Changed locking rules. Scans and appends can peacefully co-exist now.

Should solve the problems with shep-reap waiting for bucket file transmission
to finish.

diff --git a/lib/bucket.c b/lib/bucket.c
index 3516e85126eaa3b0d914245b0cc3dbcc940d236d..f1e9e8b119fcb16baba7b65108d0830be0aa0e5b 100644 (file)
--- a/lib/bucket.c
+++ b/lib/bucket.c
@@ -59,22 +59,39 @@ obuck_broken(char *msg, sh_off_t pos)
 }
 
 /*
- *  Unfortunately we cannot use flock() here since it happily permits
- *  locking a shared fd (e.g., after fork()) multiple times. The fcntl
- *  locks are very ugly and they don't support 64-bit offsets, but we
- *  can work around the problem by always locking the first header
- *  in the file.
+ *  We need several types of locks:
+ *
+ *     Read lock       reading parts of bucket file
+ *     Write lock      any write operations
+ *     Append lock     appending to the end of the file
+ *     Scan lock       reading parts which we are certain they exist
+ *
+ *  Multiple read and scan locks can co-exist together.
+ *  Scan locks can co-exist with an append lock.
+ *  There can be at most one write/append lock at a time.
+ *
+ *  These lock types map to a pair of normal read-write locks which
+ *  we represent as fcntl() locks on the first and second byte of the
+ *  bucket file. [We cannot use flock() since it happily permits
+ *  locking a shared fd (e.g., after fork()) multiple times at it also
+ *  doesn't offer multiple locks on a single file.]
+ *
+ *                     byte0           byte1
+ *     Read            <read>          <read>
+ *     Write           <write>         <write>
+ *     Append          -               <write>
+ *     Scan            <read>          -
  */
 
 static inline void
-obuck_do_lock(int type)
+obuck_do_lock(int type, int start, int len)
 {
   struct flock fl;
 
   fl.l_type = type;
   fl.l_whence = SEEK_SET;
-  fl.l_start = 0;
-  fl.l_len = sizeof(struct obuck_header);
+  fl.l_start = start;
+  fl.l_len = len;
   if (fcntl(obuck_fd, F_SETLKW, &fl) < 0)
     die("fcntl lock: %m");
 }
@@ -82,19 +99,31 @@ obuck_do_lock(int type)
 inline void
 obuck_lock_read(void)
 {
-  obuck_do_lock(F_RDLCK);
+  obuck_do_lock(F_RDLCK, 0, 2);
 }
 
 inline void
 obuck_lock_write(void)
 {
-  obuck_do_lock(F_WRLCK);
+  obuck_do_lock(F_WRLCK, 0, 2);
+}
+
+static inline void
+obuck_lock_append(void)
+{
+  obuck_do_lock(F_WRLCK, 0, 1);
+}
+
+static inline void
+obuck_lock_read_to_scan(void)
+{
+  obuck_do_lock(F_UNLCK, 0, 1);
 }
 
 inline void
 obuck_unlock(void)
 {
-  obuck_do_lock(F_UNLCK);
+  obuck_do_lock(F_UNLCK, 0, 2);
 }
 
 /*** FastIO emulation ***/
@@ -306,7 +335,7 @@ obuck_create(u32 type)
 {
   ASSERT(!obuck_write_fb);
 
-  obuck_lock_write();
+  obuck_lock_append();
   sh_off_t start = sh_seek(obuck_fd, 0, SEEK_END);
   if (start & (OBUCK_ALIGN - 1))
     obuck_broken("Misaligned file", start);
@@ -369,7 +398,7 @@ obuck_delete(oid_t oid)
 
 static struct fastbuf *obuck_rpf;
 static uns slurp_remains;
-static sh_off_t slurp_start, slurp_current;
+static sh_off_t slurp_start, slurp_current, slurp_end;
 
 static int
 obuck_slurp_refill(struct fastbuf *f)
@@ -401,6 +430,10 @@ obuck_slurp_pool(struct obuck_header *hdrp)
        {
          obuck_lock_read();
          obuck_rpf = bopen(obuck_name, O_RDONLY, obuck_slurp_buflen);
+         bseek(obuck_rpf, 0, SEEK_END);
+         slurp_end = btell(obuck_rpf);
+         bsetpos(obuck_rpf, 0);
+         obuck_lock_read_to_scan();
        }
       else
        {
@@ -409,7 +442,10 @@ obuck_slurp_pool(struct obuck_header *hdrp)
            obuck_broken("Missing trailer", slurp_start);
        }
       slurp_start = btell(obuck_rpf);
-      l = bread(obuck_rpf, hdrp, sizeof(struct obuck_header));
+      if (slurp_start < slurp_end)
+       l = bread(obuck_rpf, hdrp, sizeof(struct obuck_header));
+      else
+       l = 0;
       if (!l)
        {
          bclose(obuck_rpf);