Re: [RESENDING RFC PATCH 4/4] usb: xhci: Use temporary buffer to consolidate SG

[Greg KH <gregkh@linuxfoundation.org>]

On Fri, Mar 27, 2020 at 10:05:21AM +0000, Tejas Joglekar wrote:
> Hi,
> On 3/27/2020 2:57 PM, Greg KH wrote:
> > On Fri, Mar 27, 2020 at 02:23:46PM +0530, Tejas Joglekar wrote:
> >> The Synopsys xHC has an internal TRB cache of size TRB_CACHE_SIZE for
> >> each endpoint. The default value for TRB_CACHE_SIZE is 16 for SS and 8
> >> for HS. The controller loads and updates the TRB cache from the transfer
> >> ring in system memory whenever the driver issues a start transfer or
> >> update transfer command.
> >>
> >> For chained TRBs, the Synopsys xHC requires that the total amount of
> >> bytes for all TRBs loaded in the TRB cache be greater than or equal to 1
> >> MPS. Or the chain ends within the TRB cache (with a last TRB).
> >>
> >> If this requirement is not met, the controller will not be able to send
> >> or receive a packet and it will hang causing a driver timeout and error.
> > 
> > Sounds like broken hardware, or is this requirement in the xhci spec?
> > 
> Not a xhci spec requirement, but behavior of Synopsys xHC. We have not seen
> actual failure on Linux yet but it is possible in future if SG list with
> very small buffer size is given for transfer. 

So this is something required that is outside of the spec, meaning that
the hardware is imposing additional requirements, which implies it's a
hardware bug, or "quirk", right?

> >> This can be a problem if a class driver queues SG requests with many
> >> small-buffer entries. The XHCI driver will create a chained TRB for each
> >> entry which may trigger this issue.
> >>
> >> This patch adds logic to the XHCI driver to detect and prevent this from
> >> happening.
> >>
> >> For every (TRB_CACHE_SIZE - 2), we check the total buffer size of
> >> the SG list and if the last window of (TRB_CACHE_SIZE - 2) SG list length
> >> and we don't make up at least 1 MPS, we create a temporary buffer to
> >> consolidate full SG list into the buffer.
> >>
> >> We check at (TRB_CACHE_SIZE - 2) window because it is possible that there
> >> would be a link and/or event data TRB that take up to 2 of the cache
> >> entries.
> >>
> >> We discovered this issue with devices on other platforms but have not
> >> yet come across any device that triggers this on Linux. But it could be
> >> a real problem now or in the future. All it takes is N number of small
> >> chained TRBs. And other instances of the Synopsys IP may have smaller
> >> values for the TRB_CACHE_SIZE which would exacerbate the problem.
> >>
> >> Signed-off-by: Tejas Joglekar <joglekar@synopsys.com>
> >> ---
> >>
> >> Resending as 'umlaut' in email are not accepted by some servers.
> >>
> >>  drivers/usb/core/hcd.c       |   8 +++
> >>  drivers/usb/host/xhci-ring.c |   2 +-
> >>  drivers/usb/host/xhci.c      | 128 +++++++++++++++++++++++++++++++++++++++++++
> >>  drivers/usb/host/xhci.h      |   4 ++
> >>  4 files changed, 141 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/drivers/usb/core/hcd.c b/drivers/usb/core/hcd.c
> >> index aa45840d8273..fdd257a2b8a6 100644
> >> --- a/drivers/usb/core/hcd.c
> >> +++ b/drivers/usb/core/hcd.c
> >> @@ -1459,6 +1459,14 @@ int usb_hcd_map_urb_for_dma(struct usb_hcd *hcd, struct urb *urb,
> >>  					return -EINVAL;
> >>  				}
> >>  
> >> +				/*
> >> +				 * If SG is consolidate into single buffer
> >> +				 * return early
> > 
> > I do not understand this comment.
> 
> The SG list is copied to a temporary buffer, and buffer is DMA mapped so we should 
> not map the SG list again, and return without any mapping here. 

Please write this all out a lot more to make it more obvious.


> > 
> >> +				 */
> >> +				if ((urb->transfer_flags &
> >> +				     URB_DMA_MAP_SINGLE))
> >> +					return ret;
> > 
> > Why?  Isn't this now going to affect other host controllers (like all of
> > them?)
> > 
> When urb->num_sgs is greater than 0, other than my quirk dma_map function no one
> will set the DMA transfer flag as URB_DMA_MAP_SINGLE. So it would not be called by
> all HC's. Even when the SG list does not have very small buffer sizes this quirk will
> not set the URB_DMA_MAP_SINGLE transfer flag.

Are you sure?  :)

> >> +
> >>  				n = dma_map_sg(
> >>  						hcd->self.sysdev,
> >>  						urb->sg,
> >> diff --git a/drivers/usb/host/xhci-ring.c b/drivers/usb/host/xhci-ring.c
> >> index a78787bb5133..2fad9474912a 100644
> >> --- a/drivers/usb/host/xhci-ring.c
> >> +++ b/drivers/usb/host/xhci-ring.c
> >> @@ -3291,7 +3291,7 @@ int xhci_queue_bulk_tx(struct xhci_hcd *xhci, gfp_t mem_flags,
> >>  
> >>  	full_len = urb->transfer_buffer_length;
> >>  	/* If we have scatter/gather list, we use it. */
> >> -	if (urb->num_sgs) {
> >> +	if (urb->num_sgs && !(urb->transfer_flags & URB_DMA_MAP_SINGLE)) {
> >>  		num_sgs = urb->num_mapped_sgs;
> >>  		sg = urb->sg;
> >>  		addr = (u64) sg_dma_address(sg);
> >> diff --git a/drivers/usb/host/xhci.c b/drivers/usb/host/xhci.c
> >> index fe38275363e0..94fddbd06179 100644
> >> --- a/drivers/usb/host/xhci.c
> >> +++ b/drivers/usb/host/xhci.c
> >> @@ -1256,6 +1256,109 @@ EXPORT_SYMBOL_GPL(xhci_resume);
> >>  
> >>  /*-------------------------------------------------------------------------*/
> >>  
> >> +static int xhci_map_temp_buffer(struct usb_hcd *hcd, struct urb *urb)
> >> +{
> >> +	void *temp;
> >> +	int ret = 0;
> >> +	unsigned int len;
> >> +	unsigned int buf_len;
> >> +	enum dma_data_direction dir;
> >> +	struct xhci_hcd *xhci;
> >> +
> >> +	xhci = hcd_to_xhci(hcd);
> >> +	dir = usb_urb_dir_in(urb) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
> >> +	buf_len = urb->transfer_buffer_length;
> >> +
> >> +	temp = kzalloc_node(buf_len, GFP_ATOMIC,
> >> +			    dev_to_node(hcd->self.sysdev));
> >> +	if (!temp) {
> >> +		xhci_warn(xhci, "Failed to create temp buffer, HC may fail\n");
> > 
> > Didn't kzalloc just warn before this?
> > 
> Yes, It should.

Then do not spit out another message please.

> > And isn't this whole thing going to cause a lot more memory allocations
> > per submission than before?
> > 
> If buffer sizes for SG list are very small (less than MPS size per TRB_CACHE_SIZE)
> yes, it will have more memory allocations. 

That's not good :(

greg k-h