Unexpected behavior with DMON_SLEEP_INTERVAL 1000 on linux

[ludolpif]

Hi,
I get some infinite hangs in my SDL3 app while setting a new dmon watcher.

• DMON_SLEEP_INTERVAL 999 : nothing bad happens
• DMON_SLEEP_INTERVAL 1000 : hangs
• DMON_SLEEP_INTERVAL 1001 : hangs

I am unsure, but I suspect this:

struct timespec req = { (time_t)DMON_SLEEP_INTERVAL / 1000, (long)(DMON_SLEEP_INTERVAL * 1000000) };
nanosleep(&req, &rem);

to wait "forever" , then another part is stuck in a mutex wait as seergdb shown me. The nanosecond part should never exceed 999999999, it miss a %1000 I think (untested).

struct timespec req = { (time_t)DMON_SLEEP_INTERVAL / 1000, (long)((DMON_SLEEP_INTERVAL % 1000) * 1000000) };

Cheers,

[septag]

I cannot reproduce this. Which backend are you testing ? linux or mac ?

[ludolpif]

Hi,
Sorry about not mentioning enough details at first place in this one:

• GNU/Linux Debian 13
• dmon from 89d46c5
• gcc 4:14.2.0-1
• libsdl3-0:amd64 3.2.10+ds-1
• flecs v4.1.4

While trying to make a minimal reproducer... it does not hang yet. I will try harder. As it doesn't show up, I tried also this:

#define DMON_SLEEP_INTERVAL 10000

and it emits a compilation warnings, suggesting that the mentionned line is not really healthy, independently on the deadlock I have hard time to isolate:

gcc -Wall main.c -std=gnu99 -I/usr/include/pipewire-0.3 -I/usr/include/spa-0.2 -D_REENTRANT -I/usr/include/libdrm -I/usr/include/libdecor-0  -c -o main.o
In file included from main.c:8:
dmon.h: In function ‘_dmon_thread’:
dmon.h:1072:92: warning: integer overflow in expression of type ‘int’ results in ‘1410065408’ [-Woverflow]
 1072 |     struct timespec req = { (time_t)DMON_SLEEP_INTERVAL / 1000, (long)(DMON_SLEEP_INTERVAL * 1000000) };
      |                                                                                            ^
gcc main.o flecs.o -lSDL3  -lrt -lpthread -lm -o dmon-issue40

from:

ludolpif@lud-5490:~/git/dmon-issue40$ nl -ba main.c | head
     1	#include "flecs.h"
     2	
     3	#define SDL_MAIN_USE_CALLBACKS
     4	#include <SDL3/SDL_main.h>
     5	
     6	#define DMON_IMPL
     7	#define DMON_SLEEP_INTERVAL 10000
     8	#include "dmon.h"
     9	
    10	#include <SDL3/SDL.h>

I may have a wrong memory ownership assumption in my original hanging code, I can't really see for now how it could achieve things that behave differently with 999 and 1000 but I have to investigate harder. It seems, that with my current not hanging reproducer there is no perceptible delay between a file modification event and the report by the test program, which not the behavior I used to see in the hanging program.

Hoping give news soon on this!
Thanks for all the fish,

[ludolpif]

I can confirm that I was passing some stack-allocated char * from dmon callback to flecs, and the late one was expecting heap-allocated char * (the user program should take care, mine wasn't). So deadlock was probably there only because I got into undefined behavior elsewhere, and there is a 1 Hz ECS task active, may helping to have a side effect with DMON_INTERVAL 999 or 1000.

So the issue is only the calculation for the timespec struct. For DMON_INTERVAL 999 it will wait 999ms, for DMON_INTERVAL 1000, it will wait 1*1000+1000=2000ms , for 10000 it overflows, and it is unclear that is legal or portable to pass a ns part greater than 999 999 999.