Linux Virtualization - Jan 2013 - [PATCH 0/3] tcm_vhost: Optimize gup in vhost_scsi_map_to_sgl

Asias He (3):
  tcm_vhost: Introduce iov_num_pages
  tcm_vhost: Optimize gup in vhost_scsi_map_to_sgl
  tcm_vhost: Use iov_num_pages to calculate sgl_count

 drivers/vhost/tcm_vhost.c | 47 +++++++++++++++++++++++++++++------------------
 1 file changed, 29 insertions(+), 18 deletions(-)

-- 
1.8.1

Add a helper to calculate the number of pages needed for a iov entry.

Signed-off-by: Asias He <asias at redhat.com>
---
 drivers/vhost/tcm_vhost.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/drivers/vhost/tcm_vhost.c b/drivers/vhost/tcm_vhost.c
index 3720604..ca35c16 100644
--- a/drivers/vhost/tcm_vhost.c
+++ b/drivers/vhost/tcm_vhost.c
@@ -77,6 +77,12 @@ static struct workqueue_struct *tcm_vhost_workqueue;
 static DEFINE_MUTEX(tcm_vhost_mutex);
 static LIST_HEAD(tcm_vhost_list);
 
+static inline int iov_num_pages(struct iovec *iov)
+{
+	return (PAGE_ALIGN((unsigned long)iov->iov_base + iov->iov_len) -
+	       ((unsigned long)iov->iov_base & PAGE_MASK)) >> PAGE_SHIFT;
+}
+
 static int tcm_vhost_check_true(struct se_portal_group *se_tpg)
 {
 	return 1;
-- 
1.8.1

We can get all the pages in one time instead of calling
gup N times.

Signed-off-by: Asias He <asias at redhat.com>
---
 drivers/vhost/tcm_vhost.c | 33 ++++++++++++++++++++-------------
 1 file changed, 20 insertions(+), 13 deletions(-)

diff --git a/drivers/vhost/tcm_vhost.c b/drivers/vhost/tcm_vhost.c
index ca35c16..59be442 100644
--- a/drivers/vhost/tcm_vhost.c
+++ b/drivers/vhost/tcm_vhost.c
@@ -430,37 +430,45 @@ static struct tcm_vhost_cmd *vhost_scsi_allocate_cmd(
  * Returns the number of scatterlist entries used or -errno on error.
  */
 static int vhost_scsi_map_to_sgl(struct scatterlist *sgl,
-	unsigned int sgl_count, void __user *ptr, size_t len, int write)
+	unsigned int sgl_count, struct iovec *iov, int write)
 {
 	struct scatterlist *sg = sgl;
 	unsigned int npages = 0;
+	void __user *ptr = iov->iov_base;
+	size_t len = iov->iov_len;
 	int ret;
+	unsigned int pages_nr, offset, nbytes;
+	struct page **pages;
+
+	pages_nr = iov_num_pages(iov);
+	pages = kmalloc(pages_nr * sizeof(struct page *), GFP_ATOMIC);
+	if (!pages)
+		return -ENOMEM;
+
+	ret = get_user_pages_fast((unsigned long)ptr, pages_nr, write, pages);
+	if (ret != pages_nr)
+		goto err;
 
 	while (len > 0) {
-		struct page *page;
-		unsigned int offset = (uintptr_t)ptr & ~PAGE_MASK;
-		unsigned int nbytes = min_t(unsigned int,
-				PAGE_SIZE - offset, len);
+		offset = (uintptr_t)ptr & ~PAGE_MASK;
+		nbytes = min_t(unsigned int, PAGE_SIZE - offset, len);
 
 		if (npages == sgl_count) {
 			ret = -ENOBUFS;
 			goto err;
 		}
 
-		ret = get_user_pages_fast((unsigned long)ptr, 1, write, &page);
-		BUG_ON(ret == 0); /* we should either get our page or fail */
-		if (ret < 0)
-			goto err;
-
-		sg_set_page(sg, page, nbytes, offset);
+		sg_set_page(sg, pages[npages], nbytes, offset);
 		ptr += nbytes;
 		len -= nbytes;
 		sg++;
 		npages++;
 	}
+	kfree(pages);
 	return npages;
 
 err:
+	kfree(pages);
 	/* Put pages that we hold */
 	for (sg = sgl; sg != &sgl[npages]; sg++)
 		put_page(sg_page(sg));
@@ -498,8 +506,7 @@ static int vhost_scsi_map_iov_to_sgl(struct tcm_vhost_cmd
*tv_cmd,
 
 	pr_debug("Mapping %u iovecs for %u pages\n", niov, sgl_count);
 	for (i = 0; i < niov; i++) {
-		ret = vhost_scsi_map_to_sgl(sg, sgl_count, iov[i].iov_base,
-					iov[i].iov_len, write);
+		ret = vhost_scsi_map_to_sgl(sg, sgl_count, &iov[i], write);
 		if (ret < 0) {
 			for (i = 0; i < tv_cmd->tvc_sgl_count; i++)
 				put_page(sg_page(&tv_cmd->tvc_sgl[i]));
-- 
1.8.1

Signed-off-by: Asias He <asias at redhat.com>
---
 drivers/vhost/tcm_vhost.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/drivers/vhost/tcm_vhost.c b/drivers/vhost/tcm_vhost.c
index 59be442..5b77469 100644
--- a/drivers/vhost/tcm_vhost.c
+++ b/drivers/vhost/tcm_vhost.c
@@ -487,11 +487,9 @@ static int vhost_scsi_map_iov_to_sgl(struct tcm_vhost_cmd
*tv_cmd,
 	 * Find out how long sglist needs to be
 	 */
 	sgl_count = 0;
-	for (i = 0; i < niov; i++) {
-		sgl_count += (((uintptr_t)iov[i].iov_base + iov[i].iov_len +
-				PAGE_SIZE - 1) >> PAGE_SHIFT) -
-				((uintptr_t)iov[i].iov_base >> PAGE_SHIFT);
-	}
+	for (i = 0; i < niov; i++)
+		sgl_count += iov_num_pages(&iov[i]);
+
 	/* TODO overflow checking */
 
 	sg = kmalloc(sizeof(tv_cmd->tvc_sgl[0]) * sgl_count, GFP_ATOMIC);
-- 
1.8.1

On Mon, Jan 21, 2013 at 04:05:27PM +0800, Asias He
wrote:
> We can get all the pages in one time instead of calling
> gup N times.
> 
> Signed-off-by: Asias He <asias at redhat.com>
> ---
>  drivers/vhost/tcm_vhost.c | 33 ++++++++++++++++++++-------------
>  1 file changed, 20 insertions(+), 13 deletions(-)
> 
> diff --git a/drivers/vhost/tcm_vhost.c b/drivers/vhost/tcm_vhost.c
> index ca35c16..59be442 100644
> --- a/drivers/vhost/tcm_vhost.c
> +++ b/drivers/vhost/tcm_vhost.c
> @@ -430,37 +430,45 @@ static struct tcm_vhost_cmd *vhost_scsi_allocate_cmd(
>   * Returns the number of scatterlist entries used or -errno on error.
>   */
>  static int vhost_scsi_map_to_sgl(struct scatterlist *sgl,
> -	unsigned int sgl_count, void __user *ptr, size_t len, int write)
> +	unsigned int sgl_count, struct iovec *iov, int write)
>  {
>  	struct scatterlist *sg = sgl;
>  	unsigned int npages = 0;
> +	void __user *ptr = iov->iov_base;
> +	size_t len = iov->iov_len;
>  	int ret;
> +	unsigned int pages_nr, offset, nbytes;
> +	struct page **pages;
> +
> +	pages_nr = iov_num_pages(iov);
> +	pages = kmalloc(pages_nr * sizeof(struct page *), GFP_ATOMIC);
> +	if (!pages)
> +		return -ENOMEM;
> +
> +	ret = get_user_pages_fast((unsigned long)ptr, pages_nr, write, pages);
> +	if (ret != pages_nr)
> +		goto err;
1. Why GFP_ATOMIC? get_user_pages_fast can sleep, so this path must not
be atomic (if it is, should use __get_user_pages_fast).

GFP_ATOMIC should be avoided.

2. Should drop reference to pages whose refcount has been increased,
if ret > 0 && ret != pages_nr (see last phrase of get_user_pages_fast
commentary).

On 01/22/2013 02:57 AM, Marcelo Tosatti wrote:
> On Mon, Jan 21, 2013 at 04:05:27PM +0800, Asias He wrote:
>> We can get all the pages in one time instead of calling
>> gup N times.
>>
>> Signed-off-by: Asias He <asias at redhat.com>
>> ---
>>  drivers/vhost/tcm_vhost.c | 33 ++++++++++++++++++++-------------
>>  1 file changed, 20 insertions(+), 13 deletions(-)
>>
>> diff --git a/drivers/vhost/tcm_vhost.c b/drivers/vhost/tcm_vhost.c
>> index ca35c16..59be442 100644
>> --- a/drivers/vhost/tcm_vhost.c
>> +++ b/drivers/vhost/tcm_vhost.c
>> @@ -430,37 +430,45 @@ static struct tcm_vhost_cmd
*vhost_scsi_allocate_cmd(
>>   * Returns the number of scatterlist entries used or -errno on error.
>>   */
>>  static int vhost_scsi_map_to_sgl(struct scatterlist *sgl,
>> -	unsigned int sgl_count, void __user *ptr, size_t len, int write)
>> +	unsigned int sgl_count, struct iovec *iov, int write)
>>  {
>>  	struct scatterlist *sg = sgl;
>>  	unsigned int npages = 0;
>> +	void __user *ptr = iov->iov_base;
>> +	size_t len = iov->iov_len;
>>  	int ret;
>> +	unsigned int pages_nr, offset, nbytes;
>> +	struct page **pages;
>> +
>> +	pages_nr = iov_num_pages(iov);
>> +	pages = kmalloc(pages_nr * sizeof(struct page *), GFP_ATOMIC);
>> +	if (!pages)
>> +		return -ENOMEM;
>> +
>> +	ret = get_user_pages_fast((unsigned long)ptr, pages_nr, write,
pages);
>> +	if (ret != pages_nr)
>> +		goto err;
> 
> 1. Why GFP_ATOMIC? get_user_pages_fast can sleep, so this path must not
> be atomic (if it is, should use __get_user_pages_fast).
> 
> GFP_ATOMIC should be avoided.
> 
> 2. Should drop reference to pages whose refcount has been increased,
> if ret > 0 && ret != pages_nr (see last phrase of
get_user_pages_fast
> commentary).
Thanks Marcelo. V2 is on the way.

-- 
Asias