Re: [PATCH 1/2] uas: revert from scsi_add_host_with_dma() to scsi_add_host()

From: Hans de Goede <hdegoede@redhat.com>
To: Tom Yan <tom.ty89@gmail.com>,
	hch@lst.de, gregkh@linuxfoundation.org,
	linux-usb@vger.kernel.org
Cc: mathias.nyman@intel.com, linux-kernel@vger.kernel.org,
	linux-pci@vger.kernel.org, baolu.lu@linux.intel.com
Subject: Re: [PATCH 1/2] uas: revert from scsi_add_host_with_dma() to scsi_add_host()
Date: Mon, 30 Nov 2020 10:48:41 +0100	[thread overview]
Message-ID: <62e0d5ea-e665-b913-5482-a75db0ac1368@redhat.com> (raw)
In-Reply-To: <20201128154849.3193-1-tom.ty89@gmail.com>

Hi,

On 11/28/20 4:48 PM, Tom Yan wrote:
> Apparently the former (with the chosen dma_dev) may cause problem in certain
> case (e.g. where thunderbolt dock and intel iommu are involved). The error
> observed was:
> 
> XHCI swiotlb buffer is full / DMAR: Device bounce map failed
> 
> For now we retain the clamp for hw_max_sectors against the dma_max_mapping_size.
> Since the device/size for the clamp that is applied when the scsi request queue
> is initialized/allocated is different than the one used here, we invalidate the
> early clamping by making a fallback blk_queue_max_hw_sectors() call.
> 
> Signed-off-by: Tom Yan <tom.ty89@gmail.com>

I can confirm that this fixes the network performance on a Lenovo Thunderbolt
dock generation 2, which uses an USB attach NIC.

With this patch added on top of 5.10-rc5 scp performance to another machine
on the local gbit LAN goes back from the regressed 1 MB/s to its original 100MB/s
as it should be:

Tested-by: Hans de Goede <hdegoede@redhat.com>

Regards,

Hans

> ---
>  drivers/usb/storage/uas.c | 11 +++++++----
>  1 file changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/usb/storage/uas.c b/drivers/usb/storage/uas.c
> index c8a577309e8f..5db1325cea20 100644
> --- a/drivers/usb/storage/uas.c
> +++ b/drivers/usb/storage/uas.c
> @@ -843,18 +843,21 @@ static int uas_slave_alloc(struct scsi_device *sdev)
>  static int uas_slave_configure(struct scsi_device *sdev)
>  {
>  	struct uas_dev_info *devinfo = sdev->hostdata;
> -	struct device *dev = sdev->host->dma_dev;
> +	struct usb_device *udev = devinfo->udev;
>  
>  	if (devinfo->flags & US_FL_MAX_SECTORS_64)
>  		blk_queue_max_hw_sectors(sdev->request_queue, 64);
>  	else if (devinfo->flags & US_FL_MAX_SECTORS_240)
>  		blk_queue_max_hw_sectors(sdev->request_queue, 240);
> -	else if (devinfo->udev->speed >= USB_SPEED_SUPER)
> +	else if (udev->speed >= USB_SPEED_SUPER)
>  		blk_queue_max_hw_sectors(sdev->request_queue, 2048);
> +	else
> +		blk_queue_max_hw_sectors(sdev->request_queue,
> +					 SCSI_DEFAULT_MAX_SECTORS);
>  
>  	blk_queue_max_hw_sectors(sdev->request_queue,
>  		min_t(size_t, queue_max_hw_sectors(sdev->request_queue),
> -		      dma_max_mapping_size(dev) >> SECTOR_SHIFT));
> +		      dma_max_mapping_size(udev->bus->sysdev) >> SECTOR_SHIFT));
>  
>  	if (devinfo->flags & US_FL_NO_REPORT_OPCODES)
>  		sdev->no_report_opcodes = 1;
> @@ -1040,7 +1043,7 @@ static int uas_probe(struct usb_interface *intf, const struct usb_device_id *id)
>  	shost->can_queue = devinfo->qdepth - 2;
>  
>  	usb_set_intfdata(intf, shost);
> -	result = scsi_add_host_with_dma(shost, &intf->dev, udev->bus->sysdev);
> +	result = scsi_add_host(shost, &intf->dev);
>  	if (result)
>  		goto free_streams;
>  
> 