On Thursday 03 April 2008 09:54:51 am Rene Herman wrote:However, now that you made me look closer and in context -- there's actually a possibly somewhat serious problem here.
isapnp_read_resources() stores the resources as read from the hardware at the index in the table that matches the actual index in the hardware and isapnp_set_resources() stores them back into those same hardware indices.
Now by using pnp_add_foo_resource() which just scans for the first _UNSET resource, the resources might not end up in the same linear position in table/list if intermediate resources were unset in hardware (!ret). A subsequent isapnp_set_resources() would them restore the value to the wrong hardware index.
The IORESOURCE_ flags currently reserve too few bits (IORESOURCE_BITS, 8) to be able to store the hardware index: IORESOURCE_MEM and IORESOURCE_DMA need 2 and 1 respectively and there are 1 and 0 available respectively. It's ofcourse possible to hijack a few more bits in IORESOURCE_ flags but you're turning this into a list. I suppose the idea is to make it a simple list of struct resource, but perhaps a resource-private "driver_data" sort of field comes in handy for more than this already? Swiping more of IORESOURCE_ is a bit ugly...
In any case, I missed this, but ISAPnP is still (at least in principle) broken with the current set therefore.
Hmm... you're right. And I think it could bite PNPBIOS and PNPACPI
as well -- they don't read/write hardware registers directly, but the
firmware still depends on preserving the resource order. I'll have to
ponder that for a while.