[PATCH v2]Resend:videobuf_dma_sg: a new implementation for mmap

[PATCH v2]Resend:videobuf_dma_sg: a new implementation for mmap

[Figo.zhang]

a mmap issue for videobuf-dma-sg: it will alloc a new page for mmaping when it encounter
page fault at video_vm_ops->fault(). pls see http://www.spinics.net/lists/linux-media/msg21243.html

a new implementation for mmap, it translate to vmalloc to page at video_vm_ops->fault().

in v2, if mem->dma.vmalloc is NULL at video_vm_ops->fault(), it will alloc memory by 
vmlloc_32().

Signed-off-by: Figo.zhang <figo1802@gmail.com>
---
drivers/media/video/videobuf-dma-sg.c |   50 +++++++++++++++++++++++++++------
 1 files changed, 41 insertions(+), 9 deletions(-)

diff --git a/drivers/media/video/videobuf-dma-sg.c b/drivers/media/video/videobuf-dma-sg.c
index 8359e6b..f7295da 100644
--- a/drivers/media/video/videobuf-dma-sg.c
+++ b/drivers/media/video/videobuf-dma-sg.c
@@ -201,10 +201,11 @@ int videobuf_dma_init_kernel(struct videobuf_dmabuf *dma, int direction,
 	dprintk(1, "init kernel [%d pages]\n", nr_pages);
 
 	dma->direction = direction;
-	dma->vmalloc = vmalloc_32(nr_pages << PAGE_SHIFT);
-	if (NULL == dma->vmalloc) {
-		dprintk(1, "vmalloc_32(%d pages) failed\n", nr_pages);
-		return -ENOMEM;
+	if (!dma->vmalloc)
+		dma->vmalloc = vmalloc_32(nr_pages << PAGE_SHIFT);
+		if (NULL == dma->vmalloc) {
+			dprintk(1, "vmalloc_32(%d pages) failed\n", nr_pages);
+			return -ENOMEM;
 	}
 
 	dprintk(1, "vmalloc is at addr 0x%08lx, size=%d\n",
@@ -397,16 +398,47 @@ static void videobuf_vm_close(struct vm_area_struct *vma)
  */
 static int videobuf_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 {
-	struct page *page;
+	struct page *page = NULL;
+	struct videobuf_mapping *map = vma->vm_private_data;
+	struct videobuf_queue *q = map->q;
+	struct videobuf_dma_sg_memory *mem = NULL;
+
+	unsigned long offset;
+	unsigned long page_nr;
+	int first;
 
 	dprintk(3, "fault: fault @ %08lx [vma %08lx-%08lx]\n",
 		(unsigned long)vmf->virtual_address,
 		vma->vm_start, vma->vm_end);
 
-	page = alloc_page(GFP_USER | __GFP_DMA32);
-	if (!page)
-		return VM_FAULT_OOM;
-	clear_user_highpage(page, (unsigned long)vmf->virtual_address);
+	mutex_lock(&q->vb_lock);
+	offset = (unsigned long)vmf->virtual_address - vma->vm_start;
+	page_nr = offset >> PAGE_SHIFT;
+
+	for (first = 0; first < VIDEO_MAX_FRAME; first++) {
+			if (NULL == q->bufs[first])
+				continue;
+
+			MAGIC_CHECK(mem->magic, MAGIC_SG_MEM);
+
+			if (q->bufs[first]->map == map)
+				break;
+	}
+
+	mem = q->bufs[first]->priv;
+	if (!mem)
+		return VM_FAULT_SIGBUS;
+	if (!mem->dma.vmalloc) {
+		mem->dma.vmalloc = vmalloc_32(PAGE_ALIGN(q->bufs[first]->size));
+		if (NULL == mem->dma.vmalloc) {
+			dprintk(1, "%s: vmalloc_32() failed\n", __func__);
+			return VM_FAULT_OOM;
+		}
+	} else
+		page = vmalloc_to_page(mem->dma.vmalloc+
+				(offset & (~PAGE_MASK)));
+	mutex_unlock(&q->vb_lock);
+
 	vmf->page = page;
 
 	return 0;

Re: [PATCH v2]Resend:videobuf_dma_sg: a new implementation for mmap

[Laurent Pinchart]

Hi,

On Friday 30 July 2010 02:08:02 Figo.zhang wrote:
> a mmap issue for videobuf-dma-sg: it will alloc a new page for mmaping when
> it encounter page fault at video_vm_ops->fault(). pls see
> http://www.spinics.net/lists/linux-media/msg21243.html
> 
> a new implementation for mmap, it translate to vmalloc to page at
> video_vm_ops->fault().
> 
> in v2, if mem->dma.vmalloc is NULL at video_vm_ops->fault(), it will alloc
> memory by vmlloc_32().

You're replacing allocation in videobuf_vm_fault by allocationg in 
videobuf_vm_fault. I don't see the point. videobuf_vm_fault needs to go away 
completely.

This has been discussed previously: fixing videobuf is not really possible. A 
videobuf2 implementation is needed and is (slowly) being worked on. I wouldn't 
bother with this patch, just drop it.

> Signed-off-by: Figo.zhang <figo1802@gmail.com>
> ---
> drivers/media/video/videobuf-dma-sg.c |   50
> +++++++++++++++++++++++++++------ 1 files changed, 41 insertions(+), 9
> deletions(-)
> 
> diff --git a/drivers/media/video/videobuf-dma-sg.c
> b/drivers/media/video/videobuf-dma-sg.c index 8359e6b..f7295da 100644
> --- a/drivers/media/video/videobuf-dma-sg.c
> +++ b/drivers/media/video/videobuf-dma-sg.c
> @@ -201,10 +201,11 @@ int videobuf_dma_init_kernel(struct videobuf_dmabuf
> *dma, int direction, dprintk(1, "init kernel [%d pages]\n", nr_pages);
> 
>  	dma->direction = direction;
> -	dma->vmalloc = vmalloc_32(nr_pages << PAGE_SHIFT);
> -	if (NULL == dma->vmalloc) {
> -		dprintk(1, "vmalloc_32(%d pages) failed\n", nr_pages);
> -		return -ENOMEM;
> +	if (!dma->vmalloc)
> +		dma->vmalloc = vmalloc_32(nr_pages << PAGE_SHIFT);
> +		if (NULL == dma->vmalloc) {
> +			dprintk(1, "vmalloc_32(%d pages) failed\n", nr_pages);
> +			return -ENOMEM;
>  	}
> 
>  	dprintk(1, "vmalloc is at addr 0x%08lx, size=%d\n",
> @@ -397,16 +398,47 @@ static void videobuf_vm_close(struct vm_area_struct
> *vma) */
>  static int videobuf_vm_fault(struct vm_area_struct *vma, struct vm_fault
> *vmf) {
> -	struct page *page;
> +	struct page *page = NULL;
> +	struct videobuf_mapping *map = vma->vm_private_data;
> +	struct videobuf_queue *q = map->q;
> +	struct videobuf_dma_sg_memory *mem = NULL;
> +
> +	unsigned long offset;
> +	unsigned long page_nr;
> +	int first;
> 
>  	dprintk(3, "fault: fault @ %08lx [vma %08lx-%08lx]\n",
>  		(unsigned long)vmf->virtual_address,
>  		vma->vm_start, vma->vm_end);
> 
> -	page = alloc_page(GFP_USER | __GFP_DMA32);
> -	if (!page)
> -		return VM_FAULT_OOM;
> -	clear_user_highpage(page, (unsigned long)vmf->virtual_address);
> +	mutex_lock(&q->vb_lock);
> +	offset = (unsigned long)vmf->virtual_address - vma->vm_start;
> +	page_nr = offset >> PAGE_SHIFT;
> +
> +	for (first = 0; first < VIDEO_MAX_FRAME; first++) {
> +			if (NULL == q->bufs[first])
> +				continue;
> +
> +			MAGIC_CHECK(mem->magic, MAGIC_SG_MEM);
> +
> +			if (q->bufs[first]->map == map)
> +				break;
> +	}
> +
> +	mem = q->bufs[first]->priv;
> +	if (!mem)
> +		return VM_FAULT_SIGBUS;
> +	if (!mem->dma.vmalloc) {
> +		mem->dma.vmalloc = vmalloc_32(PAGE_ALIGN(q->bufs[first]->size));
> +		if (NULL == mem->dma.vmalloc) {
> +			dprintk(1, "%s: vmalloc_32() failed\n", __func__);
> +			return VM_FAULT_OOM;
> +		}
> +	} else
> +		page = vmalloc_to_page(mem->dma.vmalloc+
> +				(offset & (~PAGE_MASK)));
> +	mutex_unlock(&q->vb_lock);
> +
>  	vmf->page = page;
> 
>  	return 0;

-- 
Regards,

Laurent Pinchart

Re: [PATCH v2]Resend:videobuf_dma_sg: a new implementation for mmap

[Figo.zhang]

On Fri, 2010-07-30 at 11:31 +0200, Laurent Pinchart wrote:
> Hi,
> 
> On Friday 30 July 2010 02:08:02 Figo.zhang wrote:
> > a mmap issue for videobuf-dma-sg: it will alloc a new page for mmaping when
> > it encounter page fault at video_vm_ops->fault(). pls see
> > http://www.spinics.net/lists/linux-media/msg21243.html
> > 
> > a new implementation for mmap, it translate to vmalloc to page at
> > video_vm_ops->fault().
> > 
> > in v2, if mem->dma.vmalloc is NULL at video_vm_ops->fault(), it will alloc
> > memory by vmlloc_32().
> 
> You're replacing allocation in videobuf_vm_fault by allocationg in 
> videobuf_vm_fault. I don't see the point. videobuf_vm_fault needs to go away 
> completely.
in now videobuf code, the mmap alloc a new buf, the capture dma buffer
using vmalloc() alloc buffer, how is
relationship with them? in usrspace , the mmap region will not the
actual capture dma data, how is work? 

> 
> This has been discussed previously: fixing videobuf is not really possible. A 
> videobuf2 implementation is needed and is (slowly) being worked on. 
hi Mauro, do you think this is a issue for videobuf-dma-sg?

> I wouldn't 
> bother with this patch, just drop it.
> 
> > Signed-off-by: Figo.zhang <figo1802@gmail.com>
> > ---
> > drivers/media/video/videobuf-dma-sg.c |   50
> > +++++++++++++++++++++++++++------ 1 files changed, 41 insertions(+), 9
> > deletions(-)
> > 
> > diff --git a/drivers/media/video/videobuf-dma-sg.c
> > b/drivers/media/video/videobuf-dma-sg.c index 8359e6b..f7295da 100644
> > --- a/drivers/media/video/videobuf-dma-sg.c
> > +++ b/drivers/media/video/videobuf-dma-sg.c
> > @@ -201,10 +201,11 @@ int videobuf_dma_init_kernel(struct videobuf_dmabuf
> > *dma, int direction, dprintk(1, "init kernel [%d pages]\n", nr_pages);
> > 
> >  	dma->direction = direction;
> > -	dma->vmalloc = vmalloc_32(nr_pages << PAGE_SHIFT);
> > -	if (NULL == dma->vmalloc) {
> > -		dprintk(1, "vmalloc_32(%d pages) failed\n", nr_pages);
> > -		return -ENOMEM;
> > +	if (!dma->vmalloc)
> > +		dma->vmalloc = vmalloc_32(nr_pages << PAGE_SHIFT);
> > +		if (NULL == dma->vmalloc) {
> > +			dprintk(1, "vmalloc_32(%d pages) failed\n", nr_pages);
> > +			return -ENOMEM;
> >  	}
> > 
> >  	dprintk(1, "vmalloc is at addr 0x%08lx, size=%d\n",
> > @@ -397,16 +398,47 @@ static void videobuf_vm_close(struct vm_area_struct
> > *vma) */
> >  static int videobuf_vm_fault(struct vm_area_struct *vma, struct vm_fault
> > *vmf) {
> > -	struct page *page;
> > +	struct page *page = NULL;
> > +	struct videobuf_mapping *map = vma->vm_private_data;
> > +	struct videobuf_queue *q = map->q;
> > +	struct videobuf_dma_sg_memory *mem = NULL;
> > +
> > +	unsigned long offset;
> > +	unsigned long page_nr;
> > +	int first;
> > 
> >  	dprintk(3, "fault: fault @ %08lx [vma %08lx-%08lx]\n",
> >  		(unsigned long)vmf->virtual_address,
> >  		vma->vm_start, vma->vm_end);
> > 
> > -	page = alloc_page(GFP_USER | __GFP_DMA32);
> > -	if (!page)
> > -		return VM_FAULT_OOM;
> > -	clear_user_highpage(page, (unsigned long)vmf->virtual_address);
> > +	mutex_lock(&q->vb_lock);
> > +	offset = (unsigned long)vmf->virtual_address - vma->vm_start;
> > +	page_nr = offset >> PAGE_SHIFT;
> > +
> > +	for (first = 0; first < VIDEO_MAX_FRAME; first++) {
> > +			if (NULL == q->bufs[first])
> > +				continue;
> > +
> > +			MAGIC_CHECK(mem->magic, MAGIC_SG_MEM);
> > +
> > +			if (q->bufs[first]->map == map)
> > +				break;
> > +	}
> > +
> > +	mem = q->bufs[first]->priv;
> > +	if (!mem)
> > +		return VM_FAULT_SIGBUS;
> > +	if (!mem->dma.vmalloc) {
> > +		mem->dma.vmalloc = vmalloc_32(PAGE_ALIGN(q->bufs[first]->size));
> > +		if (NULL == mem->dma.vmalloc) {
> > +			dprintk(1, "%s: vmalloc_32() failed\n", __func__);
> > +			return VM_FAULT_OOM;
> > +		}
> > +	} else
> > +		page = vmalloc_to_page(mem->dma.vmalloc+
> > +				(offset & (~PAGE_MASK)));
> > +	mutex_unlock(&q->vb_lock);
> > +
> >  	vmf->page = page;
> > 
> >  	return 0;
> 

Re: [PATCH v2]Resend:videobuf_dma_sg: a new implementation for mmap

[Figo.zhang]

On Fri, 2010-07-30 at 11:31 +0200, Laurent Pinchart wrote:
> Hi,
> 
> On Friday 30 July 2010 02:08:02 Figo.zhang wrote:
> > a mmap issue for videobuf-dma-sg: it will alloc a new page for mmaping when
> > it encounter page fault at video_vm_ops->fault(). pls see
> > http://www.spinics.net/lists/linux-media/msg21243.html
> > 
> > a new implementation for mmap, it translate to vmalloc to page at
> > video_vm_ops->fault().
> > 
> > in v2, if mem->dma.vmalloc is NULL at video_vm_ops->fault(), it will alloc
> > memory by vmlloc_32().
> 
> You're replacing allocation in videobuf_vm_fault by allocationg in 
> videobuf_vm_fault. I don't see the point. videobuf_vm_fault needs to go away 
> completely.
in now videobuf code, the mmap alloc a new buf, the capture dma buffer
using vmalloc() alloc buffer, how is
relationship with them? in usrspace , the mmap region will not the
actual capture dma data, how is work? 

> 
> This has been discussed previously: fixing videobuf is not really possible. A 
> videobuf2 implementation is needed and is (slowly) being worked on. 
hi Mauro, do you think this is a issue for videobuf-dma-sg?

> I wouldn't 
> bother with this patch, just drop it.
> 
> > Signed-off-by: Figo.zhang <figo1802@gmail.com>
> > ---
> > drivers/media/video/videobuf-dma-sg.c |   50
> > +++++++++++++++++++++++++++------ 1 files changed, 41 insertions(+), 9
> > deletions(-)
> > 
> > diff --git a/drivers/media/video/videobuf-dma-sg.c
> > b/drivers/media/video/videobuf-dma-sg.c index 8359e6b..f7295da 100644
> > --- a/drivers/media/video/videobuf-dma-sg.c
> > +++ b/drivers/media/video/videobuf-dma-sg.c
> > @@ -201,10 +201,11 @@ int videobuf_dma_init_kernel(struct videobuf_dmabuf
> > *dma, int direction, dprintk(1, "init kernel [%d pages]\n", nr_pages);
> > 
> >  	dma->direction = direction;
> > -	dma->vmalloc = vmalloc_32(nr_pages << PAGE_SHIFT);
> > -	if (NULL == dma->vmalloc) {
> > -		dprintk(1, "vmalloc_32(%d pages) failed\n", nr_pages);
> > -		return -ENOMEM;
> > +	if (!dma->vmalloc)
> > +		dma->vmalloc = vmalloc_32(nr_pages << PAGE_SHIFT);
> > +		if (NULL == dma->vmalloc) {
> > +			dprintk(1, "vmalloc_32(%d pages) failed\n", nr_pages);
> > +			return -ENOMEM;
> >  	}
> > 
> >  	dprintk(1, "vmalloc is at addr 0x%08lx, size=%d\n",
> > @@ -397,16 +398,47 @@ static void videobuf_vm_close(struct vm_area_struct
> > *vma) */
> >  static int videobuf_vm_fault(struct vm_area_struct *vma, struct vm_fault
> > *vmf) {
> > -	struct page *page;
> > +	struct page *page = NULL;
> > +	struct videobuf_mapping *map = vma->vm_private_data;
> > +	struct videobuf_queue *q = map->q;
> > +	struct videobuf_dma_sg_memory *mem = NULL;
> > +
> > +	unsigned long offset;
> > +	unsigned long page_nr;
> > +	int first;
> > 
> >  	dprintk(3, "fault: fault @ %08lx [vma %08lx-%08lx]\n",
> >  		(unsigned long)vmf->virtual_address,
> >  		vma->vm_start, vma->vm_end);
> > 
> > -	page = alloc_page(GFP_USER | __GFP_DMA32);
> > -	if (!page)
> > -		return VM_FAULT_OOM;
> > -	clear_user_highpage(page, (unsigned long)vmf->virtual_address);
> > +	mutex_lock(&q->vb_lock);
> > +	offset = (unsigned long)vmf->virtual_address - vma->vm_start;
> > +	page_nr = offset >> PAGE_SHIFT;
> > +
> > +	for (first = 0; first < VIDEO_MAX_FRAME; first++) {
> > +			if (NULL == q->bufs[first])
> > +				continue;
> > +
> > +			MAGIC_CHECK(mem->magic, MAGIC_SG_MEM);
> > +
> > +			if (q->bufs[first]->map == map)
> > +				break;
> > +	}
> > +
> > +	mem = q->bufs[first]->priv;
> > +	if (!mem)
> > +		return VM_FAULT_SIGBUS;
> > +	if (!mem->dma.vmalloc) {
> > +		mem->dma.vmalloc = vmalloc_32(PAGE_ALIGN(q->bufs[first]->size));
> > +		if (NULL == mem->dma.vmalloc) {
> > +			dprintk(1, "%s: vmalloc_32() failed\n", __func__);
> > +			return VM_FAULT_OOM;
> > +		}
> > +	} else
> > +		page = vmalloc_to_page(mem->dma.vmalloc+
> > +				(offset & (~PAGE_MASK)));
> > +	mutex_unlock(&q->vb_lock);
> > +
> >  	vmf->page = page;
> > 
> >  	return 0;
> 

Re: [PATCH v2]Resend:videobuf_dma_sg: a new implementation for mmap

[Figo.zhang]

On Fri, 2010-07-30 at 19:39 +0800, Figo.zhang wrote:
> On Fri, 2010-07-30 at 11:31 +0200, Laurent Pinchart wrote:
> > Hi,
> > 
> > On Friday 30 July 2010 02:08:02 Figo.zhang wrote:
> > > a mmap issue for videobuf-dma-sg: it will alloc a new page for mmaping when
> > > it encounter page fault at video_vm_ops->fault(). pls see
> > > http://www.spinics.net/lists/linux-media/msg21243.html
> > > 
> > > a new implementation for mmap, it translate to vmalloc to page at
> > > video_vm_ops->fault().
> > > 
> > > in v2, if mem->dma.vmalloc is NULL at video_vm_ops->fault(), it will alloc
> > > memory by vmlloc_32().
> > 
> > You're replacing allocation in videobuf_vm_fault by allocationg in 
> > videobuf_vm_fault. I don't see the point. videobuf_vm_fault needs to go away 
> > completely.
> in now videobuf code, the mmap alloc a new buf, the capture dma buffer
> using vmalloc() alloc buffer, how is
> relationship with them? in usrspace , the mmap region will not the
> actual capture dma data, how is work? 


hmm, I understand with some mistake before.for mmap in videobuf-sg, it
is not call vmalloc_32() to alloc memory, it call
videobuf_dma_init_user_locked()->get_user_pages(), and the end, it will
handle_mm_fault()-> vm_ops->fault(). Because in mmap(), it have assigned
vaule for "baddr" variable.

it is too obscure for videobuf-sg, hope for videbuf2. ~~