Sometimes blocks when launched via NickelMenu

[baskerville]

I've tried launching Plato via NickelMenu using the following entry:

menu_item  :main  :Start Plato      :cmd_spawn  :quiet:/usr/local/Plato/plato.sh

And it sometimes results in Plato blocking (for a minute or so) on this line! What blocks in load_json is the second line.

[pgaskin]

Can you try using strace on Plato?

Also, does this also happen if you use cmd_output and & to background Plato?

[baskerville]

Here's the relevant excerpt from the output of strace:

open("/mnt/onboard/books/.metadata.json", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = 5</mnt/onboard/books/.metadata.json> <0.000>
read(5</mnt/onboard/books/.metadata.json>, "{", 1) = 1 <0.000>
getrandom(NULL, 0, GRND_NONBLOCK)       = -1 ENOSYS (Function not implemented) <0.000>
open("/dev/random", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = 6</dev/random> <0.000>
ioctl(6</dev/random>, FIOCLEX)          = 0 <0.000>
poll([{fd=6</dev/random>, events=POLLIN}], 1, -1) = 1 ([{fd=6, revents=POLLIN}]) <172.034>
close(6</dev/random>)                   = 0 <0.000>
open("/dev/urandom", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = 6</dev/urandom> <0.000>
ioctl(6</dev/urandom>, FIOCLEX)         = 0 <0.000>
read(6</dev/urandom>, "\342\2367\240\370\334D\372\323\306\243\33\324n\312\"p\201Y\206\267\2456i-\314\27D^\1\320r", 32) = 32 <0.000>
read(5</mnt/onboard/books/.metadata.json>, "\n", 1) = 1 <0.000>
read(5</mnt/onboard/books/.metadata.json>, " ", 1) = 1 <0.000>
read(5</mnt/onboard/books/.metadata.json>, " ", 1) = 1 <0.000>
read(5</mnt/onboard/books/.metadata.json>, "\"", 1) = 1 <0.000>
read(5</mnt/onboard/books/.metadata.json>, "5", 1) = 1 <0.000>
read(5</mnt/onboard/books/.metadata.json>, "0", 1) = 1 <0.000>
read(5</mnt/onboard/books/.metadata.json>, "9", 1) = 1 <0.000>

The call that blocks is poll([{fd=6</dev/random>, events=POLLIN}], 1, -1). It seems to be blocking for 172 seconds!

[pgaskin]

That seems to be blocking due to low entropy before falling back to weaker random numbers. How long have you booted for? Also, can you record the value of /proc/sys/kernel/random/entropy_avail every second or so starting from right after Nickel's launch?

Although, I wonder what's needing truly random numbers in Plato... I suspect it's another thread doing it rather than the JSON parser.

[baskerville]

Although, I wonder what's needing truly random numbers in Plato... I suspect it's another thread doing it rather than the JSON parser.

I'm afraid it might be the fancy hasher I'm using for the DB's hash map.

[pgaskin]

So probably shepmaster/twox-hash#13.

[NiLuJe]

/dev/random is apparently quirky on Kobo, even long after the boot (possibly especially on older devices).

I've had to neuter an OpenSSL feature after a similar issue: koreader/koreader-base#1099

[baskerville]

I thought it wasn't worth reporting a bug, since the trace is clearly reflecting the behavior of the getrandom crate.

The fact that the function might block is also acknowledged.

I've decided, as consequence, to replace things that might directly or indirectly depend on OsRng, so as to never poll /dev/random.

[NiLuJe]

Yeah, I'm no expert, but what's surprising is the amount of time it can block, and the fact that it does so long after the early boot.

Might be worse just after a resume from suspend?

[pgaskin]

Well, on most devices there are more sources of entropy, but the Kobo is pretty bare-bones. And, the lib does use GRND_NONBLOCK (and from a quick look, would fall back to pseudo random numbers afterwards), but that's only supported on 3.14+, so it falls back to a blocking random read.