2013-09-17 07:19:03

by Michael S. Tsirkin

[permalink] [raw]
Subject: [PATCH] vhost/scsi: use vmalloc for order-10 allocation

As vhost scsi device struct is large, if the device is
created on a busy system, kzalloc() might fail, so this patch does a
fallback to vzalloc().

As vmalloc() adds overhead on data-path, add __GFP_REPEAT
to kzalloc() flags to do this fallback only when really needed.

Reported-by: Dan Aloni <[email protected]>
Signed-off-by: Michael S. Tsirkin <[email protected]>
---

I put this on my vhost fixes branch, intend to merge for 3.12.
Dan, could you please confirm this works for you?

drivers/vhost/scsi.c | 41 +++++++++++++++++++++++++++--------------
1 file changed, 27 insertions(+), 14 deletions(-)

diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
index 4b79a1f..2c30bb0 100644
--- a/drivers/vhost/scsi.c
+++ b/drivers/vhost/scsi.c
@@ -1373,21 +1373,30 @@ static int vhost_scsi_set_features(struct vhost_scsi *vs, u64 features)
return 0;
}

+static void vhost_scsi_free(struct vhost_scsi *vs)
+{
+ if (is_vmalloc_addr(vs))
+ vfree(vs);
+ else
+ kfree(vs);
+}
+
static int vhost_scsi_open(struct inode *inode, struct file *f)
{
struct vhost_scsi *vs;
struct vhost_virtqueue **vqs;
- int r, i;
+ int r = -ENOMEM, i;

- vs = kzalloc(sizeof(*vs), GFP_KERNEL);
- if (!vs)
- return -ENOMEM;
+ vs = kzalloc(sizeof(*vs), GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
+ if (!vs) {
+ vs = vzalloc(sizeof(*vs));
+ if (!vs)
+ goto err_vs;
+ }

vqs = kmalloc(VHOST_SCSI_MAX_VQ * sizeof(*vqs), GFP_KERNEL);
- if (!vqs) {
- kfree(vs);
- return -ENOMEM;
- }
+ if (!vqs)
+ goto err_vqs;

vhost_work_init(&vs->vs_completion_work, vhost_scsi_complete_cmd_work);
vhost_work_init(&vs->vs_event_work, tcm_vhost_evt_work);
@@ -1407,14 +1416,18 @@ static int vhost_scsi_open(struct inode *inode, struct file *f)

tcm_vhost_init_inflight(vs, NULL);

- if (r < 0) {
- kfree(vqs);
- kfree(vs);
- return r;
- }
+ if (r < 0)
+ goto err_init;

f->private_data = vs;
return 0;
+
+err_init:
+ kfree(vqs);
+err_vqs:
+ vhost_scsi_free(vs);
+err_vs:
+ return r;
}

static int vhost_scsi_release(struct inode *inode, struct file *f)
@@ -1431,7 +1444,7 @@ static int vhost_scsi_release(struct inode *inode, struct file *f)
/* Jobs can re-queue themselves in evt kick handler. Do extra flush. */
vhost_scsi_flush(vs);
kfree(vs->dev.vqs);
- kfree(vs);
+ vhost_scsi_free(vs);
return 0;
}

--
MST


2013-09-17 08:55:38

by Asias He

[permalink] [raw]
Subject: Re: [PATCH] vhost/scsi: use vmalloc for order-10 allocation

On Tue, Sep 17, 2013 at 10:21:07AM +0300, Michael S. Tsirkin wrote:
> As vhost scsi device struct is large, if the device is
> created on a busy system, kzalloc() might fail, so this patch does a
> fallback to vzalloc().
>
> As vmalloc() adds overhead on data-path, add __GFP_REPEAT
> to kzalloc() flags to do this fallback only when really needed.
>
> Reported-by: Dan Aloni <[email protected]>
> Signed-off-by: Michael S. Tsirkin <[email protected]>

Reviewed-by: Asias He <[email protected]>

> ---
>
> I put this on my vhost fixes branch, intend to merge for 3.12.
> Dan, could you please confirm this works for you?
>
> drivers/vhost/scsi.c | 41 +++++++++++++++++++++++++++--------------
> 1 file changed, 27 insertions(+), 14 deletions(-)
>
> diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
> index 4b79a1f..2c30bb0 100644
> --- a/drivers/vhost/scsi.c
> +++ b/drivers/vhost/scsi.c
> @@ -1373,21 +1373,30 @@ static int vhost_scsi_set_features(struct vhost_scsi *vs, u64 features)
> return 0;
> }
>
> +static void vhost_scsi_free(struct vhost_scsi *vs)
> +{
> + if (is_vmalloc_addr(vs))
> + vfree(vs);
> + else
> + kfree(vs);
> +}
> +
> static int vhost_scsi_open(struct inode *inode, struct file *f)
> {
> struct vhost_scsi *vs;
> struct vhost_virtqueue **vqs;
> - int r, i;
> + int r = -ENOMEM, i;
>
> - vs = kzalloc(sizeof(*vs), GFP_KERNEL);
> - if (!vs)
> - return -ENOMEM;
> + vs = kzalloc(sizeof(*vs), GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
> + if (!vs) {
> + vs = vzalloc(sizeof(*vs));
> + if (!vs)
> + goto err_vs;
> + }
>
> vqs = kmalloc(VHOST_SCSI_MAX_VQ * sizeof(*vqs), GFP_KERNEL);
> - if (!vqs) {
> - kfree(vs);
> - return -ENOMEM;
> - }
> + if (!vqs)
> + goto err_vqs;
>
> vhost_work_init(&vs->vs_completion_work, vhost_scsi_complete_cmd_work);
> vhost_work_init(&vs->vs_event_work, tcm_vhost_evt_work);
> @@ -1407,14 +1416,18 @@ static int vhost_scsi_open(struct inode *inode, struct file *f)
>
> tcm_vhost_init_inflight(vs, NULL);
>
> - if (r < 0) {
> - kfree(vqs);
> - kfree(vs);
> - return r;
> - }
> + if (r < 0)
> + goto err_init;
>
> f->private_data = vs;
> return 0;
> +
> +err_init:
> + kfree(vqs);
> +err_vqs:
> + vhost_scsi_free(vs);
> +err_vs:
> + return r;
> }
>
> static int vhost_scsi_release(struct inode *inode, struct file *f)
> @@ -1431,7 +1444,7 @@ static int vhost_scsi_release(struct inode *inode, struct file *f)
> /* Jobs can re-queue themselves in evt kick handler. Do extra flush. */
> vhost_scsi_flush(vs);
> kfree(vs->dev.vqs);
> - kfree(vs);
> + vhost_scsi_free(vs);
> return 0;
> }
>
> --
> MST

--
Asias

2013-09-17 18:14:41

by Sergei Shtylyov

[permalink] [raw]
Subject: Re: [PATCH] vhost/scsi: use vmalloc for order-10 allocation

Hello.

On 09/17/2013 11:21 AM, Michael S. Tsirkin wrote:

> As vhost scsi device struct is large, if the device is
> created on a busy system, kzalloc() might fail, so this patch does a
> fallback to vzalloc().

> As vmalloc() adds overhead on data-path, add __GFP_REPEAT
> to kzalloc() flags to do this fallback only when really needed.

> Reported-by: Dan Aloni <[email protected]>
> Signed-off-by: Michael S. Tsirkin <[email protected]>
> ---

> I put this on my vhost fixes branch, intend to merge for 3.12.
> Dan, could you please confirm this works for you?

> drivers/vhost/scsi.c | 41 +++++++++++++++++++++++++++--------------
> 1 file changed, 27 insertions(+), 14 deletions(-)

> diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
> index 4b79a1f..2c30bb0 100644
> --- a/drivers/vhost/scsi.c
> +++ b/drivers/vhost/scsi.c
> @@ -1373,21 +1373,30 @@ static int vhost_scsi_set_features(struct vhost_scsi *vs, u64 features)
> return 0;
> }
>
> +static void vhost_scsi_free(struct vhost_scsi *vs)
> +{
> + if (is_vmalloc_addr(vs))
> + vfree(vs);
> + else
> + kfree(vs);

Indent with the tabs ISO spaces, please.

> +}
> +
> static int vhost_scsi_open(struct inode *inode, struct file *f)
> {
> struct vhost_scsi *vs;
> struct vhost_virtqueue **vqs;
> - int r, i;
> + int r = -ENOMEM, i;
>
> - vs = kzalloc(sizeof(*vs), GFP_KERNEL);
> - if (!vs)
> - return -ENOMEM;
> + vs = kzalloc(sizeof(*vs), GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);

Indent here with a tab, please.

> + if (!vs) {
> + vs = vzalloc(sizeof(*vs));
> + if (!vs)
> + goto err_vs;
> + }

WBR, Sergei

2013-09-17 19:49:51

by Michael S. Tsirkin

[permalink] [raw]
Subject: Re: [PATCH] vhost/scsi: use vmalloc for order-10 allocation

On Tue, Sep 17, 2013 at 10:14:37PM +0400, Sergei Shtylyov wrote:
> Hello.
>
> On 09/17/2013 11:21 AM, Michael S. Tsirkin wrote:
>
> >As vhost scsi device struct is large, if the device is
> >created on a busy system, kzalloc() might fail, so this patch does a
> >fallback to vzalloc().
>
> >As vmalloc() adds overhead on data-path, add __GFP_REPEAT
> >to kzalloc() flags to do this fallback only when really needed.
>
> >Reported-by: Dan Aloni <[email protected]>
> >Signed-off-by: Michael S. Tsirkin <[email protected]>
> >---
>
> >I put this on my vhost fixes branch, intend to merge for 3.12.
> >Dan, could you please confirm this works for you?
>
> > drivers/vhost/scsi.c | 41 +++++++++++++++++++++++++++--------------
> > 1 file changed, 27 insertions(+), 14 deletions(-)
>
> >diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
> >index 4b79a1f..2c30bb0 100644
> >--- a/drivers/vhost/scsi.c
> >+++ b/drivers/vhost/scsi.c
> >@@ -1373,21 +1373,30 @@ static int vhost_scsi_set_features(struct vhost_scsi *vs, u64 features)
> > return 0;
> > }
> >
> >+static void vhost_scsi_free(struct vhost_scsi *vs)
> >+{
> >+ if (is_vmalloc_addr(vs))
> >+ vfree(vs);
> >+ else
> >+ kfree(vs);
>
> Indent with the tabs ISO spaces, please.
>
> >+}
> >+
> > static int vhost_scsi_open(struct inode *inode, struct file *f)
> > {
> > struct vhost_scsi *vs;
> > struct vhost_virtqueue **vqs;
> >- int r, i;
> >+ int r = -ENOMEM, i;
> >
> >- vs = kzalloc(sizeof(*vs), GFP_KERNEL);
> >- if (!vs)
> >- return -ENOMEM;
> >+ vs = kzalloc(sizeof(*vs), GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
>
> Indent here with a tab, please.
>
> >+ if (!vs) {
> >+ vs = vzalloc(sizeof(*vs));
> >+ if (!vs)
> >+ goto err_vs;
> >+ }
>
> WBR, Sergei

Thanks, I'll fix this up.