status: add fallback to slow crawl on watchman failure

Summary:
The Python "status" code would fall back to non-Watchman status on a Watchman error by default. Previously I didn't reproduce this in the Rust status, hoping it wasn't necessary. However, various cases have popped up where Watchman can't be used in various environments for various reasons, so let's add the fallback to Rust as well.

Note that some of the Watchman errors have been real bugs in Watchman (e.g. startup race condition in Windows), and potential "environment" bugs (e.g. Watchman simply not installed), but it does seem like we should err on the side of not blocking users with errors. I'm going to disable fsmonitor.fallback-on-watchman-exception for tests and local development so we don't hide new issues.

Reviewed By: quark-zju

Differential Revision: D51265962

fbshipit-source-id: d66dd47e83f4bd3811f04d237a75148030cc8526

eden/scm/lib/workingcopy/src/watchmanfs/watchmanfs.rs

@@ -156,52 +156,7 @@ impl WatchmanFileSystem {
 
         Ok(result)
     }
-}
-
-async fn crawl_progress(root: PathBuf, approx_file_count: u64) -> Result<()> {
-    let client = {
-        let _bar = ProgressBar::new_detached("connecting watchman", 0, "");
-
-        // If watchman just started (and we issued "watch-project" from
-        // query_files), this connect gets stuck indefinitely. Work around by
-        // timing out and retrying until we get through.
-        loop {
-            match tokio::time::timeout(Duration::from_secs(1), Connector::new().connect()).await {
-                Ok(client) => break client?,
-                Err(_) => {}
-            };
-
-            tokio::time::sleep(Duration::from_secs(1)).await;
-        }
-    };
-
-    let mut bar = None;
-
-    let req = DebugRootStatusRequest(
-        "debug-root-status",
-        CanonicalPath::canonicalize(root)?.into_path_buf(),
-    );
-
-    loop {
-        let response: DebugRootStatusResponse = client.generic_request(req.clone()).await?;
 
-        if let Some(RootStatus {
-            recrawl_info: Some(RecrawlInfo { stats: Some(stats) }),
-        }) = response.root_status
-        {
-            bar.get_or_insert_with(|| {
-                ProgressBar::new_detached("crawling", approx_file_count, "files (approx)")
-            })
-            .set_position(stats);
-        } else if bar.is_some() {
-            return Ok(());
-        }
-
-        tokio::time::sleep(Duration::from_millis(100)).await;
-    }
-}
-
-impl FileSystem for WatchmanFileSystem {
     #[tracing::instrument(skip_all)]
     fn pending_changes(
         &self,
@@ -265,11 +220,14 @@ impl FileSystem for WatchmanFileSystem {
                         })?,
                 },
                 ignore_dirs,
-            ))?
+            ))
         };
 
+        // Make sure we always abort - even in case of error.
         progress_handle.abort();
 
+        let result = result?;
+
         tracing::debug!(
             target: "watchman_info",
             watchmanfreshinstances= if result.is_fresh_instance { 1 } else { 0 },
@@ -397,6 +355,102 @@ impl FileSystem for WatchmanFileSystem {
 
         Ok(Box::new(pending_changes.into_iter()))
     }
+}
+
+async fn crawl_progress(root: PathBuf, approx_file_count: u64) -> Result<()> {
+    let client = {
+        let _bar = ProgressBar::new_detached("connecting watchman", 0, "");
+
+        // If watchman just started (and we issued "watch-project" from
+        // query_files), this connect gets stuck indefinitely. Work around by
+        // timing out and retrying until we get through.
+        loop {
+            match tokio::time::timeout(Duration::from_secs(1), Connector::new().connect()).await {
+                Ok(client) => break client?,
+                Err(_) => {}
+            };
+
+            tokio::time::sleep(Duration::from_secs(1)).await;
+        }
+    };
+
+    let mut bar = None;
+
+    let req = DebugRootStatusRequest(
+        "debug-root-status",
+        CanonicalPath::canonicalize(root)?.into_path_buf(),
+    );
+
+    loop {
+        let response: DebugRootStatusResponse = client.generic_request(req.clone()).await?;
+
+        if let Some(RootStatus {
+            recrawl_info: Some(RecrawlInfo { stats: Some(stats) }),
+        }) = response.root_status
+        {
+            bar.get_or_insert_with(|| {
+                ProgressBar::new_detached("crawling", approx_file_count, "files (approx)")
+            })
+            .set_position(stats);
+        } else if bar.is_some() {
+            return Ok(());
+        }
+
+        tokio::time::sleep(Duration::from_millis(100)).await;
+    }
+}
+
+impl FileSystem for WatchmanFileSystem {
+    fn pending_changes(
+        &self,
+        matcher: DynMatcher,
+        ignore_matcher: DynMatcher,
+        ignore_dirs: Vec<PathBuf>,
+        include_ignored: bool,
+        last_write: SystemTime,
+        config: &dyn Config,
+        io: &IO,
+    ) -> Result<Box<dyn Iterator<Item = Result<PendingChange>>>> {
+        let result = self.pending_changes(
+            matcher.clone(),
+            ignore_matcher.clone(),
+            ignore_dirs.clone(),
+            include_ignored,
+            last_write,
+            config,
+            io,
+        );
+
+        match result {
+            Ok(result) => Ok(result),
+            Err(err) if err.is::<watchman_client::Error>() => {
+                if !config.get_or("fsmonitor", "fallback-on-watchman-exception", || true)? {
+                    return Err(err);
+                }
+
+                // On watchman error, fall back to manual walk. This is important for errors such as:
+                //   - "watchman" binary not in PATH
+                //   - unsupported filesystem (e.g. NFS)
+                //
+                // A better approach might be an allowlist of errors to fall
+                // back on so we can fail hard in cases where watchman "should"
+                // work, but that is probably still an unacceptable UX in general.
+
+                tracing::debug!(target: "watchman_info", watchmanfallback=1);
+                tracing::warn!(?err, "watchman error - falling back to slow crawl");
+                self.inner.pending_changes(
+                    matcher,
+                    ignore_matcher,
+                    ignore_dirs,
+                    include_ignored,
+                    last_write,
+                    config,
+                    io,
+                )
+            }
+            Err(err) => Err(err),
+        }
+    }
 
     fn sparse_matcher(
         &self,

eden/scm/tests/test-fsmonitor-fallback.t

@@ -0,0 +1,24 @@
+#debugruntest-compatible
+#require fsmonitor
+
+  $ configure modernclient
+  $ newclientrepo
+  $ echo foo > foo
+  $ hg commit -Aqm foo
+
+  $ echo nope > $TESTTMP/watchman
+  $ chmod +x $TESTTMP/watchman
+  $ export PATH=$TESTTMP:$PATH
+  $ unset WATCHMAN_SOCK
+
+  $ echo foo >> foo
+  $ LOG=warn,watchman_info=debug hg st --config fsmonitor.fallback-on-watchman-exception=true
+  DEBUG watchman_info: watchmanfallback=1
+   WARN workingcopy::watchmanfs::watchmanfs: watchman error - falling back to slow crawl * (glob)
+  ` (?)
+  M foo
+
+  $ LOG=warn,watchman_info=debug hg st --config fsmonitor.fallback-on-watchman-exception=false
+  abort: While invoking the watchman CLI * (glob)
+  ` (?)
+  [255]