inode: Fix double locking

This was a long lurking buglet caught by the new inode locking code and
chased down by Tormod. I think this approach is correct. The old inefficiency
of doing a bfind to find the buffer does not exist and the a bread discovering
this again is still present.

diff --git a/Kernel/inode.c b/Kernel/inode.c
index a3cbb3a..dacfb7c 100644 (file)
--- a/Kernel/inode.c
+++ b/Kernel/inode.c
@@ -92,8 +92,10 @@ void readi(regptr inoptr ino, uint8_t flag)
                while (udata.u_count) {
                        pblk = mapcalc(ino, &amount, 1);
 
-#if defined(read_direct)
-                       if (!ispipe && pblk != NULLBLK && amount == BLKSIZE && read_direct(flag) && bfind(dev, pblk) == 0) {
+#if !defined(read_direct)
+                       bp = NULL;
+#else
+                       if (pblk != NULLBLK && (bp = bfind(dev, pblk)) == NULL && !ispipe && amount == BLKSIZE && read_direct(flag) == 0) {
                                /* we can transfer direct from disk to the userspace buffer */
                                /* FIXME: allow for async queued I/O here. We want
                                   an API something like breadasync() that either
@@ -117,10 +119,11 @@ void readi(regptr inoptr ino, uint8_t flag)
                                /* we transfer through the buffer pool */
                                if (pblk == NULLBLK)
                                        bp = zerobuf();
-                               else
+                               else if (bp == NULL) {
                                        bp = bread(dev, pblk, 0);
-                               if (bp == NULL)
-                                       break;
+                                       if (bp == NULL)
+                                               break;
+                               }
                                uputblk(bp, uoff(), amount);
 
                                brelse(bp);