]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
virtio/console: Add splice_write support
authorMasami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Thu, 9 Aug 2012 12:30:39 +0000 (21:30 +0900)
committerStephen Rothwell <sfr@canb.auug.org.au>
Wed, 22 Aug 2012 00:47:10 +0000 (10:47 +1000)
Enable to use splice_write from pipe to virtio-console port.
This steals pages from pipe and directly send it to host.

Note that this may accelerate only the guest to host path.

Changes in v2:
 - Use GFP_KERNEL instead of GFP_ATOMIC in syscall context function.

Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
drivers/char/virtio_console.c

index cdf2f5451c767ff0584c3739e8a35e4327cc52e1..730816cdeb45c46fb0517ceccec72ecb5993aaa0 100644 (file)
@@ -24,6 +24,8 @@
 #include <linux/err.h>
 #include <linux/freezer.h>
 #include <linux/fs.h>
+#include <linux/splice.h>
+#include <linux/pagemap.h>
 #include <linux/init.h>
 #include <linux/list.h>
 #include <linux/poll.h>
@@ -227,6 +229,7 @@ struct port {
        bool guest_connected;
 };
 
+#define MAX_SPLICE_PAGES       32
 /* This is the very early arch-specified put chars function. */
 static int (*early_put_chars)(u32, const char *, int);
 
@@ -474,26 +477,52 @@ static ssize_t send_control_msg(struct port *port, unsigned int event,
        return 0;
 }
 
+struct buffer_token {
+       union {
+               void *buf;
+               struct scatterlist *sg;
+       } u;
+       bool sgpages;
+};
+
+static void reclaim_sg_pages(struct scatterlist *sg)
+{
+       int i;
+       struct page *page;
+
+       for (i = 0; i < MAX_SPLICE_PAGES; i++) {
+               page = sg_page(&sg[i]);
+               if (!page)
+                       break;
+               put_page(page);
+       }
+       kfree(sg);
+}
+
 /* Callers must take the port->outvq_lock */
 static void reclaim_consumed_buffers(struct port *port)
 {
-       void *buf;
+       struct buffer_token *tok;
        unsigned int len;
 
        if (!port->portdev) {
                /* Device has been unplugged.  vqs are already gone. */
                return;
        }
-       while ((buf = virtqueue_get_buf(port->out_vq, &len))) {
-               kfree(buf);
+       while ((tok = virtqueue_get_buf(port->out_vq, &len))) {
+               if (tok->sgpages)
+                       reclaim_sg_pages(tok->u.sg);
+               else
+                       kfree(tok->u.buf);
+               kfree(tok);
                port->outvq_full = false;
        }
 }
 
-static ssize_t send_buf(struct port *port, void *in_buf, size_t in_count,
-                       bool nonblock)
+static ssize_t __send_to_port(struct port *port, struct scatterlist *sg,
+                             int nents, size_t in_count,
+                             struct buffer_token *tok, bool nonblock)
 {
-       struct scatterlist sg[1];
        struct virtqueue *out_vq;
        ssize_t ret;
        unsigned long flags;
@@ -505,8 +534,7 @@ static ssize_t send_buf(struct port *port, void *in_buf, size_t in_count,
 
        reclaim_consumed_buffers(port);
 
-       sg_init_one(sg, in_buf, in_count);
-       ret = virtqueue_add_buf(out_vq, sg, 1, 0, in_buf, GFP_ATOMIC);
+       ret = virtqueue_add_buf(out_vq, sg, nents, 0, tok, GFP_ATOMIC);
 
        /* Tell Host to go! */
        virtqueue_kick(out_vq);
@@ -544,6 +572,37 @@ done:
        return in_count;
 }
 
+static ssize_t send_buf(struct port *port, void *in_buf, size_t in_count,
+                       bool nonblock)
+{
+       struct scatterlist sg[1];
+       struct buffer_token *tok;
+
+       tok = kmalloc(sizeof(*tok), GFP_ATOMIC);
+       if (!tok)
+               return -ENOMEM;
+       tok->sgpages = false;
+       tok->u.buf = in_buf;
+
+       sg_init_one(sg, in_buf, in_count);
+
+       return __send_to_port(port, sg, 1, in_count, tok, nonblock);
+}
+
+static ssize_t send_pages(struct port *port, struct scatterlist *sg, int nents,
+                         size_t in_count, bool nonblock)
+{
+       struct buffer_token *tok;
+
+       tok = kmalloc(sizeof(*tok), GFP_ATOMIC);
+       if (!tok)
+               return -ENOMEM;
+       tok->sgpages = true;
+       tok->u.sg = sg;
+
+       return __send_to_port(port, sg, nents, in_count, tok, nonblock);
+}
+
 /*
  * Give out the data that's requested from the buffer that we have
  * queued up.
@@ -725,6 +784,66 @@ out:
        return ret;
 }
 
+struct sg_list {
+       unsigned int n;
+       size_t len;
+       struct scatterlist *sg;
+};
+
+static int pipe_to_sg(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
+                       struct splice_desc *sd)
+{
+       struct sg_list *sgl = sd->u.data;
+       unsigned int len = 0;
+
+       if (sgl->n == MAX_SPLICE_PAGES)
+               return 0;
+
+       /* Try lock this page */
+       if (buf->ops->steal(pipe, buf) == 0) {
+               /* Get reference and unlock page for moving */
+               get_page(buf->page);
+               unlock_page(buf->page);
+
+               len = min(buf->len, sd->len);
+               sg_set_page(&(sgl->sg[sgl->n]), buf->page, len, buf->offset);
+               sgl->n++;
+               sgl->len += len;
+       }
+
+       return len;
+}
+
+/* Faster zero-copy write by splicing */
+static ssize_t port_fops_splice_write(struct pipe_inode_info *pipe,
+                                     struct file *filp, loff_t *ppos,
+                                     size_t len, unsigned int flags)
+{
+       struct port *port = filp->private_data;
+       struct sg_list sgl;
+       ssize_t ret;
+       struct splice_desc sd = {
+               .total_len = len,
+               .flags = flags,
+               .pos = *ppos,
+               .u.data = &sgl,
+       };
+
+       sgl.n = 0;
+       sgl.len = 0;
+       sgl.sg = kmalloc(sizeof(struct scatterlist) * MAX_SPLICE_PAGES,
+                        GFP_KERNEL);
+       if (unlikely(!sgl.sg))
+               return -ENOMEM;
+
+       sg_init_table(sgl.sg, MAX_SPLICE_PAGES);
+       ret = __splice_from_pipe(pipe, &sd, pipe_to_sg);
+       if (likely(ret > 0))
+               ret = send_pages(port, sgl.sg, sgl.n, sgl.len, true);
+
+       return ret;
+}
+
 static unsigned int port_fops_poll(struct file *filp, poll_table *wait)
 {
        struct port *port;
@@ -856,6 +975,7 @@ static const struct file_operations port_fops = {
        .open  = port_fops_open,
        .read  = port_fops_read,
        .write = port_fops_write,
+       .splice_write = port_fops_splice_write,
        .poll  = port_fops_poll,
        .release = port_fops_release,
        .fasync = port_fops_fasync,