summaryrefslogtreecommitdiff
path: root/include
diff options
context:
space:
mode:
authorPete Zaitcev <zaitcev@redhat.com>2009-06-11 08:53:20 -0600
committerGreg Kroah-Hartman <gregkh@suse.de>2009-09-23 06:46:19 -0700
commit4e9e92003529e5c7bb11281f7c2c9b3fe8858403 (patch)
tree07169c9a996a119aebb5865a76ff1177afe90a22 /include
parentf4e2332cfcf900e0a926c4e0fc35f751bcbcaa1b (diff)
downloadlwn-4e9e92003529e5c7bb11281f7c2c9b3fe8858403.tar.gz
lwn-4e9e92003529e5c7bb11281f7c2c9b3fe8858403.zip
USB: usbmon: end ugly tricks with DMA peeking
This patch fixes crashes when usbmon attempts to access GART aperture. The old code attempted to take a bus address and convert it into a virtual address, which clearly was impossible on systems with actual IOMMUs. Let us not persist in this foolishness, and use transfer_buffer in all cases instead. I think downsides are negligible. The ones I see are: - A driver may pass an address of one buffer down as transfer_buffer, and entirely different entity mapped for DMA, resulting in misleading output of usbmon. Note, however, that PIO based controllers would do transfer the same data that usbmon sees here. - Out of tree drivers may crash usbmon if they store garbage in transfer_buffer. I inspected the in-tree drivers, and clarified the documentation in comments. - Drivers that use get_user_pages will not be possible to monitor. I only found one driver with this problem (drivers/staging/rspiusb). - Same happens with with usb_storage transferring from highmem, but it works fine on 64-bit systems, so I think it's not a concern. At least we don't crash anymore. Why didn't we do this in 2.6.10? That's because back in those days it was popular not to fill in transfer_buffer, so almost all traffic would be invisible (e.g. all of HID was like that). But now, the tree is almost 100% PIO friendly, so we can do the right thing at last. Signed-off-by: Pete Zaitcev <zaitcev@redhat.com> Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Diffstat (limited to 'include')
-rw-r--r--include/linux/usb.h19
1 files changed, 13 insertions, 6 deletions
diff --git a/include/linux/usb.h b/include/linux/usb.h
index 19fabc487beb..3b45a0d27b80 100644
--- a/include/linux/usb.h
+++ b/include/linux/usb.h
@@ -1036,9 +1036,10 @@ typedef void (*usb_complete_t)(struct urb *);
* @transfer_flags: A variety of flags may be used to affect how URB
* submission, unlinking, or operation are handled. Different
* kinds of URB can use different flags.
- * @transfer_buffer: This identifies the buffer to (or from) which
- * the I/O request will be performed (unless URB_NO_TRANSFER_DMA_MAP
- * is set). This buffer must be suitable for DMA; allocate it with
+ * @transfer_buffer: This identifies the buffer to (or from) which the I/O
+ * request will be performed unless URB_NO_TRANSFER_DMA_MAP is set
+ * (however, do not leave garbage in transfer_buffer even then).
+ * This buffer must be suitable for DMA; allocate it with
* kmalloc() or equivalent. For transfers to "in" endpoints, contents
* of this buffer will be modified. This buffer is used for the data
* stage of control transfers.
@@ -1104,9 +1105,15 @@ typedef void (*usb_complete_t)(struct urb *);
* allocate a DMA buffer with usb_buffer_alloc() or call usb_buffer_map().
* When these transfer flags are provided, host controller drivers will
* attempt to use the dma addresses found in the transfer_dma and/or
- * setup_dma fields rather than determining a dma address themselves. (Note
- * that transfer_buffer and setup_packet must still be set because not all
- * host controllers use DMA, nor do virtual root hubs).
+ * setup_dma fields rather than determining a dma address themselves.
+ *
+ * Note that transfer_buffer must still be set if the controller
+ * does not support DMA (as indicated by bus.uses_dma) and when talking
+ * to root hub. If you have to trasfer between highmem zone and the device
+ * on such controller, create a bounce buffer or bail out with an error.
+ * If transfer_buffer cannot be set (is in highmem) and the controller is DMA
+ * capable, assign NULL to it, so that usbmon knows not to use the value.
+ * The setup_packet must always be set, so it cannot be located in highmem.
*
* Initialization:
*