On 2026-05-18 21:50, Aurelien Jarno wrote: > control: tag -1 + patch > control: forwarded -1 https://github.com/microsoft/mimalloc/pull/1296 > > Hi: > > On 2026-05-15 00:06, Aurelien Jarno wrote: > > control: retitle -1 mimalloc: riscv64: build-time MMU detection causes > > crashes on SV48 systems > > control: reassign -1 src:mimalloc > > control: found -1 3.0.1+ds-1 > > > > Hi, > > > > On 2026-05-14 22:09, Aurelien Jarno wrote: > > > On 2026-05-14 13:02, Aurelien Jarno wrote: > > > > Hi, > > > > > > > > On 2026-05-14 01:02, Peter Green wrote: > > > > > found 1136400 0.1.40-1 > > > > > thanks > > > > > > > > > > I believe you've quoted the wrong part of the log, > > > > > quoting output from a test that is already marked as "broken". > > > > > > > > > > I belive the actual failure is. > > > > > > > > > > 142s process didn't exit successfully: `CARGO=/usr/bin/cargo > > > > > CARGO_MANIFEST_DIR=/usr/share/cargo/registry/libmimalloc-sys-0.1.40 > > > > > CARGO_MANIFEST_PATH=/usr/share/cargo/registry/libmimalloc-sys-0.1.40/Cargo.toml > > > > > CARGO_PKG_AUTHORS='Octavian Oncescu <[email protected]>' > > > > > CARGO_PKG_DESCRIPTION='Sys crate wrapping the mimalloc allocator' > > > > > CARGO_PKG_HOMEPAGE='' CARGO_PKG_LICENSE=MIT CARGO_PKG_LICENSE_FILE='' > > > > > CARGO_PKG_NAME=libmimalloc-sys CARGO_PKG_README='' > > > > > CARGO_PKG_REPOSITORY='https://github.com/purpleprotocol/mimalloc_rust/tree/master/libmimalloc-sys' > > > > > CARGO_PKG_RUST_VERSION='' CARGO_PKG_VERSION=0.1.40 > > > > > CARGO_PKG_VERSION_MAJOR=0 CARGO_PKG_VERSION_MINOR=1 > > > > > CARGO_PKG_VERSION_PATCH=40 CARGO_PKG_VERSION_PRE='' > > > > > LD_LIBRARY_PATH='/tmp/tmp.YhgzhQ6wzU/target/riscv64gc-unknown-linux-gnu/debug:/tmp/tmp.YhgzhQ6wzU/target/riscv64gc-unknown-linux-gnu/debug/deps:/usr/lib/rustlib/riscv64gc-unknown-linux-gnu/lib' > > > > > > > > > > OUT_DIR=/tmp/tmp.YhgzhQ6wzU/target/riscv64gc-unknown-linux-gnu/debug/build/libmimalloc-sys-4d9f02f6027bf000/out > > > > > > > > > > /tmp/tmp.YhgzhQ6wzU/target/riscv64gc-unknown-linux-gnu/debug/deps/libmimalloc_sys-ceadfc72acec09f8` > > > > > (signal: 11, SIGSEGV: invalid memory reference) > > > > > 142s autopkgtest [05:06:41]: test librust-libmimalloc-sys-dev:arena: > > > > > -----------------------] > > > > > > > > I am personally unable to reproduce the issue on two different boards, > > > > using JH7110 and SpacemiT K1 CPU. That said after retrying the > > > > autopkgtest a few time on debci (it runs fast), I have noticed the > > > > following pattern: > > > > - Tests pass on debci-10 running Linux 6.12.86+deb13-riscv64 > > > > - Tests fail on debci-31, debci-32 and debci-33 running Linux > > > > 6.6.88-win2030 > > > > > > > > Therefore I believe it could be a kernel or a hardware issue. I guess > > > > the boards running the 6.6.88-win2030 kernel are the P550 ones. > > > > > > I am able to reproduce the issue on a P550 board, either running kernel > > > 6.6.77, 6.12.33 or 6.12.78. I am also able to reproduce it on the same > > > hardware in a VM running kernel 6.12.86+deb13-riscv64. > > > > > > On the other hand I am unable to reproduce it on the riscv64 porterbox > > > (ricci.d.o). > > > > > > I therefore believe there is something fishy with the hardware that > > > needs more investigation. > > > > I finally found the issue. It is in mimalloc, which contains the > > following build time check: > > > > | # Check /proc/cpuinfo for an SV39 MMU and limit the virtual address bits. > > | # (this will skip the aligned hinting in that case. Issue #939, #949) > > | if (EXISTS /proc/cpuinfo) > > | file(STRINGS /proc/cpuinfo mi_sv39_mmu REGEX "^mmu[ \t]+:[ \t]+sv39$") > > | if (mi_sv39_mmu) > > | MESSAGE( STATUS "Set virtual address bits to 39 (SV39 MMU detected)" ) > > | list(APPEND mi_defines MI_DEFAULT_VIRTUAL_ADDRESS_BITS=39) > > | endif() > > | endif() > > > > Our build daemons are using a sv39 nmu: > > > > | aurel32@ricci:~$ grep ^mmu /proc/cpuinfo | head -n1 > > | mmu : sv39 > > > > While the P550 defaults to a sv48 nmu: > > | aurel32@p550:~$ grep ^mmu /proc/cpuinfo | head -n1 > > | mmu : sv48 > > > > Therefore libmimalloc.so.3 is built with a 39-bit virtual address space > > support. When the binary is run on a sv48 system, things explodes. > > Rebuilding libmimalloc.so.3 on the P550 enables sv48 support and things > > work again, even on a sv39 mmu. > > > > It's not clear to me if there is a drawback to always enable sv48 > > support, if yes I guess a runtime detection should be implemented. > > Anyway for now reassigning the bug to the correct package. > > I have send a patch upstream: > https://github.com/microsoft/mimalloc/pull/1296
The upstream patch got merged, but doesn't apply cleanly. Please find attached a patch to drop into debian/patches. Regards Aurelien -- Aurelien Jarno GPG: 4096R/1DDD8C9B [email protected] http://aurel32.net
commit 7abd6bfcf2c56a9729dc2204aba8e3e0369a5910 Author: Aurelien Jarno <[email protected]> Date: Mon May 25 11:41:28 2026 +0200 RISC-V: detect virtual address space at runtime using hwprobe The virtual address space on RISC-V is currently detected at build time by parsing /proc/cpuinfo. This works when the binary runs on the same hardware it was built, however running such a binary on a system with a larger address space just causes a segmentation fault. Replace the build time check with a runtime detection using the hwprobe interface (available since Linux 6.11, commit c9b8cd139c1d "riscv: hwprobe export highest virtual userspace address"), with a fallback to parsing /proc/cpuinfo on older kernels. diff --git a/CMakeLists.txt b/CMakeLists.txt index c33c3046..a23169a5 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -475,13 +475,14 @@ else() # list(APPEND mi_defines MI_WIN_INIT_USE_CRT_TLS=1) endif() -# Check /proc/cpuinfo for an SV39 MMU and limit the virtual address bits. -# (this will skip the aligned hinting in that case. Issue #939, #949) -if (EXISTS /proc/cpuinfo) - file(STRINGS /proc/cpuinfo mi_sv39_mmu REGEX "^mmu[ \t]+:[ \t]+sv39$") - if (mi_sv39_mmu) - MESSAGE( STATUS "Set virtual address bits to 39 (SV39 MMU detected)" ) - list(APPEND mi_defines MI_DEFAULT_VIRTUAL_ADDRESS_BITS=39) +if(MI_ARCH MATCHES "riscv") + CHECK_INCLUDE_FILES("asm/hwprobe.h;" MI_HAS_ASM_HWPROBEH) + if (MI_HAS_ASM_HWPROBEH) + list(APPEND mi_defines MI_HAS_ASM_HWPROBEH=1) + endif() + CHECK_INCLUDE_FILES("sys/hwprobe.h;" MI_HAS_SYS_HWPROBEH) + if (MI_HAS_SYS_HWPROBEH) + list(APPEND mi_defines MI_HAS_SYS_HWPROBEH=1) endif() endif() diff --git a/src/prim/unix/prim.c b/src/prim/unix/prim.c index 7781a37a..fc08712d 100644 --- a/src/prim/unix/prim.c +++ b/src/prim/unix/prim.c @@ -41,6 +41,13 @@ terms of the MIT license. A copy of the license can be found in the file #else #include <sys/mman.h> #endif + #if defined(__riscv) || defined(_M_RISCV) + #if defined(MI_HAS_SYS_HWPROBEH) + #include <sys/hwprobe.h> + #elif defined(MI_HAS_ASM_HWPROBEH) + #include <asm/hwprobe.h> + #endif + #endif #elif defined(__APPLE__) #include <AvailabilityMacros.h> #include <TargetConditionals.h> @@ -206,6 +213,52 @@ static void unix_detect_physical_memory( size_t page_size, size_t* physical_memo #endif } +// Detect the virtual address bits (currently Linux/RISC-V only) +static size_t unix_detect_virtual_address_bits(void) +{ + #if defined(__riscv) || defined(_M_RISCV) + #if defined(RISCV_HWPROBE_KEY_HIGHEST_VIRT_ADDRESS) + struct riscv_hwprobe probe = { + .key = RISCV_HWPROBE_KEY_HIGHEST_VIRT_ADDRESS, + }; + + // Prefer the GNU libc interface if available, as it can also use the VDSO + #if defined(MI_HAS_SYS_HWPROBEH) + if (__riscv_hwprobe(&probe, 1, 0, NULL, 0) == 0) { + #else + if (syscall(__NR_riscv_hwprobe, &probe, 1, 0, NULL, 0) == 0) { + #endif + // If a key is unknown to the kernel, its key field will be cleared to -1. + if (probe.key != -1) { + return MI_SIZE_BITS - mi_clz((uintptr_t)probe.value); + } + } + #endif + + // Fallback to checking /proc/cpuinfo for older kernels + const int fd = mi_prim_open("/proc/cpuinfo", O_RDONLY); + if (fd >= 0) { + char buf[2048]; + ssize_t nread = mi_prim_read(fd, &buf, sizeof(buf)); + mi_prim_close(fd); + if ((nread >= 1) && (nread <= (ssize_t)sizeof(buf))) { + if (_mi_strnstr(buf, nread, "sv39")) { + return 39; + } + if (_mi_strnstr(buf, nread, "sv48")) { + return 48; + } + if (_mi_strnstr(buf, nread, "sv57")) { + return 57; + } + } + } + #endif + + // default: MI_MAX_VABITS + return MI_MAX_VABITS; +} + void _mi_prim_mem_init( mi_os_mem_config_t* config ) { long psize = sysconf(_SC_PAGESIZE); @@ -219,6 +272,7 @@ void _mi_prim_mem_init( mi_os_mem_config_t* config ) config->has_partial_free = true; // mmap can free in parts config->has_virtual_reserve = true; // todo: check if this true for NetBSD? (for anonymous mmap with PROT_NONE) config->has_transparent_huge_pages = unix_detect_thp(); + config->virtual_address_bits = unix_detect_virtual_address_bits(); // disable transparent huge pages for this process? #if (defined(__linux__) || defined(__ANDROID__)) && defined(PR_GET_THP_DISABLE)
