diff mbox series

usb-storage: Revert commit 747668dbc061 ("usb-storage: Set virt_boundary_mask to avoid SG overflows")

Message ID Pine.LNX.4.44L0.1910211145520.1673-100000@iolanthe.rowland.org (mailing list archive)
State New, archived
Headers show
Series usb-storage: Revert commit 747668dbc061 ("usb-storage: Set virt_boundary_mask to avoid SG overflows") | expand

Commit Message

Alan Stern Oct. 21, 2019, 3:48 p.m. UTC
Commit 747668dbc061 ("usb-storage: Set virt_boundary_mask to avoid SG
overflows") attempted to solve a problem involving scatter-gather I/O
and USB/IP by setting the virt_boundary_mask for mass-storage devices.

However, it now turns out that this interacts badly with commit
09324d32d2a0 ("block: force an unlimited segment size on queues with a
virt boundary"), which was added later.  A typical error message is:

	ehci-pci 0000:00:13.2: swiotlb buffer is full (sz: 327680 bytes),
	total 32768 (slots), used 97 (slots)

There is no longer any reason to keep the virt_boundary_mask setting
for usb-storage.  It was needed in the first place only for handling
devices with a block size smaller than the maxpacket size and where
the host controller was not capable of fully general scatter-gather
operation (that is, able to merge two SG segments into a single USB
packet).  But:

	High-speed or slower connections never use a bulk maxpacket
	value larger than 512;

	The SCSI layer does not handle block devices with a block size
	smaller than 512 bytes;

	All the host controllers capable of SuperSpeed operation can
	handle fully general SG;

	Since commit ea44d190764b ("usbip: Implement SG support to
	vhci-hcd and stub driver") was merged, the USB/IP driver can
	also handle SG.

Therefore all supported device/controller combinations should be okay
with no need for any special virt_boundary_mask.  So in order to fix
the swiotlb problem, this patch reverts commit 747668dbc061.

Reported-and-tested-by: Piergiorgio Sartor <piergiorgio.sartor@nexgo.de>
Link: https://marc.info/?l=linux-usb&m=157134199501202&w=2
Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
CC: Seth Bollinger <Seth.Bollinger@digi.com>
CC: <stable@vger.kernel.org>

---


[as1922]


 drivers/usb/storage/scsiglue.c |   10 ----------
 1 file changed, 10 deletions(-)

Comments

Christoph Hellwig Oct. 23, 2019, 1:53 a.m. UTC | #1
On Mon, Oct 21, 2019 at 11:48:06AM -0400, Alan Stern wrote:
> There is no longer any reason to keep the virt_boundary_mask setting
> for usb-storage.  It was needed in the first place only for handling
> devices with a block size smaller than the maxpacket size and where
> the host controller was not capable of fully general scatter-gather
> operation (that is, able to merge two SG segments into a single USB
> packet).  But:
> 
> 	High-speed or slower connections never use a bulk maxpacket
> 	value larger than 512;
> 
> 	The SCSI layer does not handle block devices with a block size
> 	smaller than 512 bytes;
> 
> 	All the host controllers capable of SuperSpeed operation can
> 	handle fully general SG;
> 
> 	Since commit ea44d190764b ("usbip: Implement SG support to
> 	vhci-hcd and stub driver") was merged, the USB/IP driver can
> 	also handle SG.
> 
> Therefore all supported device/controller combinations should be okay
> with no need for any special virt_boundary_mask.  So in order to fix
> the swiotlb problem, this patch reverts commit 747668dbc061.

That's great to know.  The same should also apply to uas, shouldn't
it?

Otherwise:

Acked-by: Christoph Hellwig <hch@lst.de>
Alan Stern Oct. 23, 2019, 2:12 p.m. UTC | #2
On Wed, 23 Oct 2019, Christoph Hellwig wrote:

> On Mon, Oct 21, 2019 at 11:48:06AM -0400, Alan Stern wrote:
> > There is no longer any reason to keep the virt_boundary_mask setting
> > for usb-storage.  It was needed in the first place only for handling
> > devices with a block size smaller than the maxpacket size and where
> > the host controller was not capable of fully general scatter-gather
> > operation (that is, able to merge two SG segments into a single USB
> > packet).  But:
> > 
> > 	High-speed or slower connections never use a bulk maxpacket
> > 	value larger than 512;
> > 
> > 	The SCSI layer does not handle block devices with a block size
> > 	smaller than 512 bytes;
> > 
> > 	All the host controllers capable of SuperSpeed operation can
> > 	handle fully general SG;
> > 
> > 	Since commit ea44d190764b ("usbip: Implement SG support to
> > 	vhci-hcd and stub driver") was merged, the USB/IP driver can
> > 	also handle SG.
> > 
> > Therefore all supported device/controller combinations should be okay
> > with no need for any special virt_boundary_mask.  So in order to fix
> > the swiotlb problem, this patch reverts commit 747668dbc061.
> 
> That's great to know.  The same should also apply to uas, shouldn't
> it?

Ah, yes, excellent point -- I had forgotten about it.  Additional patch 
coming up soon...

Alan Stern

> Otherwise:
> 
> Acked-by: Christoph Hellwig <hch@lst.de>
diff mbox series

Patch

Index: usb-devel/drivers/usb/storage/scsiglue.c
===================================================================
--- usb-devel.orig/drivers/usb/storage/scsiglue.c
+++ usb-devel/drivers/usb/storage/scsiglue.c
@@ -68,7 +68,6 @@  static const char* host_info(struct Scsi
 static int slave_alloc (struct scsi_device *sdev)
 {
 	struct us_data *us = host_to_us(sdev->host);
-	int maxp;
 
 	/*
 	 * Set the INQUIRY transfer length to 36.  We don't use any of
@@ -78,15 +77,6 @@  static int slave_alloc (struct scsi_devi
 	sdev->inquiry_len = 36;
 
 	/*
-	 * USB has unusual scatter-gather requirements: the length of each
-	 * scatterlist element except the last must be divisible by the
-	 * Bulk maxpacket value.  Fortunately this value is always a
-	 * power of 2.  Inform the block layer about this requirement.
-	 */
-	maxp = usb_maxpacket(us->pusb_dev, us->recv_bulk_pipe, 0);
-	blk_queue_virt_boundary(sdev->request_queue, maxp - 1);
-
-	/*
 	 * Some host controllers may have alignment requirements.
 	 * We'll play it safe by requiring 512-byte alignment always.
 	 */