Hi,

I am posting anonymously because I cannot find my password for the forums. My user name is thebondo.

I use fossil on windows (although not as a server) so I thought I would give this a look. Hopefully this will be helpful. The changes below have only been tested on windows.

After rummaging around the fossil source code and tracing what happens for running the backoffice in windows, the primary cause appears to be in main.c in the function cmd_http. Here is a unified diff with a proposed fix for getting the backoffice to run.

  --- src/main.c
  +++ src/main.c
  @@ -2756,10 +2756,10 @@
     if( zInFile ){
  -    backoffice_disable();
  +//  backoffice_disable();
       g.httpIn = fossil_fopen(zInFile, "rb");
       if( g.httpIn==0 ) fossil_fatal("cannot open \"%s\" for reading", zInFile);
     }else{
       g.httpIn = stdin;
   #if defined(_WIN32)
      _setmode(_fileno(stdin), _O_BINARY);
   #endif
     }

The function win32_http_request from winhttp.c handles requests for the windows server (either running in a command prompt or running as a service). This function serves the request by running a child process for a "fossil http" command using fossil_system. The options used for this command always use --in and --out. So when cmd_http is called from "fossil server" on windows, zInFile is always set, which means the line backoffice_disable() is always executed. This effectively stops the backoffice from running. On Linux this is not an issue because fossil server handles requests by forking and ultimately calling process_one_web_page from the child process (I think).

Initially, I tried just this fix. However, I happened to be working in a path with spaces (C:fossil test), which caused a separate issue that also prevented the backoffice from running as expected.

After some more exploring, I found that the important block is in the function backoffice_run_if_needed from backoffice.c (shown below with a diff for a proposed fix). By tracing the execution, I found that _wspawnv splits the elements of the third argument on unquoted spaces. It does not split the second argument on unquoted spaces. So "C:fossil testfossil.exe" was getting executed, but it was being given an argument list of "C:fossil" "testfossil.exe" "backoffice" "-R" "C:fossil" "testserver.fossil". Just fixing the quoting of g.nameOfExe was not sufficient. Then, the backoffice would run, but it would silently fail because it would not find either "C:fossil" or "testserver.fossil" as repositories.

  --- src/backoffice.c
  +++ src/backoffice.c
  @@ -822,20 +822,23 @@
   #if defined(_WIN32)
     {
       int i;
       intptr_t x;
  -    char *argv[4];
  -    wchar_t *ax[5];
  +    char *argv[5];
  +    wchar_t *ax[6];
       argv[0] = g.nameOfExe;
  -    argv[1] = "backoffice";
  -    argv[2] = "-R";
  -    argv[3] = backofficeDb;
  -    ax[4] = 0;
  -    for(i=0; i<=3; i++) ax[i] = fossil_utf8_to_unicode(argv[i]);
  -    x = _wspawnv(_P_NOWAIT, ax[0], (const wchar_t * const *)ax);
  -    for(i=0; i<=3; i++) fossil_unicode_free(ax[i]);
  +    argv[1] = mprintf("%$", g.nameOfExe);
  +    argv[2] = "backoffice";
  +    argv[3] = "-R";
  +    argv[4] = mprintf("%$", backofficeDb);
  +    ax[5] = 0;
  +    for(i=0; i<=4; i++) ax[i] = fossil_utf8_to_unicode(argv[i]);
  +    x = _wspawnv(_P_NOWAIT, ax[0], (const wchar_t * const *)(ax+1));
  +    for(i=0; i<=4; i++) fossil_unicode_free(ax[i]);
  +    fossil_free(argv[1]);
  +    fossil_free(argv[4]);
       backofficeTrace(
         "/***** Subprocess %d creates backoffice child %lu *****/\n",
         GETPID(), GetProcessId((HANDLE)x));
       if( x>=0 ) return;
     }
   #else /* unix */

The preceding diff fixes the issue with spaces by

1. inserting a new argv[1] item with a quoted version of g.nameOfExe
2. quoting backofficeDb for argv[4]
3. Updateing _wspawnv to use ax[1..5] instead of ax[0..4]
4. Updating the loop limits
5. Adding code to free argv[1] and argv[4]

I did a global code search to see where else "--in" might be used internally, and I did not find any uses in src that looked like there would be issues. However, it does look like tests might be affected. My guess is that the line in main.c was originally added so that test_fossil_http from testtester.tcl would not try to invoke the backoffice. Perhaps a separate flag could be added to fossil http to prevent the backoffice from running when needed?