ISS-176 postgres execution failed with segmentation fault and sometimes --auth requires a value

[stormalf]

postgres execution failed with segmentation fault and sometimes --auth requires a value. Not sure yet, what is this segmentation fault and why when retrying with gdb instead of segmentation fault we receive FATAL --auth requires a value bucket: 162 sshift: 8 hashvalue: 2696752290 segment_num: 0 hashp->dsize: 256 hashp->dir: 0x33e53c8 1===bucket=162,Segmentation fault child process exited with exit code 139 2024-08-22 09:24:21.464 GMT [269657] FATAL: --auth requires a value [Inferior 3 (process 269657) exited with code 01] (gdb) no idea for now about the root cause: 0x00007ffff7ce1887 in __GI___libc_write (fd=4, buf=0x584020, nbytes=155) at ../sysdeps/unix/sysv/linux/write.c:26 26 ../sysdeps/unix/sysv/linux/write.c: No such file or directory. (gdb) bt

0 0x00007ffff7ce1887 in __GI___libc_write (fd=4, buf=0x584020, nbytes=155)

at ../sysdeps/unix/sysv/linux/write.c:26

1 0x00007ffff7c57eed in _IO_new_file_write (f=0x478d10, data=0x584020, n=155) at ./libio/fileops.c:1180

2 0x00007ffff7c599e1 in new_do_write (to_do=155,

data=0x584020 "insert ( 752 overlay 11 10 12 1 0 0 0 f f f t f i s 3 0 17 '17 17 23' _null_ _null_ _null_ _null_ _null_ byteaoverlay_no_len _null_ _null_ _null_ _null_ )\n\n _null_ )\n\n )\n_ )\n _null_ )\n )\n",

fp=0x478d10) at ./libio/libioP.h:947

3 _IO_new_do_write (to_do=155,

data=0x584020 "insert ( 752 overlay 11 10 12 1 0 0 0 f f f t f i s 3 0 17 '17 17 23' _null_ _null_ _null_ _null_ _null_ byteaoverlay_no_len _null_ _null_ _null_ _null_ )\n\n _null_ )\n\n )\n_ )\n _null_ )\n )\n",

fp=0x478d10) at ./libio/fileops.c:425

4 _IO_new_do_write (fp=0x478d10,

data=0x584020 "insert ( 752 overlay 11 10 12 1 0 0 0 f f f t f i s 3 0 17 '17 17 23' _null_ _null_ _null_ _null_ _null_ byteaoverlay_no_len _null_ _null_ _null_ _null_ )\n\n _null_ )\n\n )\n_ )\n _null_ )\n )\n",

to_do=155) at ./libio/fileops.c:422

5 0x00007ffff7c574d8 in _IO_new_file_sync (fp=0x478d10) at ./libio/fileops.c:798

6 0x00007ffff7c4c1aa in GIIO_fflush (fp=0x478d10) at ./libio/libioP.h:947

7 0x0000000000410f12 in bootstrap_template1 () at initdb.c:331

8 0x0000000000409764 in initialize_data_directory () at initdb.c:3029

9 0x0000000000407bb8 in main () at initdb.c:3430

Perhaps a builtin function that doesn't work as expected ?

[stormalf]

it was caused by incorrect extended assembly generated by chibicc that causes a segmentation fault :

0x000000000147addd in pg_atomic_compare_exchange_u32_impl () at ../../../../src/include/port/atomics/arch-x86.h:171 171 asm volatile( (gdb) bt

0 0x000000000147addd in pg_atomic_compare_exchange_u32_impl ()

at ../../../../src/include/port/atomics/arch-x86.h:171

1 0x000000000147ab07 in pg_atomic_compare_exchange_u32 () at ../../../../src/include/port/atomics.h:355

2 0x000000000147614f in LWLockAttemptLock () at lwlock.c:827

3 0x000000000147a5b0 in LWLockAcquire () at lwlock.c:1233

4 0x0000000001440b31 in ShmemInitStruct () at shmem.c:392

5 0x0000000001441684 in ShmemInitHash () at shmem.c:353

6 0x00000000014417b4 in InitShmemIndex () at shmem.c:298

7 0x0000000001413000 in CreateOrAttachShmemStructs () at ipci.c:273

8 0x000000000141388b in CreateSharedMemoryAndSemaphores () at ipci.c:234

9 0x00000000007ef8ba in BootstrapModeMain () at bootstrap.c:313

10 0x0000000000ef5d2f in main () at main.c:184

[stormalf]

New segmentation fault after fixing the issue with atomic compare exchange : Thread 5.1 "postgres" received signal SIGSEGV, Segmentation fault. [Switching to Thread 0x7ffff7c7fb80 (LWP 532639)] 0x0000000000000000 in ?? () (gdb) bt

0 0x0000000000000000 in ?? ()

1 0x000000000156e93d in StatsShmemInit () at pgstat_shmem.c:234

2 0x0000000001412909 in CreateOrAttachShmemStructs () at ipci.c:338

3 0x0000000001412e31 in CreateSharedMemoryAndSemaphores () at ipci.c:234

4 0x00000000007ee8f1 in BootstrapModeMain () at bootstrap.c:313

5 0x0000000000ef4e07 in main () at main.c:184

[stormalf]

the issue was in pg_stat.c compiled by chibicc the pointer function defined in a struct is always null. Compiling the same program pg_stat.c with gcc, the function pointer in a struct has a valid address!

[stormalf]

After investigation the root cause was the offsetof management by chibicc that was not correct. With help of chatgpt, rewritten the builtin_offsetof parsing to manage all the case we can find in postgres. Will be solved in 1.0.23.