On Jan 4, 2011, at 4:57 AM, Manuel Bouyer wrote: > On Tue, Jan 04, 2011 at 02:32:02AM -0800, Matt Thomas wrote: >> Not really. A lot of device can only do 32bit DMA transfers so without some >> type assistance (like the alpha has) you are restricted to DMA to the first >> 4GB of RAM. If you are doing DMA to an address >= 4GB, the system will >> stage the data in a buffer < 4GB for the DMA and then move the data to its >> ultimate destination (for a read). For a write it copies to the "bounce" >> buffer, and then does a DMA. >> >> The point is if you are allocating a mbuf or a usb buffer, it makes sense to >> allocate from the first 4GB if you can to avoid using bounce buffers. >> >> It's just a preference. > > Is there some provision to always keep some <4GB pages free (with some being > quite large for e.g. USB descriptors) so there is always some bounce buffers > available ? AFAIK, if you allocate from the <4Gb free list, you can eat all > DMA-capable RAM. > I fixed this issue in arch/x86/x86/x86_machdep.c 1.37, I wonder if your change > has reintroduced this problem ...
Since it's not turned on by default, I doubt it.