Re: [Xen-devel] [PATCH v2 5/9] xen/gntdev: Allow mappings for DMA buffers

From: Oleksandr Andrushchenko
Date: Fri Jun 08 2018 - 07:33:56 EST

On 06/08/2018 12:46 AM, Boris Ostrovsky wrote:
(Stefano, question for you at the end)

On 06/07/2018 02:39 AM, Oleksandr Andrushchenko wrote:
On 06/07/2018 12:19 AM, Boris Ostrovsky wrote:
On 06/06/2018 04:14 AM, Oleksandr Andrushchenko wrote:
On 06/04/2018 11:12 PM, Boris Ostrovsky wrote:
On 06/01/2018 07:41 AM, Oleksandr Andrushchenko wrote:
@@ -121,8 +146,27 @@ static void gntdev_free_map(struct grant_map
ÂÂÂÂÂÂ if (map == NULL)
*Option 1: kfree(map->frames);*
+ÂÂÂ if (map->dma_vaddr) {
+ÂÂÂÂÂÂÂ struct gnttab_dma_alloc_args args;
+ÂÂÂÂÂÂÂ = map->dma_dev;
+ÂÂÂÂÂÂÂ args.coherent = map->dma_flags & GNTDEV_DMA_FLAG_COHERENT;
+ÂÂÂÂÂÂÂ args.nr_pages = map->count;
+ÂÂÂÂÂÂÂ args.pages = map->pages;
+ÂÂÂÂÂÂÂ args.frames = map->frames;
+ÂÂÂÂÂÂÂ args.vaddr = map->dma_vaddr;
+ÂÂÂÂÂÂÂ args.dev_bus_addr = map->dma_bus_addr;
+ÂÂÂÂÂÂÂ gnttab_dma_free_pages(&args);
*Option 2: kfree(map->frames);*
+ÂÂÂ } else
ÂÂÂÂÂÂ if (map->pages)
ÂÂÂÂÂÂÂÂÂÂ gnttab_free_pages(map->count, map->pages);
+ÂÂÂ kfree(map->frames);

Can this be done under if (map->dma_vaddr) ?
ÂÂ In other words, is it
possible for dma_vaddr to be NULL and still have unallocated frames
It is possible to have vaddr == NULL and frames != NULL as we
allocate frames outside of gnttab_dma_alloc_pages which
may fail. Calling kfree on NULL pointer is safe,
I am not questioning safety of the code, I would like avoid another
Ah, I now understand, so you are asking if we can have
that kfree(map->frames); in the place *Option 2* I marked above.
Unfortunately no: map->frames is allocated before we try to
allocate DMA memory, e.g. before dma_vaddr is set:
ÂÂÂ ÂÂÂ add->frames = kcalloc(count, sizeof(add->frames[0]),
ÂÂÂ ÂÂÂ if (!add->frames)
ÂÂÂ ÂÂÂ ÂÂÂ goto err;

ÂÂÂ ÂÂÂ if (gnttab_dma_alloc_pages(&args))
ÂÂÂ ÂÂÂ ÂÂÂ goto err;

ÂÂÂ ÂÂÂ add->dma_vaddr = args.vaddr;
ÂÂÂ gntdev_free_map(add);

So, it is possible to enter gntdev_free_map with
frames != NULL and dma_vaddr == NULL. Option 1 above cannot be used
as map->frames is needed for gnttab_dma_free_pages(&args);
and Option 2 cannot be used as frames != NULL and dma_vaddr == NULL.
Thus, I think that unfortunately we need that #ifdef.
Option 3 below can also be considered, but that seems to be not good
as we free resources in different places which looks inconsistent.

I was only thinking of option 2. But if it is possible to have frames !=
NULL and dma_vaddr == NULL then perhaps we indeed will have to live with
the extra ifdef.

Sorry if I'm still missing your point.
I see no reason to change this code.
ÂÂÂÂÂÂ kfree(map->pages);
ÂÂÂÂÂÂ kfree(map->grants);
ÂÂÂÂÂÂ kfree(map->map_ops);
@@ -132,7 +176,8 @@ static void gntdev_free_map(struct grant_map
ÂÂÂÂÂÂ kfree(map);
ÂÂ }
ÂÂ -static struct grant_map *gntdev_alloc_map(struct gntdev_priv
*priv, int count)
+static struct grant_map *gntdev_alloc_map(struct gntdev_priv *priv,
int count,
ÂÂ {
ÂÂÂÂÂÂ struct grant_map *add;
ÂÂÂÂÂÂ int i;
@@ -155,6 +200,37 @@ static struct grant_map
*gntdev_alloc_map(struct gntdev_priv *priv, int count)
ÂÂÂÂÂÂÂÂÂÂ NULL == add->pages)
ÂÂÂÂÂÂÂÂÂÂ goto err;
+ÂÂÂ add->dma_flags = dma_flags;
+ÂÂÂ /*
+ÂÂÂÂ * Check if this mapping is requested to be backed
+ÂÂÂÂ * by a DMA buffer.
+ÂÂÂÂ */
+ÂÂÂ if (dma_flags & (GNTDEV_DMA_FLAG_WC |
+ÂÂÂÂÂÂÂ struct gnttab_dma_alloc_args args;
+ÂÂÂÂÂÂÂ add->frames = kcalloc(count, sizeof(add->frames[0]),
+ÂÂÂÂÂÂÂ if (!add->frames)
+ÂÂÂÂÂÂÂÂÂÂÂ goto err;
+ÂÂÂÂÂÂÂ /* Remember the device, so we can free DMA memory. */
+ÂÂÂÂÂÂÂ add->dma_dev = priv->dma_dev;
+ÂÂÂÂÂÂÂ = priv->dma_dev;
+ÂÂÂÂÂÂÂ args.coherent = dma_flags & GNTDEV_DMA_FLAG_COHERENT;
+ÂÂÂÂÂÂÂ args.nr_pages = count;
+ÂÂÂÂÂÂÂ args.pages = add->pages;
+ÂÂÂÂÂÂÂ args.frames = add->frames;
+ÂÂÂÂÂÂÂ if (gnttab_dma_alloc_pages(&args))
*Option 3: kfree(map->frames);*
+ÂÂÂÂÂÂÂÂÂÂÂ goto err;
+ÂÂÂÂÂÂÂ add->dma_vaddr = args.vaddr;
+ÂÂÂÂÂÂÂ add->dma_bus_addr = args.dev_bus_addr;
+ÂÂÂ } else
ÂÂÂÂÂÂ if (gnttab_alloc_pages(count, add->pages))
ÂÂÂÂÂÂÂÂÂÂ goto err;
ÂÂ @@ -325,6 +401,14 @@ static int map_grant_pages(struct grant_map
ÂÂÂÂÂÂÂÂÂÂ map->unmap_ops[i].handle = map->map_ops[i].handle;
ÂÂÂÂÂÂÂÂÂÂ if (use_ptemod)
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
+ÂÂÂÂÂÂÂ else if (map->dma_vaddr) {
+ÂÂÂÂÂÂÂÂÂÂÂ unsigned long mfn;
+ÂÂÂÂÂÂÂÂÂÂÂ mfn = __pfn_to_mfn(page_to_pfn(map->pages[i]));
Not pfn_to_mfn()?
I'd love to, but pfn_to_mfn is only defined for x86, not ARM: [1]
and [2]

drivers/xen/gntdev.c:408:10: error: implicit declaration of function
âpfn_to_mfnâ [-Werror=implicit-function-declaration]
ÂÂÂÂ mfn = pfn_to_mfn(page_to_pfn(map->pages[i]));

So, I'll keep __pfn_to_mfn
How will this work on non-PV x86?
So, you mean I need:
#ifdef CONFIG_X86
mfn = pfn_to_mfn(page_to_pfn(map->pages[i]));
mfn = __pfn_to_mfn(page_to_pfn(map->pages[i]));

I'd rather fix it in ARM code. Stefano, why does ARM uses the
underscored version?
Do you want me to add one more patch for ARM to wrap __pfn_to_mfn
with static inline for ARM? e.g.
static inline ...pfn_to_mfn(...)
ÂÂÂ __pfn_to_mfn();


Thank you,