Use /proc/self/maps when available instead of std::env::current_exe (#488)

pnkfelix · web-flow · commit e1d31f2f27f7 · 2022-10-25T11:26:02.000-05:00
* standalone proc/self/maps parsing code to try instead of less reliable current_exe.

(updated to only define the parse_running_mmaps module in cases where we are
also defining libs_dl_iterate_phdr, since that is the only place it is currently
needed.)

* checkpoint regression test demonstrating failure of rust issue 101913.

* fix code to work for cargo test --no-default-features .

* fix test to ignore itself when it cannot invoke readelf.

* Conditionalize 64-bit test. Add test that works for 32-bit targets.

* placate cargo fmt.

* switch to pure `Read` rather than `BufRead` based implementation, to save a bit on code size.

ls -sk before:
1052 target/release/libbacktrace.rlib

ls -sk after:
1044 target/release/libbacktrace.rlib

* Avoid using `str::split_once` method, since the MSRV is 1.42 but that was added in 1.52.

* placate cargo fmt.

Co-authored-by: Felix Klock &lt;pnkfelix@amazon.com&gt;
diff --git a/Cargo.toml b/Cargo.toml
@@ -130,3 +130,8 @@ edition = '2018'
 name = "concurrent-panics"
 required-features = ["std"]
 harness = false
+
+[[test]]
+name = "current-exe-mismatch"
+required-features = ["std"]
+harness = false
diff --git a/src/symbolize/gimli.rs b/src/symbolize/gimli.rs
@@ -184,6 +184,8 @@ cfg_if::cfg_if! {
     ))] {
         mod libs_dl_iterate_phdr;
         use libs_dl_iterate_phdr::native_libraries;
+        #[path = "gimli/parse_running_mmaps_unix.rs"]
+        mod parse_running_mmaps;
     } else if #[cfg(target_env = "libnx")] {
         mod libs_libnx;
         use libs_libnx::native_libraries;
diff --git a/src/symbolize/gimli/libs_dl_iterate_phdr.rs b/src/symbolize/gimli/libs_dl_iterate_phdr.rs
@@ -17,6 +17,20 @@ pub(super) fn native_libraries() -> Vec<Library> {
     return ret;
 }
 
+fn infer_current_exe(base_addr: usize) -> OsString {
+    if let Ok(entries) = super::parse_running_mmaps::parse_maps() {
+        let opt_path = entries
+            .iter()
+            .find(|e| e.ip_matches(base_addr) && e.pathname().len() > 0)
+            .map(|e| e.pathname())
+            .cloned();
+        if let Some(path) = opt_path {
+            return path;
+        }
+    }
+    env::current_exe().map(|e| e.into()).unwrap_or_default()
+}
+
 // `info` should be a valid pointers.
 // `vec` should be a valid pointer to a `std::Vec`.
 unsafe extern "C" fn callback(
@@ -28,8 +42,12 @@ unsafe extern "C" fn callback(
     let libs = &mut *(vec as *mut Vec<Library>);
     let is_main_prog = info.dlpi_name.is_null() || *info.dlpi_name == 0;
     let name = if is_main_prog {
+        // The man page for dl_iterate_phdr says that the first object visited by
+        // callback is the main program; so the first time we encounter a
+        // nameless entry, we can assume its the main program and try to infer its path.
+        // After that, we cannot continue that assumption, and we use an empty string.
         if libs.is_empty() {
-            env::current_exe().map(|e| e.into()).unwrap_or_default()
+            infer_current_exe(info.dlpi_addr as usize)
         } else {
             OsString::new()
         }
diff --git a/src/symbolize/gimli/parse_running_mmaps_unix.rs b/src/symbolize/gimli/parse_running_mmaps_unix.rs
@@ -0,0 +1,242 @@
+// Note: This file is only currently used on targets that call out to the code
+// in `mod libs_dl_iterate_phdr` (e.g. linux, freebsd, ...); it may be more
+// general purpose, but it hasn't been tested elsewhere.
+
+use super::mystd::fs::File;
+use super::mystd::io::Read;
+use super::mystd::str::FromStr;
+use super::{OsString, String, Vec};
+
+#[derive(PartialEq, Eq, Debug)]
+pub(super) struct MapsEntry {
+    /// start (inclusive) and limit (exclusive) of address range.
+    address: (usize, usize),
+    /// The perms field are the permissions for the entry
+    ///
+    /// r = read
+    /// w = write
+    /// x = execute
+    /// s = shared
+    /// p = private (copy on write)
+    perms: [char; 4],
+    /// Offset into the file (or "whatever").
+    offset: usize,
+    /// device (major, minor)
+    dev: (usize, usize),
+    /// inode on the device. 0 indicates that no inode is associated with the memory region (e.g. uninitalized data aka BSS).
+    inode: usize,
+    /// Usually the file backing the mapping.
+    ///
+    /// Note: The man page for proc includes a note about "coordination" by
+    /// using readelf to see the Offset field in ELF program headers. pnkfelix
+    /// is not yet sure if that is intended to be a comment on pathname, or what
+    /// form/purpose such coordination is meant to have.
+    ///
+    /// There are also some pseudo-paths:
+    /// "[stack]": The initial process's (aka main thread's) stack.
+    /// "[stack:<tid>]": a specific thread's stack. (This was only present for a limited range of Linux verisons; it was determined to be too expensive to provide.)
+    /// "[vdso]": Virtual dynamically linked shared object
+    /// "[heap]": The process's heap
+    ///
+    /// The pathname can be blank, which means it is an anonymous mapping
+    /// obtained via mmap.
+    ///
+    /// Newlines in pathname are replaced with an octal escape sequence.
+    ///
+    /// The pathname may have "(deleted)" appended onto it if the file-backed
+    /// path has been deleted.
+    ///
+    /// Note that modifications like the latter two indicated above imply that
+    /// in general the pathname may be ambiguous. (I.e. you cannot tell if the
+    /// denoted filename actually ended with the text "(deleted)", or if that
+    /// was added by the maps rendering.
+    pathname: OsString,
+}
+
+pub(super) fn parse_maps() -> Result<Vec<MapsEntry>, &'static str> {
+    let mut v = Vec::new();
+    let mut proc_self_maps =
+        File::open("/proc/self/maps").map_err(|_| "Couldn't open /proc/self/maps")?;
+    let mut buf = String::new();
+    let _bytes_read = proc_self_maps
+        .read_to_string(&mut buf)
+        .map_err(|_| "Couldn't read /proc/self/maps")?;
+    for line in buf.lines() {
+        v.push(line.parse()?);
+    }
+
+    Ok(v)
+}
+
+impl MapsEntry {
+    pub(super) fn pathname(&self) -> &OsString {
+        &self.pathname
+    }
+
+    pub(super) fn ip_matches(&self, ip: usize) -> bool {
+        self.address.0 <= ip && ip < self.address.1
+    }
+}
+
+impl FromStr for MapsEntry {
+    type Err = &'static str;
+
+    // Format: address perms offset dev inode pathname
+    // e.g.: "ffffffffff600000-ffffffffff601000 --xp 00000000 00:00 0                  [vsyscall]"
+    // e.g.: "7f5985f46000-7f5985f48000 rw-p 00039000 103:06 76021795                  /usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2"
+    // e.g.: "35b1a21000-35b1a22000 rw-p 00000000 00:00 0"
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        let mut parts = s
+            .split(' ') // space-separated fields
+            .filter(|s| s.len() > 0); // multiple spaces implies empty strings that need to be skipped.
+        let range_str = parts.next().ok_or("Couldn't find address")?;
+        let perms_str = parts.next().ok_or("Couldn't find permissions")?;
+        let offset_str = parts.next().ok_or("Couldn't find offset")?;
+        let dev_str = parts.next().ok_or("Couldn't find dev")?;
+        let inode_str = parts.next().ok_or("Couldn't find inode")?;
+        let pathname_str = parts.next().unwrap_or(""); // pathname may be omitted.
+
+        let hex = |s| usize::from_str_radix(s, 16).map_err(|_| "Couldn't parse hex number");
+        let address = {
+            // This could use `range_str.split_once('-')` once the MSRV passes 1.52.
+            if let Some(idx) = range_str.find('-') {
+                let (start, rest) = range_str.split_at(idx);
+                let (_div, limit) = rest.split_at(1);
+                (hex(start)?, hex(limit)?)
+            } else {
+                return Err("Couldn't parse address range");
+            }
+        };
+        let perms: [char; 4] = {
+            let mut chars = perms_str.chars();
+            let mut c = || chars.next().ok_or("insufficient perms");
+            let perms = [c()?, c()?, c()?, c()?];
+            if chars.next().is_some() {
+                return Err("too many perms");
+            }
+            perms
+        };
+        let offset = hex(offset_str)?;
+        let dev = {
+            // This could use `dev_str.split_once(':')` once the MSRV passes 1.52.
+            if let Some(idx) = dev_str.find(':') {
+                let (major, rest) = dev_str.split_at(idx);
+                let (_div, minor) = rest.split_at(1);
+                (hex(major)?, hex(minor)?)
+            } else {
+                return Err("Couldn't parse dev")?;
+            }
+        };
+        let inode = hex(inode_str)?;
+        let pathname = pathname_str.into();
+
+        Ok(MapsEntry {
+            address,
+            perms,
+            offset,
+            dev,
+            inode,
+            pathname,
+        })
+    }
+}
+
+// Make sure we can parse 64-bit sample output if we're on a 64-bit target.
+#[cfg(target_pointer_width = "64")]
+#[test]
+fn check_maps_entry_parsing_64bit() {
+    assert_eq!(
+        "ffffffffff600000-ffffffffff601000 --xp 00000000 00:00 0                  \
+                [vsyscall]"
+            .parse::<MapsEntry>()
+            .unwrap(),
+        MapsEntry {
+            address: (0xffffffffff600000, 0xffffffffff601000),
+            perms: ['-', '-', 'x', 'p'],
+            offset: 0x00000000,
+            dev: (0x00, 0x00),
+            inode: 0x0,
+            pathname: "[vsyscall]".into(),
+        }
+    );
+
+    assert_eq!(
+        "7f5985f46000-7f5985f48000 rw-p 00039000 103:06 76021795                  \
+                /usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2"
+            .parse::<MapsEntry>()
+            .unwrap(),
+        MapsEntry {
+            address: (0x7f5985f46000, 0x7f5985f48000),
+            perms: ['r', 'w', '-', 'p'],
+            offset: 0x00039000,
+            dev: (0x103, 0x06),
+            inode: 0x76021795,
+            pathname: "/usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2".into(),
+        }
+    );
+    assert_eq!(
+        "35b1a21000-35b1a22000 rw-p 00000000 00:00 0"
+            .parse::<MapsEntry>()
+            .unwrap(),
+        MapsEntry {
+            address: (0x35b1a21000, 0x35b1a22000),
+            perms: ['r', 'w', '-', 'p'],
+            offset: 0x00000000,
+            dev: (0x00, 0x00),
+            inode: 0x0,
+            pathname: Default::default(),
+        }
+    );
+}
+
+// (This output was taken from a 32-bit machine, but will work on any target)
+#[test]
+fn check_maps_entry_parsing_32bit() {
+    /* Example snippet of output:
+    08056000-08077000 rw-p 00000000 00:00 0          [heap]
+    b7c79000-b7e02000 r--p 00000000 08:01 60662705   /usr/lib/locale/locale-archive
+    b7e02000-b7e03000 rw-p 00000000 00:00 0
+        */
+    assert_eq!(
+        "08056000-08077000 rw-p 00000000 00:00 0          \
+                [heap]"
+            .parse::<MapsEntry>()
+            .unwrap(),
+        MapsEntry {
+            address: (0x08056000, 0x08077000),
+            perms: ['r', 'w', '-', 'p'],
+            offset: 0x00000000,
+            dev: (0x00, 0x00),
+            inode: 0x0,
+            pathname: "[heap]".into(),
+        }
+    );
+
+    assert_eq!(
+        "b7c79000-b7e02000 r--p 00000000 08:01 60662705   \
+                /usr/lib/locale/locale-archive"
+            .parse::<MapsEntry>()
+            .unwrap(),
+        MapsEntry {
+            address: (0xb7c79000, 0xb7e02000),
+            perms: ['r', '-', '-', 'p'],
+            offset: 0x00000000,
+            dev: (0x08, 0x01),
+            inode: 0x60662705,
+            pathname: "/usr/lib/locale/locale-archive".into(),
+        }
+    );
+    assert_eq!(
+        "b7e02000-b7e03000 rw-p 00000000 00:00 0"
+            .parse::<MapsEntry>()
+            .unwrap(),
+        MapsEntry {
+            address: (0xb7e02000, 0xb7e03000),
+            perms: ['r', 'w', '-', 'p'],
+            offset: 0x00000000,
+            dev: (0x00, 0x00),
+            inode: 0x0,
+            pathname: Default::default(),
+        }
+    );
+}
diff --git a/tests/common/mod.rs b/tests/common/mod.rs
@@ -0,0 +1,14 @@
+/// Some tests only make sense in contexts where they can re-exec the test
+/// itself. Not all contexts support this, so you can call this method to find
+/// out which case you are in.
+pub fn cannot_reexec_the_test() -> bool {
+    // These run in docker containers on CI where they can't re-exec the test,
+    // so just skip these for CI. No other reason this can't run on those
+    // platforms though.
+    // Miri does not have support for re-execing a file
+    cfg!(unix)
+        && (cfg!(target_arch = "arm")
+            || cfg!(target_arch = "aarch64")
+            || cfg!(target_arch = "s390x"))
+        || cfg!(miri)
+}
diff --git a/tests/concurrent-panics.rs b/tests/concurrent-panics.rs
@@ -9,17 +9,11 @@ const PANICS: usize = 100;
 const THREADS: usize = 8;
 const VAR: &str = "__THE_TEST_YOU_ARE_LUKE";
 
+mod common;
+
 fn main() {
-    // These run in docker containers on CI where they can't re-exec the test,
-    // so just skip these for CI. No other reason this can't run on those
-    // platforms though.
-    // Miri does not have support for re-execing a file
-    if cfg!(unix)
-        && (cfg!(target_arch = "arm")
-            || cfg!(target_arch = "aarch64")
-            || cfg!(target_arch = "s390x"))
-        || cfg!(miri)
-    {
+    // If we cannot re-exec this test, there's no point in trying to do it.
+    if common::cannot_reexec_the_test() {
         println!("test result: ok");
         return;
     }
diff --git a/tests/current-exe-mismatch.rs b/tests/current-exe-mismatch.rs
