On Wed, Jan 09, 2013 at 09:28:47PM +0000, Russell King - ARM Linux wrote: > On Wed, Jan 09, 2013 at 09:43:05PM +0100, Thierry Reding wrote: > > The I/O map cache is used to map large regions of physical memory in > > smaller chunks to avoid running out of vmalloc()/ioremap() space. > > > > Signed-off-by: Thierry Reding <thierry.reding@xxxxxxxxxxxxxxxxx> > > We already have a means where we record the mappings which ioremap() > creates. If you look at /proc/vmallocinfo, you'll notice lines such > as: > > 0xf7b72000-0xf7b74000 8192 e1000_probe+0x291/0xa68 [e1000e] phys=fc025000 ioremap > > which gives you the virtual address range, physical address and type > of the mapping. Why do we need a duplicated data structure? > > Moreover, you seem to suggest that you want to break up a large > ioremap() mapping into several smaller mappings. Why? The idea > behind ioremap() is that this relationship holds true: > > ptr = ioremap(cookie + n, size); > > For any 'n' in the range 0 .. size, the location shall be accessible > via ptr + n when using the IO space accessors. If you're going to > break up a mapping into several smaller ones, this no longer holds > true. > > If the problem is that you're ioremapping huge address ranges because > you're passing larger-than-required resources to devices, then that's > part of the problem too. The resource is not larger than required. It's just that in order to address the extended configuration space of all PCI devices we require a full 256 MiB window. What this patch tries to achieve is allow a driver to take a large region and ioremap() it page by page on an as needed basis. Thierry
Attachment:
pgpkDFXFGRBbz.pgp
Description: PGP signature
