2.6.17 merge additions
cover 2.6.17 merge additions
Following will be a set of patches that adjust 2.6.17 merges we had done differently in our internal tree, and that I think shouldn't get lost. There is no particular ordering among them, so if any seems questionable, others can still be applied.
Jan
Quoted 5 lines
Quoted 1 line
Quoted 1 line
"Jan Beulich" <[email protected]> 14.02.07 16:56 >>>
Following will be a set of patches that adjust 2.6.17 merges we had done differently in our internal tree, and that I think shouldn't get lost. There is no particular ordering among them, so if any seems questionable, others can still be applied.
Before submitting 2.6.18 stuff we were missing from the -unstable merge, is it possible to get an understanding why patches 5 (pc speaker device registration in domU) and 8 (early DMI scan) of the 2.6.17 set were not taken?
Thanks, Jan
Quoted 4 lines · Jan Beulich
Before submitting 2.6.18 stuff we were missing from the -unstable merge, is it possible to get an understanding why patches 5 (pc speaker device registration in domU) and 8 (early DMI scan) of the 2.6.17 set were not taken?
I'm undecided on those. Keeping pcspkr is pretty harmless (actually I'm not sure what the device registration is even for!) and removing it would just take us further from native. I don't have a really strong opinion on this one however. For the second patch: the principle of moving DMI scan earlier is nice, but this approach makes a horrible mess of the mm init code (which is quite nasty enough already!). I wonder whether we could come up with a two-stage mm init for x86/64 -- some very early Xen-specific stuff to get is into a state that is a bit more like native, would allow us to do things like DMI scan at a more appropriate time, and might clean up the paging_init() mess a bit rather than making it worse.
-- Keir
Quoted 11 lines
Quoted 1 line
Quoted 1 line
Keir Fraser <[email protected]> 02.03.07 10:19 >>>
Quoted 4 lines · Jan Beulich
Before submitting 2.6.18 stuff we were missing from the -unstable merge, is it possible to get an understanding why patches 5 (pc speaker device registration in domU) and 8 (early DMI scan) of the 2.6.17 set were not taken?
I'm undecided on those. Keeping pcspkr is pretty harmless (actually I'm not sure what the device registration is even for!) and removing it would just take us further from native. I don't have a really strong opinion on this
I'm not certain on this one either, as (like you) I'm not entirely clear what the consequences of registering this is without a real device underneath. Clearly, drivers/input/misc/pcspkr.c can be prevented from trying to play with port 61 (and registering an unusable input device) by not registering this device.
Quoted 7 lines
one however. For the second patch: the principle of moving DMI scan earlier is nice, but this approach makes a horrible mess of the mm init code (which is quite nasty enough already!). I wonder whether we could come up with a two-stage mm init for x86/64 -- some very early Xen-specific stuff to get is into a state that is a bit more like native, would allow us to do things like DMI scan at a more appropriate time, and might clean up the
paging_init()mess a bit rather than making it worse.
Isn't that patch just doing this - paging_init() is almost identical to native after the patch (the sole difference is the setting of init_mm.context.pinned). The only real addition (to find_early_table_space()) is the space reservation for the fixmap tables (so these can be set up earlier) and the stuff moved out of paging_init() into init_memory_mapping(). And I don't think it is reasonable to expect init_memory_mapping() to get very close to native.
Jan
Quoted 8 lines · Jan Beulich
Isn't that patch just doing this -
paging_init()is almost identical to native
after the patch (the sole difference is the setting of init_mm.context.pinned).
The only real addition (tofind_early_table_space()) is the space reservation for the fixmap tables (so these can be set up earlier) and the stuff moved out ofpaging_init()intoinit_memory_mapping(). And I don't think it is reasonable to expectinit_memory_mapping()to get very close to native.
Okay I probably should have done a three-way comparison with the native code as well, then. My first impression was definitely that the code looked more complicated rather than less. :-) I'll take another look.
-- Keir