Merge pull request #479 from pythonspeed/474-recent-versions-likely-h…

…ave-source-code-less-jupyter-tracebacks

Recent versions likely have source-code-less Jupyter traceback

CHANGELOG.md

@@ -1,5 +1,11 @@
 # Release notes
 
+## 2023.1.0 (2023-1-20)
+
+### Bugfixes
+
+- Fix regression where source code would sometimes be missing from flamegraphs, most notably in Jupyter profiling. ([#474](https://github.com/pythonspeed/filprofiler/issues/474))
+
 ## 2022.11.0 (2022-11-07)
 
 ### Features

filpreload/Cargo.toml

@@ -18,7 +18,7 @@ path = "../memapi"
 features = []
 
 [dependencies.pyo3]
-version = "0.17"
+version = "0.18"
 default-features = false
 
 [build-dependencies]

filpreload/src/lib.rs

@@ -2,12 +2,14 @@
 use parking_lot::Mutex;
 use pymemprofile_api::memorytracking::LineNumberInfo::LineNumber;
 use pymemprofile_api::memorytracking::{
-    AllocationTracker, CallSiteId, Callstack, FunctionId, VecFunctionLocations, PARENT_PROCESS,
+    AllocationTracker, CallSiteId, Callstack, FunctionId, IdentityCleaner, VecFunctionLocations,
+    PARENT_PROCESS,
 };
 use pymemprofile_api::oom::{InfiniteMemory, OutOfMemoryEstimator, RealMemoryInfo};
 use std::cell::RefCell;
 use std::ffi::CStr;
 use std::os::raw::{c_char, c_int, c_void};
+use std::path::Path;
 
 #[macro_use]
 extern crate lazy_static;
@@ -87,6 +89,7 @@ fn set_current_callstack(callstack: &Callstack) {
 }
 
 extern "C" {
+    fn _exit(exit_code: std::os::raw::c_int);
     fn free(address: *mut c_void);
 }
 
@@ -152,7 +155,23 @@ fn add_allocation(
 
     if oom {
         // Uh-oh, we're out of memory.
-        allocations.oom_dump();
+        eprintln!(
+            "=fil-profile= We'll try to dump out SVGs. Note that no HTML file will be written."
+        );
+        let default_path = allocations.default_path.clone();
+        // Release the lock, since dumping the flamegraph will reacquire it:
+        drop(tracker_state);
+
+        dump_to_flamegraph(
+            &default_path,
+            false,
+            "out-of-memory",
+            "Current allocations at out-of-memory time",
+            false,
+        );
+        unsafe {
+            _exit(53);
+        }
     };
     Ok(())
 }
@@ -180,11 +199,55 @@ fn reset(default_path: String) {
     tracker_state.allocations.reset(default_path);
 }
 
+fn dump_to_flamegraph(
+    path: &str,
+    peak: bool,
+    base_filename: &str,
+    title: &str,
+    to_be_post_processed: bool,
+) {
+    // In order to render the flamegraph, we want to load source code using
+    // Python's linecache. That means calling into Python, which might release
+    // the GIL, allowing another thread to run, and it will try to allocation
+    // and hit the TRACKER_STATE mutex. And now we're deadlocked. So we make
+    // sure flamegraph rendering does not require TRACKER_STATE to be locked.
+    let (allocated_bytes, flamegraph_callstacks) = {
+        let mut tracker_state = TRACKER_STATE.lock();
+        let allocations = &mut tracker_state.allocations;
+
+        // Print warning if we're missing allocations.
+        allocations.warn_on_problems(peak);
+        let allocated_bytes = if peak {
+            allocations.get_peak_allocated_bytes()
+        } else {
+            allocations.get_current_allocated_bytes()
+        };
+        let flamegraph_callstacks = allocations.combine_callstacks(peak, IdentityCleaner);
+        (allocated_bytes, flamegraph_callstacks)
+    };
+
+    eprintln!("=fil-profile= Preparing to write to {}", path);
+    let directory_path = Path::new(path);
+
+    let title = format!(
+        "{} ({:.1} MiB)",
+        title,
+        allocated_bytes as f64 / (1024.0 * 1024.0)
+    );
+    let subtitle = r#"Made with the Fil profiler. <a href="https://pythonspeed.com/fil/" style="text-decoration: underline;" target="_parent">Try it on your code!</a>"#;
+    flamegraph_callstacks.write_flamegraphs(
+        directory_path,
+        base_filename,
+        &title,
+        subtitle,
+        "bytes",
+        to_be_post_processed,
+    )
+}
+
 /// Dump all callstacks in peak memory usage to format used by flamegraph.
 fn dump_peak_to_flamegraph(path: &str) {
-    let mut tracker_state = TRACKER_STATE.lock();
-    let allocations = &mut tracker_state.allocations;
-    allocations.dump_peak_to_flamegraph(path);
+    dump_to_flamegraph(path, true, "peak-memory", "Peak Tracked Memory Usage", true);
 }
 
 #[no_mangle]
@@ -318,8 +381,8 @@ extern "C" {
     fn is_initialized() -> c_int;
 
     // Increment/decrement reentrancy counter.
-    fn fil_increment_reentrancy();
-    fn fil_decrement_reentrancy();
+    //fn fil_increment_reentrancy();
+    //fn fil_decrement_reentrancy();
 }
 
 struct FilMmapAPI;
@@ -337,7 +400,7 @@ impl pymemprofile_api::mmap::MmapAPI for FilMmapAPI {
     }
 
     fn is_initialized(&self) -> bool {
-        return unsafe { is_initialized() == 1 };
+        unsafe { is_initialized() == 1 }
     }
 }
 

filprofiler/_report.py

@@ -76,19 +76,16 @@ def render_report(output_path: str, now: datetime) -> str:
 
 <h2>Profiling result</h2>
 <div style="text-align: center;"><p><input type="button" onclick="fullScreen('#peak');" value="Full screen"> · <a href="peak-memory.svg" target="_blank"><button>Open in new window</button></a></p>
-<iframe id="peak" src="peak-memory.svg" width="100%" height="400" scrolling="auto" frameborder="0"></iframe>
+<iframe id="peak" src="peak-memory.svg" width="100%" height="700" scrolling="auto" frameborder="0"></iframe>
 </div>
 <br>
 <blockquote class="center">
-            <p style="text-align: center;"><em>Check out my other project:</em></p>
-            <h3>Find memory and performance bottlenecks in production!</h3>
-            <p>When your data pipeline is too slow in production, reproducing the problem
-            on your laptop is hard or impossible—which means identifying and fixing the problem can be tricky.</p>
-            <p>What if you knew the cause of the problem as soon as you realized it was happening?</p>
-            <p>That's how
-            <strong><a href="https://sciagraph.com/">the Sciagraph profiler</a></strong> can help you:
-            it's designed to <strong>find performance
-            and memory bottlenecks by continuously profiling in production.</strong></p></blockquote>
+            <h3>Find performance bottlenecks in your data processing jobs with the Sciagraph profiler</h3>
+            <p><strong><a href="https://sciagraph.com/">The Sciagraph profiler</a></strong> can help you
+            <strong>find performance
+            and memory bottlenecks with low overhead, so you can use it in both development and production.</strong></p>
+            <p>Unlike Fil, it includes performance profiling. Sciagraph's memory profiling uses sampling so it runs faster than Fil, but unlike Fil
+            it can't accurately profile small allocations or run natively on macOS.</p></blockquote>
 <br>
             <br>
 <div style="text-align: center;"><p><input type="button" onclick="fullScreen('#peak-reversed');" value="Full screen"> ·

memapi/Cargo.toml

@@ -2,7 +2,7 @@
 name = "pymemprofile_api"
 version = "0.1.0"
 authors = ["Itamar Turner-Trauring <[email protected]>"]
-edition = "2018"
+edition = "2021"
 license = "Apache-2.0"
 
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
@@ -18,6 +18,7 @@ once_cell = "1.17"
 libloading = "0.7"
 libc = "0.2"
 serde = {version = "1", features = ["derive"] }
+parking_lot = "0.12.1"
 
 [dependencies.inferno]
 version = "0.11"
@@ -29,9 +30,7 @@ default-features = false
 features = ["memory", "process"]
 
 [dependencies.pyo3]
-version = "0.17"
-default-features = false
-features = []
+version = "0.18"
 
 [target.'cfg(target_os = "linux")'.dependencies]
 cgroups-rs = "0.3.0"
@@ -40,8 +39,9 @@ cgroups-rs = "0.3.0"
 proptest = "1.0"
 proc-maps = "0.3.0"
 tempfile = "3.3.0"
+rusty-fork = "0.3.0"
 
 [features]
 default = []
 # Optimize for the production version of Fil.
-fil4prod = []
+fil4prod = []