]> git.ipfire.org Git - thirdparty/git.git/commitdiff
compat/fsmonitor/fsm-listen-win32: implement FSMonitor backend on Windows
authorJeff Hostetler <jeffhost@microsoft.com>
Fri, 25 Mar 2022 18:02:58 +0000 (18:02 +0000)
committerJunio C Hamano <gitster@pobox.com>
Fri, 25 Mar 2022 23:04:16 +0000 (16:04 -0700)
Teach the win32 backend to register a watch on the working tree
root directory (recursively).  Also watch the <gitdir> if it is
not inside the working tree.  And to collect path change notifications
into batches and publish.

Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
compat/fsmonitor/fsm-listen-win32.c

index 916cbea254f8cf1b54fa56b9d620322f57584125..5b928ab66e5e7117083fb8a0ba497fb36783cb4b 100644 (file)
 #include "config.h"
 #include "fsmonitor.h"
 #include "fsm-listen.h"
+#include "fsmonitor--daemon.h"
+
+/*
+ * The documentation of ReadDirectoryChangesW() states that the maximum
+ * buffer size is 64K when the monitored directory is remote.
+ *
+ * Larger buffers may be used when the monitored directory is local and
+ * will help us receive events faster from the kernel and avoid dropped
+ * events.
+ *
+ * So we try to use a very large buffer and silently fallback to 64K if
+ * we get an error.
+ */
+#define MAX_RDCW_BUF_FALLBACK (65536)
+#define MAX_RDCW_BUF          (65536 * 8)
+
+struct one_watch
+{
+       char buffer[MAX_RDCW_BUF];
+       DWORD buf_len;
+       DWORD count;
+
+       struct strbuf path;
+       HANDLE hDir;
+       HANDLE hEvent;
+       OVERLAPPED overlapped;
+
+       /*
+        * Is there an active ReadDirectoryChangesW() call pending.  If so, we
+        * need to later call GetOverlappedResult() and possibly CancelIoEx().
+        */
+       BOOL is_active;
+};
+
+struct fsmonitor_daemon_backend_data
+{
+       struct one_watch *watch_worktree;
+       struct one_watch *watch_gitdir;
+
+       HANDLE hEventShutdown;
+
+       HANDLE hListener[3]; /* we don't own these handles */
+#define LISTENER_SHUTDOWN 0
+#define LISTENER_HAVE_DATA_WORKTREE 1
+#define LISTENER_HAVE_DATA_GITDIR 2
+       int nr_listener_handles;
+};
+
+/*
+ * Convert the WCHAR path from the notification into UTF8 and
+ * then normalize it.
+ */
+static int normalize_path_in_utf8(FILE_NOTIFY_INFORMATION *info,
+                                 struct strbuf *normalized_path)
+{
+       int reserve;
+       int len = 0;
+
+       strbuf_reset(normalized_path);
+       if (!info->FileNameLength)
+               goto normalize;
+
+       /*
+        * Pre-reserve enough space in the UTF8 buffer for
+        * each Unicode WCHAR character to be mapped into a
+        * sequence of 2 UTF8 characters.  That should let us
+        * avoid ERROR_INSUFFICIENT_BUFFER 99.9+% of the time.
+        */
+       reserve = info->FileNameLength + 1;
+       strbuf_grow(normalized_path, reserve);
+
+       for (;;) {
+               len = WideCharToMultiByte(CP_UTF8, 0, info->FileName,
+                                         info->FileNameLength / sizeof(WCHAR),
+                                         normalized_path->buf,
+                                         strbuf_avail(normalized_path) - 1,
+                                         NULL, NULL);
+               if (len > 0)
+                       goto normalize;
+               if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
+                       error(_("[GLE %ld] could not convert path to UTF-8: '%.*ls'"),
+                             GetLastError(),
+                             (int)(info->FileNameLength / sizeof(WCHAR)),
+                             info->FileName);
+                       return -1;
+               }
+
+               strbuf_grow(normalized_path,
+                           strbuf_avail(normalized_path) + reserve);
+       }
+
+normalize:
+       strbuf_setlen(normalized_path, len);
+       return strbuf_normalize_path(normalized_path);
+}
 
 void fsm_listen__stop_async(struct fsmonitor_daemon_state *state)
 {
+       SetEvent(state->backend_data->hListener[LISTENER_SHUTDOWN]);
+}
+
+static struct one_watch *create_watch(struct fsmonitor_daemon_state *state,
+                                     const char *path)
+{
+       struct one_watch *watch = NULL;
+       DWORD desired_access = FILE_LIST_DIRECTORY;
+       DWORD share_mode =
+               FILE_SHARE_WRITE | FILE_SHARE_READ | FILE_SHARE_DELETE;
+       HANDLE hDir;
+       wchar_t wpath[MAX_PATH];
+
+       if (xutftowcs_path(wpath, path) < 0) {
+               error(_("could not convert to wide characters: '%s'"), path);
+               return NULL;
+       }
+
+       hDir = CreateFileW(wpath,
+                          desired_access, share_mode, NULL, OPEN_EXISTING,
+                          FILE_FLAG_BACKUP_SEMANTICS | FILE_FLAG_OVERLAPPED,
+                          NULL);
+       if (hDir == INVALID_HANDLE_VALUE) {
+               error(_("[GLE %ld] could not watch '%s'"),
+                     GetLastError(), path);
+               return NULL;
+       }
+
+       CALLOC_ARRAY(watch, 1);
+
+       watch->buf_len = sizeof(watch->buffer); /* assume full MAX_RDCW_BUF */
+
+       strbuf_init(&watch->path, 0);
+       strbuf_addstr(&watch->path, path);
+
+       watch->hDir = hDir;
+       watch->hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
+
+       return watch;
+}
+
+static void destroy_watch(struct one_watch *watch)
+{
+       if (!watch)
+               return;
+
+       strbuf_release(&watch->path);
+       if (watch->hDir != INVALID_HANDLE_VALUE)
+               CloseHandle(watch->hDir);
+       if (watch->hEvent != INVALID_HANDLE_VALUE)
+               CloseHandle(watch->hEvent);
+
+       free(watch);
+}
+
+static int start_rdcw_watch(struct fsmonitor_daemon_backend_data *data,
+                           struct one_watch *watch)
+{
+       DWORD dwNotifyFilter =
+               FILE_NOTIFY_CHANGE_FILE_NAME |
+               FILE_NOTIFY_CHANGE_DIR_NAME |
+               FILE_NOTIFY_CHANGE_ATTRIBUTES |
+               FILE_NOTIFY_CHANGE_SIZE |
+               FILE_NOTIFY_CHANGE_LAST_WRITE |
+               FILE_NOTIFY_CHANGE_CREATION;
+
+       ResetEvent(watch->hEvent);
+
+       memset(&watch->overlapped, 0, sizeof(watch->overlapped));
+       watch->overlapped.hEvent = watch->hEvent;
+
+       /*
+        * Queue an async call using Overlapped IO.  This returns immediately.
+        * Our event handle will be signalled when the real result is available.
+        *
+        * The return value here just means that we successfully queued it.
+        * We won't know if the Read...() actually produces data until later.
+        */
+       watch->is_active = ReadDirectoryChangesW(
+               watch->hDir, watch->buffer, watch->buf_len, TRUE,
+               dwNotifyFilter, &watch->count, &watch->overlapped, NULL);
+
+       if (watch->is_active)
+               return 0;
+
+       error(_("ReadDirectoryChangedW failed on '%s' [GLE %ld]"),
+             watch->path.buf, GetLastError());
+       return -1;
+}
+
+static int recv_rdcw_watch(struct one_watch *watch)
+{
+       DWORD gle;
+
+       watch->is_active = FALSE;
+
+       /*
+        * The overlapped result is ready.  If the Read...() was successful
+        * we finally receive the actual result into our buffer.
+        */
+       if (GetOverlappedResult(watch->hDir, &watch->overlapped, &watch->count,
+                               TRUE))
+               return 0;
+
+       gle = GetLastError();
+       if (gle == ERROR_INVALID_PARAMETER &&
+           /*
+            * The kernel throws an invalid parameter error when our
+            * buffer is too big and we are pointed at a remote
+            * directory (and possibly for other reasons).  Quietly
+            * set it down and try again.
+            *
+            * See note about MAX_RDCW_BUF at the top.
+            */
+           watch->buf_len > MAX_RDCW_BUF_FALLBACK) {
+               watch->buf_len = MAX_RDCW_BUF_FALLBACK;
+               return -2;
+       }
+
+       /*
+        * NEEDSWORK: If an external <gitdir> is deleted, the above
+        * returns an error.  I'm not sure that there's anything that
+        * we can do here other than failing -- the <worktree>/.git
+        * link file would be broken anyway.  We might try to check
+        * for that and return a better error message, but I'm not
+        * sure it is worth it.
+        */
+
+       error(_("GetOverlappedResult failed on '%s' [GLE %ld]"),
+             watch->path.buf, gle);
+       return -1;
+}
+
+static void cancel_rdcw_watch(struct one_watch *watch)
+{
+       DWORD count;
+
+       if (!watch || !watch->is_active)
+               return;
+
+       /*
+        * The calls to ReadDirectoryChangesW() and GetOverlappedResult()
+        * form a "pair" (my term) where we queue an IO and promise to
+        * hang around and wait for the kernel to give us the result.
+        *
+        * If for some reason after we queue the IO, we have to quit
+        * or otherwise not stick around for the second half, we must
+        * tell the kernel to abort the IO.  This prevents the kernel
+        * from writing to our buffer and/or signalling our event
+        * after we free them.
+        *
+        * (Ask me how much fun it was to track that one down).
+        */
+       CancelIoEx(watch->hDir, &watch->overlapped);
+       GetOverlappedResult(watch->hDir, &watch->overlapped, &count, TRUE);
+       watch->is_active = FALSE;
+}
+
+/*
+ * Process filesystem events that happen anywhere (recursively) under the
+ * <worktree> root directory.  For a normal working directory, this includes
+ * both version controlled files and the contents of the .git/ directory.
+ *
+ * If <worktree>/.git is a file, then we only see events for the file
+ * itself.
+ */
+static int process_worktree_events(struct fsmonitor_daemon_state *state)
+{
+       struct fsmonitor_daemon_backend_data *data = state->backend_data;
+       struct one_watch *watch = data->watch_worktree;
+       struct strbuf path = STRBUF_INIT;
+       struct string_list cookie_list = STRING_LIST_INIT_DUP;
+       struct fsmonitor_batch *batch = NULL;
+       const char *p = watch->buffer;
+
+       /*
+        * If the kernel gets more events than will fit in the kernel
+        * buffer associated with our RDCW handle, it drops them and
+        * returns a count of zero.
+        *
+        * Yes, the call returns WITHOUT error and with length zero.
+        * This is the documented behavior.  (My testing has confirmed
+        * that it also sets the last error to ERROR_NOTIFY_ENUM_DIR,
+        * but we do not rely on that since the function did not
+        * return an error and it is not documented.)
+        *
+        * (The "overflow" case is not ambiguous with the "no data" case
+        * because we did an INFINITE wait.)
+        *
+        * This means we have a gap in coverage.  Tell the daemon layer
+        * to resync.
+        */
+       if (!watch->count) {
+               trace2_data_string("fsmonitor", NULL, "fsm-listen/kernel",
+                                  "overflow");
+               fsmonitor_force_resync(state);
+               return LISTENER_HAVE_DATA_WORKTREE;
+       }
+
+       /*
+        * On Windows, `info` contains an "array" of paths that are
+        * relative to the root of whichever directory handle received
+        * the event.
+        */
+       for (;;) {
+               FILE_NOTIFY_INFORMATION *info = (void *)p;
+               const char *slash;
+               enum fsmonitor_path_type t;
+
+               strbuf_reset(&path);
+               if (normalize_path_in_utf8(info, &path) == -1)
+                       goto skip_this_path;
+
+               t = fsmonitor_classify_path_workdir_relative(path.buf);
+
+               switch (t) {
+               case IS_INSIDE_DOT_GIT_WITH_COOKIE_PREFIX:
+                       /* special case cookie files within .git */
+
+                       /* Use just the filename of the cookie file. */
+                       slash = find_last_dir_sep(path.buf);
+                       string_list_append(&cookie_list,
+                                          slash ? slash + 1 : path.buf);
+                       break;
+
+               case IS_INSIDE_DOT_GIT:
+                       /* ignore everything inside of "<worktree>/.git/" */
+                       break;
+
+               case IS_DOT_GIT:
+                       /* "<worktree>/.git" was deleted (or renamed away) */
+                       if ((info->Action == FILE_ACTION_REMOVED) ||
+                           (info->Action == FILE_ACTION_RENAMED_OLD_NAME)) {
+                               trace2_data_string("fsmonitor", NULL,
+                                                  "fsm-listen/dotgit",
+                                                  "removed");
+                               goto force_shutdown;
+                       }
+                       break;
+
+               case IS_WORKDIR_PATH:
+                       /* queue normal pathname */
+                       if (!batch)
+                               batch = fsmonitor_batch__new();
+                       fsmonitor_batch__add_path(batch, path.buf);
+                       break;
+
+               case IS_GITDIR:
+               case IS_INSIDE_GITDIR:
+               case IS_INSIDE_GITDIR_WITH_COOKIE_PREFIX:
+               default:
+                       BUG("unexpected path classification '%d' for '%s'",
+                           t, path.buf);
+               }
+
+skip_this_path:
+               if (!info->NextEntryOffset)
+                       break;
+               p += info->NextEntryOffset;
+       }
+
+       fsmonitor_publish(state, batch, &cookie_list);
+       batch = NULL;
+       string_list_clear(&cookie_list, 0);
+       strbuf_release(&path);
+       return LISTENER_HAVE_DATA_WORKTREE;
+
+force_shutdown:
+       fsmonitor_batch__free_list(batch);
+       string_list_clear(&cookie_list, 0);
+       strbuf_release(&path);
+       return LISTENER_SHUTDOWN;
+}
+
+/*
+ * Process filesystem events that happened anywhere (recursively) under the
+ * external <gitdir> (such as non-primary worktrees or submodules).
+ * We only care about cookie files that our client threads created here.
+ *
+ * Note that we DO NOT get filesystem events on the external <gitdir>
+ * itself (it is not inside something that we are watching).  In particular,
+ * we do not get an event if the external <gitdir> is deleted.
+ */
+static int process_gitdir_events(struct fsmonitor_daemon_state *state)
+{
+       struct fsmonitor_daemon_backend_data *data = state->backend_data;
+       struct one_watch *watch = data->watch_gitdir;
+       struct strbuf path = STRBUF_INIT;
+       struct string_list cookie_list = STRING_LIST_INIT_DUP;
+       const char *p = watch->buffer;
+
+       if (!watch->count) {
+               trace2_data_string("fsmonitor", NULL, "fsm-listen/kernel",
+                                  "overflow");
+               fsmonitor_force_resync(state);
+               return LISTENER_HAVE_DATA_GITDIR;
+       }
+
+       for (;;) {
+               FILE_NOTIFY_INFORMATION *info = (void *)p;
+               const char *slash;
+               enum fsmonitor_path_type t;
+
+               strbuf_reset(&path);
+               if (normalize_path_in_utf8(info, &path) == -1)
+                       goto skip_this_path;
+
+               t = fsmonitor_classify_path_gitdir_relative(path.buf);
+
+               switch (t) {
+               case IS_INSIDE_GITDIR_WITH_COOKIE_PREFIX:
+                       /* special case cookie files within gitdir */
+
+                       /* Use just the filename of the cookie file. */
+                       slash = find_last_dir_sep(path.buf);
+                       string_list_append(&cookie_list,
+                                          slash ? slash + 1 : path.buf);
+                       break;
+
+               case IS_INSIDE_GITDIR:
+                       goto skip_this_path;
+
+               default:
+                       BUG("unexpected path classification '%d' for '%s'",
+                           t, path.buf);
+               }
+
+skip_this_path:
+               if (!info->NextEntryOffset)
+                       break;
+               p += info->NextEntryOffset;
+       }
+
+       fsmonitor_publish(state, NULL, &cookie_list);
+       string_list_clear(&cookie_list, 0);
+       strbuf_release(&path);
+       return LISTENER_HAVE_DATA_GITDIR;
 }
 
 void fsm_listen__loop(struct fsmonitor_daemon_state *state)
 {
+       struct fsmonitor_daemon_backend_data *data = state->backend_data;
+       DWORD dwWait;
+       int result;
+
+       state->error_code = 0;
+
+       if (start_rdcw_watch(data, data->watch_worktree) == -1)
+               goto force_error_stop;
+
+       if (data->watch_gitdir &&
+           start_rdcw_watch(data, data->watch_gitdir) == -1)
+               goto force_error_stop;
+
+       for (;;) {
+               dwWait = WaitForMultipleObjects(data->nr_listener_handles,
+                                               data->hListener,
+                                               FALSE, INFINITE);
+
+               if (dwWait == WAIT_OBJECT_0 + LISTENER_HAVE_DATA_WORKTREE) {
+                       result = recv_rdcw_watch(data->watch_worktree);
+                       if (result == -1) {
+                               /* hard error */
+                               goto force_error_stop;
+                       }
+                       if (result == -2) {
+                               /* retryable error */
+                               if (start_rdcw_watch(data, data->watch_worktree) == -1)
+                                       goto force_error_stop;
+                               continue;
+                       }
+
+                       /* have data */
+                       if (process_worktree_events(state) == LISTENER_SHUTDOWN)
+                               goto force_shutdown;
+                       if (start_rdcw_watch(data, data->watch_worktree) == -1)
+                               goto force_error_stop;
+                       continue;
+               }
+
+               if (dwWait == WAIT_OBJECT_0 + LISTENER_HAVE_DATA_GITDIR) {
+                       result = recv_rdcw_watch(data->watch_gitdir);
+                       if (result == -1) {
+                               /* hard error */
+                               goto force_error_stop;
+                       }
+                       if (result == -2) {
+                               /* retryable error */
+                               if (start_rdcw_watch(data, data->watch_gitdir) == -1)
+                                       goto force_error_stop;
+                               continue;
+                       }
+
+                       /* have data */
+                       if (process_gitdir_events(state) == LISTENER_SHUTDOWN)
+                               goto force_shutdown;
+                       if (start_rdcw_watch(data, data->watch_gitdir) == -1)
+                               goto force_error_stop;
+                       continue;
+               }
+
+               if (dwWait == WAIT_OBJECT_0 + LISTENER_SHUTDOWN)
+                       goto clean_shutdown;
+
+               error(_("could not read directory changes [GLE %ld]"),
+                     GetLastError());
+               goto force_error_stop;
+       }
+
+force_error_stop:
+       state->error_code = -1;
+
+force_shutdown:
+       /*
+        * Tell the IPC thead pool to stop (which completes the await
+        * in the main thread (which will also signal this thread (if
+        * we are still alive))).
+        */
+       ipc_server_stop_async(state->ipc_server_data);
+
+clean_shutdown:
+       cancel_rdcw_watch(data->watch_worktree);
+       cancel_rdcw_watch(data->watch_gitdir);
 }
 
 int fsm_listen__ctor(struct fsmonitor_daemon_state *state)
 {
+       struct fsmonitor_daemon_backend_data *data;
+
+       CALLOC_ARRAY(data, 1);
+
+       data->hEventShutdown = CreateEvent(NULL, TRUE, FALSE, NULL);
+
+       data->watch_worktree = create_watch(state,
+                                           state->path_worktree_watch.buf);
+       if (!data->watch_worktree)
+               goto failed;
+
+       if (state->nr_paths_watching > 1) {
+               data->watch_gitdir = create_watch(state,
+                                                 state->path_gitdir_watch.buf);
+               if (!data->watch_gitdir)
+                       goto failed;
+       }
+
+       data->hListener[LISTENER_SHUTDOWN] = data->hEventShutdown;
+       data->nr_listener_handles++;
+
+       data->hListener[LISTENER_HAVE_DATA_WORKTREE] =
+               data->watch_worktree->hEvent;
+       data->nr_listener_handles++;
+
+       if (data->watch_gitdir) {
+               data->hListener[LISTENER_HAVE_DATA_GITDIR] =
+                       data->watch_gitdir->hEvent;
+               data->nr_listener_handles++;
+       }
+
+       state->backend_data = data;
+       return 0;
+
+failed:
+       CloseHandle(data->hEventShutdown);
+       destroy_watch(data->watch_worktree);
+       destroy_watch(data->watch_gitdir);
+
        return -1;
 }
 
 void fsm_listen__dtor(struct fsmonitor_daemon_state *state)
 {
+       struct fsmonitor_daemon_backend_data *data;
+
+       if (!state || !state->backend_data)
+               return;
+
+       data = state->backend_data;
+
+       CloseHandle(data->hEventShutdown);
+       destroy_watch(data->watch_worktree);
+       destroy_watch(data->watch_gitdir);
+
+       FREE_AND_NULL(state->backend_data);
 }