Re: [PATCH v2] Add support for usbfs zerocopy.

[Alan Stern]

On Wed, 6 Jan 2016, Lingzhu Xiang wrote:

> On Wed, Nov 25, 2015 at 7:19 PM, Steinar H. Gunderson <sesse@xxxxxxxxxx> wrote:
> > Add a new interface for userspace to preallocate memory that can be
> > used with usbfs. This gives two primary benefits:
> 
> I got this when trying to allocate a little bit large buffer (~4MB)
> using the new userspace libusb_dev_mem_alloc():
> 
> > [ 1706.212407] usb 2-1.1: reset SuperSpeed USB device number 3 using xhci_hcd
> > [ 1706.234823] xhci_hcd 0000:00:14.0: swiotlb buffer is full (sz: 4325376 bytes)
> > [ 1706.234827] swiotlb: coherent allocation failed for device 0000:00:14.0 size=4325376
> > [ 1706.234830] CPU: 1 PID: 3233 Comm: Protonect Tainted: G     U  W       4.4.0-rc8-amd64 #1 Debian 4.4~rc8-1~exp1
> > [ 1706.234831] Hardware name: LENOVO 20ALCTO1WW/20ALCTO1WW, BIOS GIET76WW (2.26 ) 08/27/2014
> > [ 1706.234833]  0000000000000000 000000000f50c266 ffffffff812e6019 ffffffffffffffff
> > [ 1706.234836]  ffffffff8130dc45 ffff88020000000b 0000000000420000 ffffffff81a2a0e0
> > [ 1706.234838]  ffff880206263d80 0000000000000000 ffff88021c892f40 0000000000420040
> > [ 1706.234841] Call Trace:
> > [ 1706.234847]  [<ffffffff812e6019>] ? dump_stack+0x40/0x57
> > [ 1706.234851]  [<ffffffff8130dc45>] ? swiotlb_alloc_coherent+0x135/0x150
> > [ 1706.234867]  [<ffffffffa021deb1>] ? hcd_buffer_alloc+0xb1/0x130 [usbcore]
> > [ 1706.234875]  [<ffffffffa0221ab5>] ? usbdev_mmap+0xa5/0x1b0 [usbcore]
> > [ 1706.234880]  [<ffffffff813bbc25>] ? tty_insert_flip_string_fixed_flag+0x85/0xe0
> > [ 1706.234885]  [<ffffffff8119af87>] ? mmap_region+0x3e7/0x660
> > [ 1706.234888]  [<ffffffff8119b536>] ? do_mmap+0x336/0x420
> > [ 1706.234892]  [<ffffffff8118213f>] ? vm_mmap_pgoff+0xaf/0xf0
> > [ 1706.234895]  [<ffffffff811999dd>] ? SyS_mmap_pgoff+0x1ad/0x270
> > [ 1706.234898]  [<ffffffff811d53b6>] ? SyS_write+0x76/0xc0
> > [ 1706.234903]  [<ffffffff815829f2>] ? system_call_fast_compare_end+0xc/0x67
> 
> I understand there are some requirements on the allocation such that
> large blocks are not always available. But what is the proper way to
> determine the upper limit of the size such that the user can avoid
> generating warnings like this? (Also, the application really wants to
> be able to allocate large buffers, maybe tune swiotlb=?.)

It's debatable whether this should have generated a warning.  Why 
doesn't dma_alloc_coherent() simply fail silently?

Alan Stern

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/