[not-matthias]

TLDR: Since #53 glibc reported a 4 KiB main thread stack at the default 8 MiB limit, which breaks CPython 3.14 (BufferError at startup). The startup strings now live in their own mapping so only 64 KiB sits above SP in the stack mapping; glibc reports RLIMIT_STACK - 64 KiB (native: - ~100 KiB), SP stays fixed.

Layout (initimg-linux.c)

• pthread_getattr_np computes RLIMIT_STACK - (end of VMA containing SP - SP). The fixed 8 MiB region from fix(initimg): stabilize Linux startup stack layout #53 was all inside that VMA.
• Strings grow up from a fixed strings_start in a separate anonymous mapping; argc/argv/envp/auxv tables (64 KiB granule) stay above SP. One reserved, unmapped page between the two keeps the kernel from merging the VMAs.
• Tables larger than 64 KiB take whole granules from the string mapping, so SP never moves. Hard error only when tables + page + strings exceed 8 MiB, as before.
• Reservation size szB + 8 MiB and growth-on-fault are unchanged.

m_main.c

• Main thread client_stack_szB is now sized from the segment holding SP instead of clstack_max_size, so the overflow message and setrlimit comparison no longer include the string mapping.

Test

• none/tests/linux/main_stack_attr: reported range must cover SP and be within 1 MiB of an 8 MiB limit; then re-execs itself with 20000 args under --trace-children=yes to cover the granule fallback. Fails on master (SP ... outside reported stack [.., ..+4 KiB)), passes natively.

Measured (8 MiB ulimit -s, x86_64)

case	master	this PR	native
default env	4 KiB	8132 KiB	8096 KiB
600 KB env	4 KiB	8132 KiB, same SP
20000 args	4 KiB	8004 KiB, same SP
Python 3.14 under callgrind	BufferError	OK	OK

none/tests/linux: 18/19 pass; getdents_filter fails on master too. memcheck/tests/linux: 27/28, sys-execveat fails on master too. memcheck, helgrind and drd run a program that reads all of argv[] and environ without reports.

Review notes

• argv/envp strings are no longer inside the [stack] VMA. Visible only through /proc/self/maps; in-place argv rewriting still works (mapping is RW).
• Pre-existing and unchanged: ulimit -s 64M reports 64 MiB while usable depth is clamped to 16 MiB; runtime setrlimit does not grow the reservation.

COD-4117

[codspeed]

Merging this PR will not alter performance

✅ 84 untouched benchmarks
⏩ 60 skipped benchmarks¹

---

_{Comparing cod-4117-main-stack-reported-size (ab3ed5c) with master (d034e80)}

Footnotes

1. 60 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The change affects main-stack setup for Linux programs launched under Valgrind.

Fix the Darwin stack-size calculation before merging; it breaks DRD startup.

Findings

1.  DRD startup breaks on Darwin ▶

Fix with agent prompt

### Issue 1
coregrind/m_main.c:2167-2168
On Darwin, `am_startup` returns `(Addr)-1`, and `ii_create_image` replaces `clstack_end` only in its local copy. The global `the_iicii.clstack_end` stays at `(Addr)-1`, so this unsigned subtraction wraps `stack_szB` to `clstack_max_size + seg->end + 1`.

With DRD's default settings, that oversized value makes the stack suppression range start above its end. `start_suppression` then fails its assertion before client code runs. Apply this adjustment only on Linux, or return the actual stack top from `ii_create_image`.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Moves Linux startup strings into a separate mapping so glibc reports a main-thread stack size close to RLIMIT_STACK.

• Linux programs get a main-thread stack report close to the stack limit.

Known and unchanged: not-matthias identified large inherited stack limits and runtime reservation growth as pre-existing limitations. not-matthias also explicitly noted that moving startup strings outside the stack mapping is intentional.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart BT
  R["Reserved space for stack growth"] --> S["Runtime stack below fixed SP"]
  S --> T["Startup pointer tables: 64 KiB granules"]
  T --> G["One reserved, unmapped page"]
  G --> A["Separate mapping for startup strings"]

Loading

_{Reviews (1) · Last reviewed commit: "fix(initimg): keep glibc's reported main..." · Reviewed by Greptile}

[GuillaumeLagrange]

lgtm