On Wed, 11 Apr 2018 11:12:24 +0200
Ladislav Michl <la...@linux-mips.org> wrote:

> On Wed, Apr 11, 2018 at 10:52:01AM +0200, Boris Brezillon wrote:
> > On Wed, 11 Apr 2018 10:27:46 +0200
> > Ladislav Michl <la...@linux-mips.org> wrote:
> >   
> > > On Wed, Apr 11, 2018 at 10:08:06AM +0200, Boris Brezillon wrote:  
> > > > On Wed, 11 Apr 2018 09:36:56 +0200
> > > > Ladislav Michl <la...@linux-mips.org> wrote:
> > > >     
> > > > > Hi Boris,
> > > > > 
> > > > > On Wed, Apr 11, 2018 at 09:15:28AM +0200, Boris Brezillon wrote:    
> > > [...]  
> > > > > > Not sure this approach is safe on all archs: if the cache is VIVT or
> > > > > > VIPT, you may have several entries pointing to the same phys page, 
> > > > > > and
> > > > > > then, when dma_map_page() does its cache maintenance operations, 
> > > > > > it's
> > > > > > only taking one of these entries into account.      
> > > > > 
> > > > > Hmm, I used the same approach Samsung OneNAND driver does since commit
> > > > > dcf08227e964a53a2cb39130b74842c7dcb6adde.
> > > > > Both TI OMAP3630 and Samsung S5PC110 are using Cortex-A8 which
> > > > > is VIPT. In that case samsung's driver code has the same problem.
> > > > >     
> > > > > > In other parts of the MTD subsystem, we tend to not do DMA on 
> > > > > > buffers
> > > > > > that have been vmalloc-ed.
> > > > > > 
> > > > > > You can do something like
> > > > > > 
> > > > > >             if (virt_addr_valid(buf))
> > > > > >                     /* Use DMA */
> > > > > >             else
> > > > > >                     /*
> > > > > >                      * Do not use DMA, or use a bounce buffer
> > > > > >                      * allocated with kmalloc
> > > > > >                      */      
> > > > > 
> > > > > Okay, I'll use this approach then, but first I'd like to be sure 
> > > > > above is
> > > > > correct. Anyone?    
> > > > 
> > > > See this discussion [1]. The problem came up a few times already, so
> > > > might find other threads describing why it's not safe.
> > > > 
> > > > [1]https://lists.linuxfoundation.org/pipermail/iommu/2016-March/016240.html
> > > >     
> > > 
> > > Question was more likely whenever there might exist more that one mapping
> > > of the same page.  
> > 
> > I'm clearly not an expert, so I might be wrong, but I think a physical
> > page can be both in the identity mapping and mapped in the vmalloc
> > space. Now, is there a real risk that both ends are accessed in
> > parallel thus making different cache entries pointing to the same phys
> > page dirty, I'm not sure. Or maybe there are other corner case, but
> > you'll have to ask Russell (or any other ARM expert) to get a clearer
> > answer.  
> 
> Thank you anyway. As DMA was disabled completely for all DT enabled boards
> until 4.16 let's play safe and disable it for HIGHMEM case as you suggested.
> Later, we might eventually use the same {read,write}_bufferram for both
> OMAP and S5PC110.
> 
> > > But okay, I'll disable DMA for highmem. Patch will follow.
> > > 
> > >   ladis  
> 
> Andreas, Nikolaus, could you please test patch bellow? It is against
> linux.git and should apply also against 4.16 once you modify path.
> 

Looks like it fixes the problem. Could you please rebase on top of
v4.17-rc1 when it's out and send a new version? Note that onenand code
has moved to drivers/mtd/nand/onenand, so we'll have to manually
backport the fix to 4.16.

Regards,

Boris

> Thank you,
>       ladis
> 
> --- >8 ---  
> 
> From: Ladislav Michl <la...@linux-mips.org>
> Subject: [PATCH] mtd: onenand: omap2: Disable DMA for HIGHMEM buffers
> 
> dma_map_single doesn't get the proper DMA address for vmalloced area,
> so disable DMA in this case.
> 
> Signed-off-by: Ladislav Michl <la...@linux-mips.org>
> Reported-by: "H. Nikolaus Schaller" <h...@goldelico.com>
> ---
>  drivers/mtd/nand/onenand/omap2.c | 105 +++++++++++--------------------
>  1 file changed, 38 insertions(+), 67 deletions(-)
> 
> diff --git a/drivers/mtd/nand/onenand/omap2.c 
> b/drivers/mtd/nand/onenand/omap2.c
> index 9c159f0dd9a6..321137158ff3 100644
> --- a/drivers/mtd/nand/onenand/omap2.c
> +++ b/drivers/mtd/nand/onenand/omap2.c
> @@ -375,56 +375,42 @@ static int omap2_onenand_read_bufferram(struct mtd_info 
> *mtd, int area,
>  {
>       struct omap2_onenand *c = container_of(mtd, struct omap2_onenand, mtd);
>       struct onenand_chip *this = mtd->priv;
> -     dma_addr_t dma_src, dma_dst;
> -     int bram_offset;
> +     struct device *dev = &c->pdev->dev;
>       void *buf = (void *)buffer;
> +     dma_addr_t dma_src, dma_dst;
> +     int bram_offset, err;
>       size_t xtra;
> -     int ret;
>  
>       bram_offset = omap2_onenand_bufferram_offset(mtd, area) + area + offset;
> -     if (bram_offset & 3 || (size_t)buf & 3 || count < 384)
> -             goto out_copy;
> -
> -     /* panic_write() may be in an interrupt context */
> -     if (in_interrupt() || oops_in_progress)
> +     /*
> +      * If the buffer address is not DMA-able, len is not long enough to make
> +      * DMA transfers profitable or panic_write() may be in an interrupt
> +      * context fallback to PIO mode.
> +      */
> +     if (!virt_addr_valid(buf) || bram_offset & 3 || (size_t)buf & 3 ||
> +         count < 384 || in_interrupt() || oops_in_progress )
>               goto out_copy;
>  
> -     if (buf >= high_memory) {
> -             struct page *p1;
> -
> -             if (((size_t)buf & PAGE_MASK) !=
> -                 ((size_t)(buf + count - 1) & PAGE_MASK))
> -                     goto out_copy;
> -             p1 = vmalloc_to_page(buf);
> -             if (!p1)
> -                     goto out_copy;
> -             buf = page_address(p1) + ((size_t)buf & ~PAGE_MASK);
> -     }
> -
>       xtra = count & 3;
>       if (xtra) {
>               count -= xtra;
>               memcpy(buf + count, this->base + bram_offset + count, xtra);
>       }
>  
> +     dma_dst = dma_map_single(dev, buf, count, DMA_FROM_DEVICE);
>       dma_src = c->phys_base + bram_offset;
> -     dma_dst = dma_map_single(&c->pdev->dev, buf, count, DMA_FROM_DEVICE);
> -     if (dma_mapping_error(&c->pdev->dev, dma_dst)) {
> -             dev_err(&c->pdev->dev,
> -                     "Couldn't DMA map a %d byte buffer\n",
> -                     count);
> -             goto out_copy;
> -     }
>  
> -     ret = omap2_onenand_dma_transfer(c, dma_src, dma_dst, count);
> -     dma_unmap_single(&c->pdev->dev, dma_dst, count, DMA_FROM_DEVICE);
> -
> -     if (ret) {
> -             dev_err(&c->pdev->dev, "timeout waiting for DMA\n");
> +     if (dma_mapping_error(dev, dma_dst)) {
> +             dev_err(dev, "Couldn't DMA map a %d byte buffer\n", count);
>               goto out_copy;
>       }
>  
> -     return 0;
> +     err = omap2_onenand_dma_transfer(c, dma_src, dma_dst, count);
> +     dma_unmap_single(dev, dma_dst, count, DMA_FROM_DEVICE);
> +     if (!err)
> +             return 0;
> +
> +     dev_err(dev, "timeout waiting for DMA\n");
>  
>  out_copy:
>       memcpy(buf, this->base + bram_offset, count);
> @@ -437,49 +423,34 @@ static int omap2_onenand_write_bufferram(struct 
> mtd_info *mtd, int area,
>  {
>       struct omap2_onenand *c = container_of(mtd, struct omap2_onenand, mtd);
>       struct onenand_chip *this = mtd->priv;
> -     dma_addr_t dma_src, dma_dst;
> -     int bram_offset;
> +     struct device *dev = &c->pdev->dev;
>       void *buf = (void *)buffer;
> -     int ret;
> +     dma_addr_t dma_src, dma_dst;
> +     int bram_offset, err;
>  
>       bram_offset = omap2_onenand_bufferram_offset(mtd, area) + area + offset;
> -     if (bram_offset & 3 || (size_t)buf & 3 || count < 384)
> -             goto out_copy;
> -
> -     /* panic_write() may be in an interrupt context */
> -     if (in_interrupt() || oops_in_progress)
> +     /*
> +      * If the buffer address is not DMA-able, len is not long enough to make
> +      * DMA transfers profitable or panic_write() may be in an interrupt
> +      * context fallback to PIO mode.
> +      */
> +     if (!virt_addr_valid(buf) || bram_offset & 3 || (size_t)buf & 3 ||
> +         count < 384 || in_interrupt() || oops_in_progress )
>               goto out_copy;
>  
> -     if (buf >= high_memory) {
> -             struct page *p1;
> -
> -             if (((size_t)buf & PAGE_MASK) !=
> -                 ((size_t)(buf + count - 1) & PAGE_MASK))
> -                     goto out_copy;
> -             p1 = vmalloc_to_page(buf);
> -             if (!p1)
> -                     goto out_copy;
> -             buf = page_address(p1) + ((size_t)buf & ~PAGE_MASK);
> -     }
> -
> -     dma_src = dma_map_single(&c->pdev->dev, buf, count, DMA_TO_DEVICE);
> +     dma_src = dma_map_single(dev, buf, count, DMA_TO_DEVICE);
>       dma_dst = c->phys_base + bram_offset;
> -     if (dma_mapping_error(&c->pdev->dev, dma_src)) {
> -             dev_err(&c->pdev->dev,
> -                     "Couldn't DMA map a %d byte buffer\n",
> -                     count);
> -             return -1;
> -     }
> -
> -     ret = omap2_onenand_dma_transfer(c, dma_src, dma_dst, count);
> -     dma_unmap_single(&c->pdev->dev, dma_src, count, DMA_TO_DEVICE);
> -
> -     if (ret) {
> -             dev_err(&c->pdev->dev, "timeout waiting for DMA\n");
> +     if (dma_mapping_error(dev, dma_src)) {
> +             dev_err(dev, "Couldn't DMA map a %d byte buffer\n", count);
>               goto out_copy;
>       }
>  
> -     return 0;
> +     err = omap2_onenand_dma_transfer(c, dma_src, dma_dst, count);
> +     dma_unmap_page(dev, dma_src, count, DMA_TO_DEVICE);
> +     if (!err)
> +             return 0;
> +
> +     dev_err(dev, "timeout waiting for DMA\n");
>  
>  out_copy:
>       memcpy(this->base + bram_offset, buf, count);

Reply via email to