Rewrite mmap emulation as page-granular region allocator (WARP-69)

[Arshia001]

Summary

Replaces the malloc-veneer mmap emulation in libc-bottom-half/mman/mman.c with a side-table region allocator that serves all mappings in whole 64 KiB wasm pages drawn directly from sbrk (memory.grow).

Design and rationale: WARP-69. Root fix for ECO-263.

Key change: anonymous mappings are no longer memset. Pages fresh from memory.grow are guaranteed zero by the wasm spec and are handed out untouched, so the host commits them lazily on first write instead of all-upfront. Only pages recycled from a previous munmap are zeroed on reuse.

Conformance fixes over the old shim

• mmap returns page-aligned addresses (side table instead of an inline header before the mapping)
• Partial munmap: any whole-page subrange can be unmapped, including splitting a mapping in the middle; freed pages coalesce and are recycled for future mappings
• msync no longer rejects writable mappings (the PROT_WRITE check was inverted, so MAP_SHARED file writeback silently never happened — a data-loss bug); partial unmap writes back just the unmapped subrange with correct (fd, offset) on surviving fragments
• Writes past EOF within the last page are not written back, and the tail past EOF reads as zero, per POSIX
• MAP_NORESERVE accepted (commit is lazy anyway); failure paths return MAP_FAILED with a meaningful errno instead of NULL

Still unsupported without an MMU (unchanged): PROT_NONE/PROT_EXEC, MAP_FIXED, CoW private file mappings, continuous writeback.

Measurements

Under wasmer 7.2.0-rc.1, a 128 MiB anonymous mapping (the PHP opcache SHM pattern):

	old emulation	this PR
host RSS cost of the mapping	~128 MiB (fully committed at mmap time)	~0 until pages are actually written

End-to-end on PHPIX (PHP 8.1 ZTS server): idle RSS 906 → 781 MiB (−125 MiB); additionally, partial-munmap support un-blocks re-enabling Zend MM on WASIX, which measures 2.9× faster on allocation-heavy requests vs the tracked-malloc fallback.

Tests

Integration tests live in the wasmer repo (branch warp69-mman-tests, PR to follow after a prerelease of this change is pinned there): a new libc/mmap-page-granular test covering alignment, fresh/recycled zero-fill, partial munmap (including the over-allocate-and-trim pattern used by Zend MM), coalescing, file writeback via msync/munmap, and error semantics — plus the seven previously-ignored msync-*/munmap-sync-*/read-after-munmap tests, which this PR fixes (they were ignored due to the inverted msync check; four of them also used non-page-aligned file offsets, which POSIX rejects and this implementation now correctly rejects, and were fixed accordingly).

Full wasmer wasm_tests sweep against a sysroot built from this branch: no regressions (all failures identical under the old pinned sysroot and attributable to a missing local wasm32-wasip1 rustup target).

🤖 Generated with Claude Code