device-assignment: difference between assigned_dev_iomem_map and ...map_slow
Hi,
latest qemu-kvm bails out on cleanup as it tries to call cpu_register_physical_memory with a zero-sized region of an assigned device. That made me dig into the setup/cleanup of memory mapped io regions, trying to consolidate and fix the code.
What are the differences between normal and slow mmio regions? The former are mapped directly to the physical device (via qemu_ram_alloc_from_ptr + cpu_register_physical_memory), the latter have to be dispatched in user land (thus cpu_register_io_memory + cpu_register_physical_memory), right?
But why do we need to postpone cpu_register_io_memory to assigned_dev_iomem_map_slow? It looks like that's effectively the only difference between both mapping callbacks (subtracting some bugs and dead code). Can't we set up the io region in assigned_dev_register_regions analogously to normal regions?
BTW, the current code is leaking the slow io region on cleanup.
Comments appreciated, will translate them into a cleanup patch series.
Thanks,
Jan (who wanted to do something completely different...)
Signature · 2 lines
Siemens AG, Corporate Technology, CT T DE IT 1 Corporate Competence Center Embedded Linux
Quoted 6 lines · Jan Kiszka
Hi,
latest qemu-kvm bails out on cleanup as it tries to call
cpu_register_physical_memorywith a zero-sized region of an assigned device. That made me dig into the setup/cleanup of memory mapped io regions, trying to consolidate and fix the code.
The teardown is gated by memory_index, so that means it's tearing down a region that wasn't mapped by the guest?
Quoted 5 lines
What are the differences between normal and slow mmio regions? The former are mapped directly to the physical device (via
qemu_ram_alloc_from_ptr+cpu_register_physical_memory), the latter have to be dispatched in user land (thuscpu_register_io_memory+cpu_register_physical_memory), right?
Right.
Quoted 5 lines
But why do we need to postpone
cpu_register_io_memorytoassigned_dev_iomem_map_slow? It looks like that's effectively the only difference between both mapping callbacks (subtracting some bugs and dead code). Can't we set up the io region inassigned_dev_register_regionsanalogously to normal regions?
I imagine it was an attempt not to overload memory_index, for tests like the one above, but apparently it's not working out so well. I don't see any reason we shouldn't do the cpu_register_io_memory on setup.
Quoted 1 line
BTW, the current code is leaking the slow io region on cleanup.
Yep, I don't see an cpu_unregister_io_memory() for that region either.
Quoted 1 line
Comments appreciated, will translate them into a cleanup patch series.
Thanks!
Alex