USB: usb/dma doc updates

This patch updates some of the documentation about DMA buffer management
for USB, and ways to avoid extra copying.  Our understanding of the issues
has improved over time.

 - Most drivers should *avoid* the dma-coherent allocators.  There are
   a few exceptions (like the HID driver).

 - Some methods are currently commented out; it seems folk writing
   USB drivers aren't doing performance tuning at that level yet.

 - Just avoid highmem; there's no good way to pass an "I can do highmem
   DMA" capability through a driver stack.  This is easy, everything
   already avoids highmem.  But it'd be nice if x86_32 systems with much
   physical memory could use it directly with network adapters and mass
   storage devices.  (Patch, anyone?)

Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

1 files changed, 37 insertions, 15 deletions

diff --git a/Documentation/usb/dma.txt b/Documentation/usb/dma.txt
index 62844aeba69c..e8b50b7de9d9 100644
--- a/Documentation/usb/dma.txt
+++ b/Documentation/usb/dma.txt
@@ -32,12 +32,15 @@ ELIMINATING COPIES
 It's good to avoid making CPUs copy data needlessly.  The costs can add up,
 and effects like cache-trashing can impose subtle penalties.
 
-- When you're allocating a buffer for DMA purposes anyway, use the buffer
-  primitives.  Think of them as kmalloc and kfree that give you the right
-  kind of addresses to store in urb->transfer_buffer and urb->transfer_dma,
-  while guaranteeing that no hidden copies through DMA "bounce" buffers will
-  slow things down.  You'd also set URB_NO_TRANSFER_DMA_MAP in
-  urb->transfer_flags:
+- If you're doing lots of small data transfers from the same buffer all
+  the time, that can really burn up resources on systems which use an
+  IOMMU to manage the DMA mappings.  It can cost MUCH more to set up and
+  tear down the IOMMU mappings with each request than perform the I/O!
+
+  For those specific cases, USB has primitives to allocate less expensive
+  memory.  They work like kmalloc and kfree versions that give you the right
+  kind of addresses to store in urb->transfer_buffer and urb->transfer_dma.
+  You'd also set URB_NO_TRANSFER_DMA_MAP in urb->transfer_flags:
 
 	void *usb_buffer_alloc (struct usb_device *dev, size_t size,
 		int mem_flags, dma_addr_t *dma);
@@ -45,6 +48,10 @@ and effects like cache-trashing can impose subtle penalties.
 	void usb_buffer_free (struct usb_device *dev, size_t size,
 		void *addr, dma_addr_t dma);
 
+  Most drivers should *NOT* be using these primitives; they don't need
+  to use this type of memory ("dma-coherent"), and memory returned from
+  kmalloc() will work just fine.
+
   For control transfers you can use the buffer primitives or not for each
   of the transfer buffer and setup buffer independently.  Set the flag bits
   URB_NO_TRANSFER_DMA_MAP and URB_NO_SETUP_DMA_MAP to indicate which
@@ -54,29 +61,39 @@ and effects like cache-trashing can impose subtle penalties.
   The memory buffer returned is "dma-coherent"; sometimes you might need to
   force a consistent memory access ordering by using memory barriers.  It's
   not using a streaming DMA mapping, so it's good for small transfers on
-  systems where the I/O would otherwise tie up an IOMMU mapping.  (See
+  systems where the I/O would otherwise thrash an IOMMU mapping.  (See
   Documentation/DMA-mapping.txt for definitions of "coherent" and "streaming"
   DMA mappings.)
 
   Asking for 1/Nth of a page (as well as asking for N pages) is reasonably
   space-efficient.
 
+  On most systems the memory returned will be uncached, because the
+  semantics of dma-coherent memory require either bypassing CPU caches
+  or using cache hardware with bus-snooping support.  While x86 hardware
+  has such bus-snooping, many other systems use software to flush cache
+  lines to prevent DMA conflicts.
+
 - Devices on some EHCI controllers could handle DMA to/from high memory.
-  Driver probe() routines can notice this using a generic DMA call, then
-  tell higher level code (network, scsi, etc) about it like this:
 
-	if (dma_supported (&intf->dev, 0xffffffffffffffffULL))
-		net->features |= NETIF_F_HIGHDMA;
+  Unfortunately, the current Linux DMA infrastructure doesn't have a sane
+  way to expose these capabilities ... and in any case, HIGHMEM is mostly a
+  design wart specific to x86_32.  So your best bet is to ensure you never
+  pass a highmem buffer into a USB driver.  That's easy; it's the default
+  behavior.  Just don't override it; e.g. with NETIF_F_HIGHDMA.
 
-  That can eliminate dma bounce buffering of requests that originate (or
-  terminate) in high memory, in cases where the buffers aren't allocated
-  with usb_buffer_alloc() but instead are dma-mapped.
+  This may force your callers to do some bounce buffering, copying from
+  high memory to "normal" DMA memory.  If you can come up with a good way
+  to fix this issue (for x86_32 machines with over 1 GByte of memory),
+  feel free to submit patches.
 
 
 WORKING WITH EXISTING BUFFERS
 
 Existing buffers aren't usable for DMA without first being mapped into the
-DMA address space of the device.
+DMA address space of the device.  However, most buffers passed to your
+driver can safely be used with such DMA mapping.  (See the first section
+of DMA-mapping.txt, titled "What memory is DMA-able?")
 
 - When you're using scatterlists, you can map everything at once.  On some
   systems, this kicks in an IOMMU and turns the scatterlists into single
@@ -114,3 +131,8 @@ DMA address space of the device.
   The calls manage urb->transfer_dma for you, and set URB_NO_TRANSFER_DMA_MAP
   so that usbcore won't map or unmap the buffer.  The same goes for
   urb->setup_dma and URB_NO_SETUP_DMA_MAP for control requests.
+
+Note that several of those interfaces are currently commented out, since
+they don't have current users.  See the source code.  Other than the dmasync
+calls (where the underlying DMA primitives have changed), most of them can
+easily be commented back in if you want to use them.