Log file size problem about shadow

[stevenengler]

There seems to be millions of "mmap on fd X for Y bytes failed" warnings. I think this happens because the PR added a new check to make sure it wasn't a special file, and it turns out that tor does try to mmap one of these special files.

shadow/src/main/host/syscall/handler/mman.rs

Lines 268 to 273 in 83934ce

	// make sure we don't open special files like /dev/urandom, /etc/localtime etc. in the
	// plugin via mmap
	if unsafe { c::regularfile_getType(file) } != c::_FileType_FILE_TYPE_REGULAR {
	warn_once_then_debug!("Tried to mmap a non-regular-file");
	return Err(());
	}

This causes a warning further up in the calling code.

shadow/src/main/host/syscall/handler/mman.rs

Lines 210 to 218 in 83934ce

	// this fd exists in the plugin and not shadow; make sure to close this before returning (no
	// RAII)
	let plugin_fd = file.map(|file| Self::open_plugin_file(ctx.objs, fd, file));
	// the file is None for an anonymous mapping, or a non-null Some otherwise
	let Ok(plugin_fd) = plugin_fd.transpose() else {
	log::warn!("mmap on fd {fd} for {len} bytes failed");
	return Err(Errno::EACCES.into());
	};

[shadow_rs::host::syscall::handler::mman] (LOG_ONCE) Tried to mmap a non-regular-file
[shadow_rs::host::syscall::handler::mman] mmap on fd 3 for 118 bytes failed
[shadow_rs::host::syscall::handler::mman] mmap on fd 3 for 118 bytes failed
[shadow_rs::host::syscall::handler::mman] mmap on fd 3 for 118 bytes failed
[shadow_rs::host::syscall::handler::mman] mmap on fd 3 for 118 bytes failed
[shadow_rs::host::syscall::handler::mman] mmap on fd 3 for 118 bytes failed
[shadow_rs::host::syscall::handler::mman] mmap on fd 3 for 118 bytes failed
[shadow_rs::host::syscall::handler::mman] mmap on fd 3 for 118 bytes failed

I think I'll make this second warn log a debug log instead.

[stevenengler]

On my local machine, I only see these warnings for the "fileserver" host. On the benchmark runner, I see it for all tor hosts ("4uthority2", "relay33exitguard", "relay227exit", etc).

At least on my machine, the processes are trying to mmap an fd for "/etc/localtime".

00:00:01.000000000 [tid 1000] open("/etc/localtime", (empty), Mode(0x0)) = 3
00:00:01.000000000 [tid 1000] fstat() = 0
00:00:01.000000000 [tid 1000] mmap(0x0, 118, PROT_READ, MAP_PRIVATE, 3, 0) = 0xfffffffffffffff3 ((unknown errno 13))

We don't want to mmap the native "/etc/localhost" file, but by not doing this we make a lot of unnecessary extra syscalls. Every gettimeofday() call (which tor makes a lot of) results in several syscalls trying to mmap this file. This surprisingly doesn't seem to affect the benchmark results though, which is good.

00:03:15.000000000 [tid 1000] open("/etc/localtime", (empty), Mode(0x0)) = 7
00:03:15.000000000 [tid 1000] fstat() = 0
00:03:15.000000000 [tid 1000] mmap(0x0, 118, PROT_READ, MAP_PRIVATE, 7, 0) = 0xfffffffffffffff3 ((unknown errno 13))
00:03:15.000000000 [tid 1000] close(7) = 0
000000195000000000 [tid 1000] gettimeofday(...) = 0