Messages in this thread | | | Date | Wed, 7 Apr 2010 12:04:33 -0400 (EDT) | From | Alan Stern <> | Subject | Re: USB transfer_buffer allocations on 64bit systems |
| |
On Wed, 7 Apr 2010, Greg KH wrote:
> > FWIW, most drivers I've seen in the past hours use a wild mix of > > kmalloc(), kzalloc(), kcalloc() and usb_buffer_alloc(). That should > > really be unified.
Well, kcalloc can easily be replaced by kzalloc, right? Or the equivalent.
The extra overhead of initializing the memory to 0 isn't present in kmalloc, so we need to maintain the distinction between kmalloc and kzalloc.
However usb_buffer_alloc is fundmentally different from all the others.
> Yes, if it is necessary that we have to handle this type of crappy > hardware, then it all needs to be unified. Or at least unified into 2 > types of calls, one that needs the bounce buffer fun (what > usb_buffer_alloc() does today), and one that doesn't (usb_kzalloc() > perhaps?)
usb_buffer_alloc has very little to do with bounce buffers. Its purpose is to allocate dma-consistent memory, that it, memory which does not need to be mapped for DMA before each I/O operation and unmapped afterward.
The mapping and unmapping operations aren't extremely time consuming, so in general it makes sense to avoid them only when the _same_ buffer is going to be used for many I/O operations during a short period of time. For instance, it makes sense for audio and video, where all the data streams through a small set of buffers arranged in a ring.
But for most other uses it makes no sense. Especially since some platforms have limited amounts of consistent memory available.
Alan Stern
| |