On Wed, Jun 06, 2012 at 02:10:34PM -0700, Colin Cross wrote: > On Fri, May 11, 2012 at 5:17 PM, Anton Vorontsov > <anton.vorontsov@xxxxxxxxxx> wrote: > > This includes devices' memory (e.g. framebuffers or memory mapped > > EEPROMs on a local bus), as well as the normal RAM that we don't use > > for the main memory. > > > > For the normal (but unused) ram we could use kmaps, but this assumes > > highmem support, so we don't bother and just use the memory via > > ioremap. > > > > As a side effect, the following hack is possible: when used together > > with pstore_ram (new ramoops) module, we can limit the normal RAM region > > with mem= and then point ramoops to use the rest of the memory, e.g. > > > > mem=128M ramoops.mem_address=0x8000000 > > > > Sure, we could just reserve the region with memblock_reserve() early in > > the arch/ code, and then register a pstore_ram platform device pointing > > to the reserved region. It's still a viable option if platform wants > > to do so. > > > > Also, we might want to use IO accessors in case of a real device, > > but for now we don't bother (the old ramoops wasn't using it either, so > > at least we don't make things worse). > > This is long merged, but I remembered why I moved away from using > ioremap. The current code uses atomics to track the ringbuffer > positions, which results in ldrex and strex instructions on ARM. > ldrex and strex on memory that is mapped as Device memory (which is > what ioremap maps as) is implementation defined, and is unpredictable > at the architecture level. Makes sense, thanks for sharing! Fortunately, we still map things w/ vmap if pfn appears to be valid. :-) Thanks, -- Anton Vorontsov Email: cbouatmailru@xxxxxxxxx _______________________________________________ devel mailing list devel@xxxxxxxxxxxxxxxxxxxxxx http://driverdev.linuxdriverproject.org/mailman/listinfo/devel
