior with RUST_BACKTRACE ignoring.

[brson]

Here's the strace output of rustup --help for the open and stat syscalls:

open("/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=38191, ...}) = 0
open("/lib/x86_64-linux-gnu/libz.so.1", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=104824, ...}) = 0
open("/lib/x86_64-linux-gnu/libdl.so.2", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=14608, ...}) = 0
open("/lib/x86_64-linux-gnu/libpthread.so.0", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0755, st_size=138744, ...}) = 0
open("/lib/x86_64-linux-gnu/libgcc_s.so.1", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=89696, ...}) = 0
open("/lib/x86_64-linux-gnu/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0755, st_size=1864888, ...}) = 0
open("/lib/x86_64-linux-gnu/libm.so.6", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=1088952, ...}) = 0
open("/lib/x86_64-linux-gnu/librt.so.1", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=31712, ...}) = 0
open("/sys/devices/system/cpu/online", O_RDONLY|O_CLOEXEC) = 3
open("/proc/self/maps", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0444, st_size=0, ...}) = 0
open("/proc/self/exe", O_RDONLY)        = 3
open("/lib/x86_64-linux-gnu/libz.so.1", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libdl.so.2", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libpthread.so.0", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libgcc_s.so.1", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libc.so.6", O_RDONLY) = 3
open("/lib64/ld-linux-x86-64.so.2", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libm.so.6", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/librt.so.1", O_RDONLY) = 3
stat("/home/brian/.cargo/bin/rustup-init", 0x7ffe018adbf0) = -1 ENOENT (No such file or directory)
stat("/home/brian/.cargo/bin/multirust-setup", 0x7ffe018adbf0) = -1 ENOENT (No such file or directory)

That doesn't look unexpected, though I'm not sure what the fstats are. Mostly the dynamic linker. I bet what's happening is that there are errors triggered internally, that are handled and turned into no-errors. The generation of those errors triggers a backtrace that is never seen.

So the same strace on rustup update foo (which triggers an error):

open("/etc/ld.so.cache", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=38191, ...}) = 0
open("/lib/x86_64-linux-gnu/libz.so.1", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=104824, ...}) = 0
open("/lib/x86_64-linux-gnu/libdl.so.2", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=14608, ...}) = 0
open("/lib/x86_64-linux-gnu/libpthread.so.0", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0755, st_size=138744, ...}) = 0
open("/lib/x86_64-linux-gnu/libgcc_s.so.1", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=89696, ...}) = 0
open("/lib/x86_64-linux-gnu/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0755, st_size=1864888, ...}) = 0
open("/lib/x86_64-linux-gnu/libm.so.6", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=1088952, ...}) = 0
open("/lib/x86_64-linux-gnu/librt.so.1", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=31712, ...}) = 0
open("/sys/devices/system/cpu/online", O_RDONLY|O_CLOEXEC) = 3
open("/proc/self/maps", O_RDONLY|O_CLOEXEC) = 3
fstat(3, {st_mode=S_IFREG|0444, st_size=0, ...}) = 0
open("/proc/self/exe", O_RDONLY)        = 3
open("/lib/x86_64-linux-gnu/libz.so.1", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libdl.so.2", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libpthread.so.0", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libgcc_s.so.1", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libc.so.6", O_RDONLY) = 3
open("/lib64/ld-linux-x86-64.so.2", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/libm.so.6", O_RDONLY) = 3
open("/lib/x86_64-linux-gnu/librt.so.1", O_RDONLY) = 3
stat("/home/brian/.cargo/bin/rustup-init", 0x7ffcbcd299f0) = -1 ENOENT (No such file or directory)
stat("/home/brian/.cargo/bin/multirust-setup", 0x7ffcbcd299f0) = -1 ENOENT (No such file or directory)
stat("/home/brian/.terminfo", 0x7ffcbcd28c10) = -1 ENOENT (No such file or directory)
stat("/etc/terminfo", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
stat("/etc/terminfo/s/screen", 0x7ffcbcd28c10) = -1 ENOENT (No such file or directory)
stat("/etc/terminfo/73/screen", 0x7ffcbcd28c10) = -1 ENOENT (No such file or directory)
stat("/lib/terminfo", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
stat("/lib/terminfo/s/screen", {st_mode=S_IFREG|0644, st_size=1587, ...}) = 0
open("/lib/terminfo/s/screen", O_RDONLY|O_CLOEXEC) = 3
stat("/home/brian/.multirust", {st_mode=S_IFDIR|0775, st_size=4096, ...}) = 0
stat("/home/brian/.multirust/version", 0x7ffcbcd27e50) = -1 ENOENT (No such file or directory)
stat("/home/brian/.multirust", {st_mode=S_IFDIR|0775, st_size=4096, ...}) = 0
stat("/home/brian/.multirust/settings.toml", {st_mode=S_IFREG|0664, st_size=169, ...}) = 0
open("/home/brian/.multirust/settings.toml", O_RDONLY|O_CLOEXEC) = 3
stat("/home/brian/.multirust/toolchains/foo", 0x7ffcbcd28d70) = -1 ENOENT (No such file or directory)
stat("/home/brian/.terminfo", 0x7ffcbcd32870) = -1 ENOENT (No such file or directory)
stat("/etc/terminfo", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
stat("/etc/terminfo/s/screen", 0x7ffcbcd32870) = -1 ENOENT (No such file or directory)
stat("/etc/terminfo/73/screen", 0x7ffcbcd32870) = -1 ENOENT (No such file or directory)
stat("/lib/terminfo", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
stat("/lib/terminfo/s/screen", {st_mode=S_IFREG|0644, st_size=1587, ...}) = 0
open("/lib/terminfo/s/screen", O_RDONLY|O_CLOEXEC) = 3

Suddenly some code is very interested in terminfo, but still nothing that looks particularly out of the ordinary.

But I guess since you mentioned .pdb files you are on Windows. Maybe there are different performance characteristics for obtaining backtraces there.

[alexcrichton]

@pravic when you mention that SymInitialize is on the stack, is that the symbol which seems like it's causing all of the I/O access? The backtrace crate does indeed call that and it's pretty liberal about its usage of it, there's no global cache or anything it just ends up calling it once per stack trace and once per symbol. If it's a slow function, then it should likely definitely cache that!

@brson I think that error-chain captures a stack trace when any error happens, right? That means that a backtrace would be captured for an error which is later ignored I believe? It may be worth optimizing some stuff in backtrace-rs for this use case if it's causing perf problems.

[brson]

Yes, backtraces are generated for all errors regardless of whether they are displayed to the user.

[alexcrichton]

Ok, I've pushed a commit to backtrace-rs which should greatly reduce the number of times SymInitialize is called. @brson do you want to try to get a windows build of rustup with that commit and see if @pravic can test it out?

[brson]

I'm having some issues with my windows build at the moment, but when I've resolved them I'll post a testing build.

[brson]

@pravic Here's a build that contains @alexcrichton's fixes. Maybe you can test it and see if it performs better for you: https://s3-us-west-1.amazonaws.com/static-rust-lang-org/rustup/temp-testing/rustup-init.exe

[pravic]

Just checked (rustup 0.3.0):
rustup --help (of course, without errors) produces 3977 filesystem events to query and read symbol files (pdb) which takes around 4 seconds. And this is without online symbol server access — it would take minutes in that case.

Well, I checked new build. Behavior the same, since all symbols were loaded on first SymInitialize call, because it enumerates all process modules and tries to load (or download) pdb file for each.

But yes, backtrace calls SymInitialize many times before patch:

$ rustup --help
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
dbghelp_init!
rustup 0.3.0
The Rust toolchain installer

The only solution I found, it is disable "dbghelp" feature of backtrace completely. I definitely do not need a flying error chains with linked backtraces, RUST_BACKTRACE is enough for me when it needed.

[brson]

I wonder why running rustup --help takes 4 seconds on your machine but not mine. How are you determining the filesystem event count? I'd like to try here.

[pravic]

I'm amenable to making this behavior an option at least, but I also want to figure out why you are seeing so much disk thrashing on windows

Guess, thats how windows symbol handler works.

rustup --help loads ~50 dlls in process. Symbol handler enumerates all of them and tries to find an appropriate PDB file for each module. If module not found and online symbol server is specified in _NT_SYMBOL_PATH, dbghelp will try to lookup and download PDB file from the server.