crtl is not large-file safe at ILP32: off_t is 32 bits, so >2GB files are wrong

Type: feature (Track A — lib/crtl/include/sys/_types.h, lib/crtl/include/fcntl.h, lib/crtl/src/{fcntl.c,unistd.c,sys/mman.c})
prio: 45
Status: open

The measurement

gcc -m32 -D_FILE_OFFSET_BITS=64 against pxx --target=i386, same source, _GNU_SOURCE set explicitly in the probe (feature-test macros change what the headers declare — see frankD's note on the crtl-impl macro environment):

           glibc-m32   pxx-i386
off_t       8           4
sigset_t  128          64

dev_t, ino_t and blkcnt_t were the other three rows and are fixed — bug-a-stat-returns-st-dev-and-st-rdev-in-the-kernel-internal-encoding.

rlim_t is a fourth member of the same LFS family and belongs to this ticket rather than to a separate one. Measured: glibc -m32 gives sizeof(rlim_t) 4 without LFS and 8 with _FILE_OFFSET_BITS=64; crtl's unsigned long is 4, so it matches the NON-LFS shape, the same way off_t does. It is not a defect today — <sys/resource.h> deliberately declares no functions (the PAL has no getrlimit entry), so the only path that fills a struct rlimit is a hosted glibc link, which resolves the non-LFS getrlimit symbol and agrees with the 4-byte fields. It stops agreeing the moment the rest of this ticket lands, so rlim_t and RLIM_INFINITY move with off_t or the hosted path silently starts reading two fields as one.

Why off_t is not a one-line widening

lib/crtl/src/fcntl.c's header comment is correct today and would become false the moment off_t moves: the struct flock a caller builds matches the kernel's native layout on each target because off_t == long. Widening it alone would make every F_SETLK/F_GETLK at ILP32 pass a struct the kernel reads with the wrong field offsets — sqlite's whole locking protocol, failing silently rather than erroring.

So the change is a group:

struct flock64 and F_GETLK64/F_SETLK64/F_SETLKW64 (12/13/14 on i386)
lseek → _llseek (offset split high/low, result through a pointer)
pread/pwrite → pread64/pwrite64, offset as a register pair with the target's argument-alignment rule (arm32/riscv32 differ from i386)
mmap → mmap2, offset in pages
off_t in struct stat and every crtl signature that carries one

The PAL side is already Int64 throughout, so nothing above the syscall layer needs widening — this is entirely inside lib/crtl.

Why it is worth doing

The 32-bit targets are the point of the cross axis, and a >2GB file is ordinary — sqlite databases, a DOSBox disk image, a busybox cp. The failure mode is a wrong offset, not a refusal.

sigset_t

64 bits where glibc has 128. Internally consistent (crtl's own sigprocmask/sigaction agree with each other) and never measured against the kernel's _NSIG/8 expectation. Cheap to check once someone is in this file: sigprocmask with a signal above 32 is the row that would tell.