Changeset a4627c4 in mainline

Timestamp:

2007-09-22T20:08:12Z (17 years ago)

Author:

Jakub Jermar <jakub@…>

Branches:

lfn, master, serial, ticket/834-toolchain-update, topic/msim-upgrade, topic/simplify-dev-export

Children:

56bc4a3

Parents:

8f9ede5

Message:

VFS work.

Describe the bottleneck of one VFS-FS connection in fat.c. Add a comment
outlining possible approaches.

Also in fat.c, the async_create_manager() _must_ be called as well as
async_set_client_connection(). The latter alows VFS to duplicate the connection.
Modify fat_connection() to ACK all IPC_M_CONNECT_ME_TO calls.

Location:

uspace/srv

Files:

• fs/fat/fat.c (modified) (5 diffs)
• vfs/vfs_register.c (modified) (1 diff)

uspace/srv/fs/fat/fat.c

@@ -42 +42 @@
 #include <errno.h>
 #include <unistd.h>
-#include <stdio.h>>
+#include <stdio.h>
 #include "../../vfs/vfs.h"
 
@@ -63 +63 @@
 };
 
-/*
- * This fibril processes request from the VFS server.
+/**
+ * This connection fibril processes VFS requests from VFS.
+ *
+ * In order to support simultaneous VFS requests, our design is as follows.
+ * The connection fibril accepts VFS requests from VFS. If there is only one
+ * instance of the fibril, VFS will need to serialize all VFS requests it sends
+ * to FAT. To overcome this bottleneck, VFS can send FAT the IPC_M_CONNECT_ME_TO
+ * call. In that case, a new connection fibril will be created, which in turn
+ * will accept the call. Thus, a new phone will be opened for VFS.
+ *
+ * There are few issues with this arrangement. First, VFS can run out of
+ * available phones. In that case, VFS can close some other phones or use one
+ * phone for more serialized requests. Similarly, FAT can refuse to duplicate
+ * the connection. VFS should then just make use of already existing phones and
+ * route its requests through them. To avoid paying the fibril creation price 
+ * upon each request, FAT might want to keep the connections open after the
+ * request has been completed.
  */
-void fat_connection(ipc_callid_t iid, ipc_call_t *icall)
+static void fat_connection(ipc_callid_t iid, ipc_call_t *icall)
 {
-        dprintf("Callback connection established.\n");
+        if (iid) {
+                /*
+                 * This only happens for connections opened by
+                 * IPC_M_CONNECT_ME_TO calls as opposed to callback connections
+                 * created by IPC_M_CONNECT_TO_ME.
+                 */
+                ipc_answer_fast(iid, EOK, 0, 0);
+        }
+        
+        dprintf("VFS-FAT connection established.\n");
         while (1) {
                 ipc_callid_t callid;
@@ -74 +98 @@
         
                 callid = async_get_call(&call);
-                ipc_answer_fast(callid, ENOTSUP, 0, 0);
+                switch  (IPC_GET_METHOD(call)) {
+                default:
+                        ipc_answer_fast(callid, ENOTSUP, 0, 0);
+                        break;
+                }
         }
 }
@@ -113 +141 @@
         ipc_connect_to_me(vfs_phone, 0, 0, &phonehash);
 
+        /*
+         * Create a connection fibril to handle the callback connection.
+         */
         async_new_connection(phonehash, 0, NULL, fat_connection);
+        
+        /*
+         * Tell the async framework that other connections are to be handled by
+         * the same connection fibril as well.
+         */
+        async_set_client_connection(fat_connection);
 
         /*
@@ -121 +158 @@
         dprintf("FAT filesystem registered.\n");
 
+        async_create_manager();
+
         /*
          * TODO: Interestingly, if we merely return, the only thread dies.
-         *       If the only thread dies, the whole task is destroyed. 
-         *       Prevent the thread from exiting when there are active fibrils.
+         *       If the only thread dies, the whole task is destroyed. 
+         *       Prevent the thread from exiting when there are active fibrils.
          */
         fibril_schedule_next_adv(FIBRIL_FROM_DEAD);

uspace/srv/vfs/vfs_register.c

@@ -270 +270 @@
         dprintf("\"%s\" filesystem successfully registered.\n",
             fs_info->vfs_info.name);
+
 }