]> err.no Git - linux-2.6/commitdiff
SUNRPC: switchable buffer allocation
authorChuck Lever <cel@netapp.com>
Tue, 3 Jan 2006 08:55:49 +0000 (09:55 +0100)
committerTrond Myklebust <Trond.Myklebust@netapp.com>
Fri, 6 Jan 2006 19:58:55 +0000 (14:58 -0500)
 Add RPC client transport switch support for replacing buffer management
 on a per-transport basis.

 In the current IPv4 socket transport implementation, RPC buffers are
 allocated as needed for each RPC message that is sent.  Some transport
 implementations may choose to use pre-allocated buffers for encoding,
 sending, receiving, and unmarshalling RPC messages, however.  For
 transports capable of direct data placement, the buffers can be carved
 out of a pre-registered area of memory rather than from a slab cache.

 Test-plan:
 Millions of fsx operations.  Performance characterization with "sio" and
 "iozone".  Use oprofile and other tools to look for significant regression
 in CPU utilization.

Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
include/linux/sunrpc/sched.h
include/linux/sunrpc/xprt.h
net/sunrpc/clnt.c
net/sunrpc/sched.c
net/sunrpc/xprt.c
net/sunrpc/xprtsock.c

index 94b0afa4ab05c2a38f1d2c30e932deb1cfa2d3af..8b25629accd8880e044d57558f452b4d50a15283 100644 (file)
@@ -52,8 +52,6 @@ struct rpc_task {
         * RPC call state
         */
        struct rpc_message      tk_msg;         /* RPC call info */
-       __u32 *                 tk_buffer;      /* XDR buffer */
-       size_t                  tk_bufsize;
        __u8                    tk_garb_retry;
        __u8                    tk_cred_retry;
 
@@ -268,6 +266,7 @@ struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *);
 void           rpc_wake_up_status(struct rpc_wait_queue *, int);
 void           rpc_delay(struct rpc_task *, unsigned long);
 void *         rpc_malloc(struct rpc_task *, size_t);
+void           rpc_free(struct rpc_task *);
 int            rpciod_up(void);
 void           rpciod_down(void);
 void           rpciod_wake_up(void);
index 3b8b6e823c70eab0083330be5dd8b25727b037c0..7885b9621ce342c5293d15a21fbe7fbb3f591aae 100644 (file)
@@ -79,21 +79,19 @@ struct rpc_rqst {
        void (*rq_release_snd_buf)(struct rpc_rqst *); /* release rq_enc_pages */
        struct list_head        rq_list;
 
+       __u32 *                 rq_buffer;      /* XDR encode buffer */
+       size_t                  rq_bufsize;
+
        struct xdr_buf          rq_private_buf;         /* The receive buffer
                                                         * used in the softirq.
                                                         */
        unsigned long           rq_majortimeo;  /* major timeout alarm */
        unsigned long           rq_timeout;     /* Current timeout value */
        unsigned int            rq_retries;     /* # of retries */
-       /*
-        * For authentication (e.g. auth_des)
-        */
-       u32                     rq_creddata[2];
        
        /*
         * Partial send handling
         */
-       
        u32                     rq_bytes_sent;  /* Bytes we have sent */
 
        unsigned long           rq_xtime;       /* when transmitted */
@@ -107,6 +105,8 @@ struct rpc_xprt_ops {
        int             (*reserve_xprt)(struct rpc_task *task);
        void            (*release_xprt)(struct rpc_xprt *xprt, struct rpc_task *task);
        void            (*connect)(struct rpc_task *task);
+       void *          (*buf_alloc)(struct rpc_task *task, size_t size);
+       void            (*buf_free)(struct rpc_task *task);
        int             (*send_request)(struct rpc_task *task);
        void            (*set_retrans_timeout)(struct rpc_task *task);
        void            (*timer)(struct rpc_task *task);
index b23c0d328c9c958968840db0b58e542d6b8c84ea..25cba94c5683c4fae35aa398088be222dcd29ed6 100644 (file)
@@ -644,24 +644,26 @@ call_reserveresult(struct rpc_task *task)
 
 /*
  * 2.  Allocate the buffer. For details, see sched.c:rpc_malloc.
- *     (Note: buffer memory is freed in rpc_task_release).
+ *     (Note: buffer memory is freed in xprt_release).
  */
 static void
 call_allocate(struct rpc_task *task)
 {
+       struct rpc_rqst *req = task->tk_rqstp;
+       struct rpc_xprt *xprt = task->tk_xprt;
        unsigned int    bufsiz;
 
        dprintk("RPC: %4d call_allocate (status %d)\n", 
                                task->tk_pid, task->tk_status);
        task->tk_action = call_bind;
-       if (task->tk_buffer)
+       if (req->rq_buffer)
                return;
 
        /* FIXME: compute buffer requirements more exactly using
         * auth->au_wslack */
        bufsiz = task->tk_msg.rpc_proc->p_bufsiz + RPC_SLACK_SPACE;
 
-       if (rpc_malloc(task, bufsiz << 1) != NULL)
+       if (xprt->ops->buf_alloc(task, bufsiz << 1) != NULL)
                return;
        printk(KERN_INFO "RPC: buffer allocation failed for task %p\n", task); 
 
@@ -704,14 +706,14 @@ call_encode(struct rpc_task *task)
                                task->tk_pid, task->tk_status);
 
        /* Default buffer setup */
-       bufsiz = task->tk_bufsize >> 1;
-       sndbuf->head[0].iov_base = (void *)task->tk_buffer;
+       bufsiz = req->rq_bufsize >> 1;
+       sndbuf->head[0].iov_base = (void *)req->rq_buffer;
        sndbuf->head[0].iov_len  = bufsiz;
        sndbuf->tail[0].iov_len  = 0;
        sndbuf->page_len         = 0;
        sndbuf->len              = 0;
        sndbuf->buflen           = bufsiz;
-       rcvbuf->head[0].iov_base = (void *)((char *)task->tk_buffer + bufsiz);
+       rcvbuf->head[0].iov_base = (void *)((char *)req->rq_buffer + bufsiz);
        rcvbuf->head[0].iov_len  = bufsiz;
        rcvbuf->tail[0].iov_len  = 0;
        rcvbuf->page_len         = 0;
index 48510e3ffa02a3aff8ad0714011ab5433494b971..7415406aa1ae510639e24bc103ea42d3a862db03 100644 (file)
@@ -41,8 +41,6 @@ static mempool_t      *rpc_buffer_mempool __read_mostly;
 
 static void                    __rpc_default_timer(struct rpc_task *task);
 static void                    rpciod_killall(void);
-static void                    rpc_free(struct rpc_task *task);
-
 static void                    rpc_async_schedule(void *);
 
 /*
@@ -599,7 +597,6 @@ void rpc_exit_task(struct rpc_task *task)
                        WARN_ON(RPC_ASSASSINATED(task));
                        /* Always release the RPC slot and buffer memory */
                        xprt_release(task);
-                       rpc_free(task);
                }
        }
 }
@@ -724,17 +721,19 @@ static void rpc_async_schedule(void *arg)
        __rpc_execute((struct rpc_task *)arg);
 }
 
-/*
- * Allocate memory for RPC purposes.
+/**
+ * rpc_malloc - allocate an RPC buffer
+ * @task: RPC task that will use this buffer
+ * @size: requested byte size
  *
  * We try to ensure that some NFS reads and writes can always proceed
  * by using a mempool when allocating 'small' buffers.
  * In order to avoid memory starvation triggering more writebacks of
  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
  */
-void *
-rpc_malloc(struct rpc_task *task, size_t size)
+void * rpc_malloc(struct rpc_task *task, size_t size)
 {
+       struct rpc_rqst *req = task->tk_rqstp;
        gfp_t   gfp;
 
        if (task->tk_flags & RPC_TASK_SWAPPER)
@@ -743,27 +742,33 @@ rpc_malloc(struct rpc_task *task, size_t size)
                gfp = GFP_NOFS;
 
        if (size > RPC_BUFFER_MAXSIZE) {
-               task->tk_buffer =  kmalloc(size, gfp);
-               if (task->tk_buffer)
-                       task->tk_bufsize = size;
+               req->rq_buffer = kmalloc(size, gfp);
+               if (req->rq_buffer)
+                       req->rq_bufsize = size;
        } else {
-               task->tk_buffer =  mempool_alloc(rpc_buffer_mempool, gfp);
-               if (task->tk_buffer)
-                       task->tk_bufsize = RPC_BUFFER_MAXSIZE;
+               req->rq_buffer = mempool_alloc(rpc_buffer_mempool, gfp);
+               if (req->rq_buffer)
+                       req->rq_bufsize = RPC_BUFFER_MAXSIZE;
        }
-       return task->tk_buffer;
+       return req->rq_buffer;
 }
 
-static void
-rpc_free(struct rpc_task *task)
+/**
+ * rpc_free - free buffer allocated via rpc_malloc
+ * @task: RPC task with a buffer to be freed
+ *
+ */
+void rpc_free(struct rpc_task *task)
 {
-       if (task->tk_buffer) {
-               if (task->tk_bufsize == RPC_BUFFER_MAXSIZE)
-                       mempool_free(task->tk_buffer, rpc_buffer_mempool);
+       struct rpc_rqst *req = task->tk_rqstp;
+
+       if (req->rq_buffer) {
+               if (req->rq_bufsize == RPC_BUFFER_MAXSIZE)
+                       mempool_free(req->rq_buffer, rpc_buffer_mempool);
                else
-                       kfree(task->tk_buffer);
-               task->tk_buffer = NULL;
-               task->tk_bufsize = 0;
+                       kfree(req->rq_buffer);
+               req->rq_buffer = NULL;
+               req->rq_bufsize = 0;
        }
 }
 
@@ -887,7 +892,6 @@ void rpc_release_task(struct rpc_task *task)
                xprt_release(task);
        if (task->tk_msg.rpc_cred)
                rpcauth_unbindcred(task);
-       rpc_free(task);
        if (task->tk_client) {
                rpc_release_client(task->tk_client);
                task->tk_client = NULL;
index 6dda3860351fb1502f5ee12d0cc7e00e7aebb060..069a6cbd49ea5c5e11ff1bb8bc73df0fecf117da 100644 (file)
@@ -838,6 +838,8 @@ static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
        req->rq_timeout = xprt->timeout.to_initval;
        req->rq_task    = task;
        req->rq_xprt    = xprt;
+       req->rq_buffer  = NULL;
+       req->rq_bufsize = 0;
        req->rq_xid     = xprt_alloc_xid(xprt);
        req->rq_release_snd_buf = NULL;
        dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
@@ -867,6 +869,7 @@ void xprt_release(struct rpc_task *task)
                mod_timer(&xprt->timer,
                                xprt->last_used + xprt->idle_timeout);
        spin_unlock_bh(&xprt->transport_lock);
+       xprt->ops->buf_free(task);
        task->tk_rqstp = NULL;
        if (req->rq_release_snd_buf)
                req->rq_release_snd_buf(req);
index 77e8800d412799c088bb14d25538093216e27232..51f07c9a751b64e5f4e0342ae48e7067edf37888 100644 (file)
@@ -28,6 +28,7 @@
 #include <linux/udp.h>
 #include <linux/tcp.h>
 #include <linux/sunrpc/clnt.h>
+#include <linux/sunrpc/sched.h>
 #include <linux/file.h>
 
 #include <net/sock.h>
@@ -1161,6 +1162,8 @@ static struct rpc_xprt_ops xs_udp_ops = {
        .reserve_xprt           = xprt_reserve_xprt_cong,
        .release_xprt           = xprt_release_xprt_cong,
        .connect                = xs_connect,
+       .buf_alloc              = rpc_malloc,
+       .buf_free               = rpc_free,
        .send_request           = xs_udp_send_request,
        .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
        .timer                  = xs_udp_timer,
@@ -1173,6 +1176,8 @@ static struct rpc_xprt_ops xs_tcp_ops = {
        .reserve_xprt           = xprt_reserve_xprt,
        .release_xprt           = xprt_release_xprt,
        .connect                = xs_connect,
+       .buf_alloc              = rpc_malloc,
+       .buf_free               = rpc_free,
        .send_request           = xs_tcp_send_request,
        .set_retrans_timeout    = xprt_set_retrans_timeout_def,
        .close                  = xs_close,