fix(native): fix incorrect stacktraces on Linux (ELF segment merging + DWARF unwinding)

[mujacica]

Fixes two root causes of incorrect stacktraces reported by customers on Linux:

1. ELF segment merging: /proc/pid/maps shows separate mappings per ELF LOAD segment. Without merging, base_of_image pointed to the code segment instead of the real ELF load base, causing all RVA computations to be off by the .text section offset. This broke server-side CFI unwinding.

2. DWARF-based backtrace capture: The daemon's FP-based stack walking fails on x86_64 (and any -fomit-frame-pointer build) because RBP is used as a general-purpose register. Now captures backtrace using libunwind (DWARF .eh_frame) in the signal handler and stores it in shared memory. The daemon prefers this over FP-walking when available. Also primes libunwind's cache at init time to avoid dl_iterate_phdr deadlocks.

[supervacuus]

Since this duplicates code from the Linux module finder, why not extract what you need from there?

[mujacica]

@supervacuus that is the plan - just did a quick and dirty version to share with customer to try it out. Need to extract it from Unit testing code

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

Autofix Details

Bugbot Autofix prepared fixes for both issues found in the latest run.

• ✅ Fixed: Memory leak: resolved not freed in error paths
  • Added sentry_free(resolved) to both error paths at lines 1634 and 1646 to prevent memory leak.
• ✅ Fixed: Redundant uname() syscall in system info stream
  • Removed second uname() call and moved CSD version string generation inside the first uname() success block.

Or push these changes by commenting:

@cursor push d5f42a5262

Preview (d5f42a5262)

diff --git a/src/backends/native/minidump/sentry_minidump_linux.c b/src/backends/native/minidump/sentry_minidump_linux.c
--- a/src/backends/native/minidump/sentry_minidump_linux.c
+++ b/src/backends/native/minidump/sentry_minidump_linux.c
@@ -515,17 +515,14 @@
 
     // Populate OS version from uname(), matching Crashpad behavior
     struct utsname uts;
+    char csd_version[256] = "";
     if (uname(&uts) == 0) {
         int major = 0, minor = 0, patch = 0;
         sscanf(uts.release, "%d.%d.%d", &major, &minor, &patch);
         sysinfo.major_version = (uint32_t)major;
         sysinfo.minor_version = (uint32_t)minor;
         sysinfo.build_number = (uint32_t)patch;
-    }
 
-    // Write CSD version string with full uname info, matching Crashpad
-    char csd_version[256] = "";
-    if (uname(&uts) == 0) {
         snprintf(csd_version, sizeof(csd_version), "%s %s %s %s", uts.sysname,
             uts.release, uts.version, uts.machine);
     }
@@ -1634,6 +1631,7 @@
         SENTRY_WARN("failed to write module list structure");
         sentry_free(mod_infos);
         sentry_free(module_list);
+        sentry_free(resolved);
         return -1;
     }
 
@@ -1645,6 +1643,7 @@
                     "name/CV patching");
         sentry_free(mod_infos);
         sentry_free(module_list);
+        sentry_free(resolved);
         return dir->rva ? 0 : -1;
     }

_{This Bugbot Autofix run was free. To enable autofix for future PRs, go to the Cursor dashboard.}

[mujacica]

@cursor push d5f42a5

[cursor]

ELF parsing functions duplicate existing modulefinder logic

Low Severity

compute_elf_size_from_phdrs and read_elf_soname duplicate ELF parsing patterns already present in sentry_modulefinder_linux.c (program header reading, section header traversal, etc.). The reviewer flagged this, and the author acknowledged it's a "quick and dirty version" that needs extraction. Duplicating this low-level ELF parsing increases the risk that a fix in one copy won't be applied to the other.

Additional Locations (1)

• src/backends/native/minidump/sentry_minidump_linux.c#L960-L1096

 

[jpnurmi]

Left as a future exercise

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 3 total unresolved issues (including 2 from previous reviews).

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}