Re: [PATCH] pci-dma: Fix x86 dma_alloc_coherent to fully clear all pages returned

[Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>]

On 01/30/2015 10:54 PM, Tim Chen wrote:

> Commit d92ef66c4f8f ("x86: make dma_alloc_coherent() return zeroed memory
> if CMA is enabled") changed the dma_alloc_coherent page clearance from
> using an __GFP_ZERO in page allocation to not setting the flag but doing
> an explicit memory clear at the end.

> However the memory clear only covered the memory size that
> was requested, but may not be up to the full extent of the
> last page, if the total pages returned exceed the
> memory size requested.  This behavior has caused problem with XHCI
> and caused it to hang and froze usb:

> kernel: xhci_hcd 0000:00:14.0: Stopped the command ring failed, maybe the host is dead
> kernel: xhci_hcd 0000:00:14.0: Abort command ring failed
> kernel: xhci_hcd 0000:00:14.0: HC died; cleaning up
> kernel: xhci_hcd 0000:00:14.0: Error while assigning device slot ID
> kernel: xhci_hcd 0000:00:14.0: Max number of devices this xHCI host supports is 64.

> Other drivers may have similar issue if it assumes that the pages
> allocated are completely zeroed.

> This patch ensures that the pages returned are fully cleared.

> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
> ---
>   arch/x86/kernel/pci-dma.c | 2 ++
>   1 file changed, 2 insertions(+)

> diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
> index a25e202..534470f 100644
> --- a/arch/x86/kernel/pci-dma.c
> +++ b/arch/x86/kernel/pci-dma.c
> @@ -125,6 +125,8 @@ again:
>
>   		return NULL;
>   	}
> +	/* round up to full page size */
> +	size = (((size-1) >> PAGE_SHIFT) + 1) * PAGE_SIZE;

    This is quite suboptimal formula, we can do without shifts and 
multiplications (hopefully, still converted to shifts by gcc):

	size = (size + ~PAGE_MASK) & PAGE_MASK;

WBR, Sergei