Re: [RFC v2 0/1] memory: Delete assertion in memory_region_unregister_iommu_notifier

[Paolo Bonzini <pbonzini@redhat.com>]

On 03/07/20 09:39, Eugenio Perez Martin wrote:
> #4  0x0000555555888171 in memory_region_notify_one
> (notifier=0x7ffde0487fa8, entry=0x7ffde5dfe200) at
> /home/qemu/memory.c:1918
> 1918        assert(entry->iova >= notifier->start && entry_end <=
> notifier->end);
> (gdb) p *entry
> $1 = {target_as = 0x555556f6c050, iova = 0, translated_addr = 0,
> addr_mask = 18446744073709551615, perm = IOMMU_NONE}

Oh, I see now.  I am worried that an IOMMU notifier could interpret the
IOMMUTLBEntry incorrectly if there is only partial overlap.  There are
various possibilities:

1) create another IOMMUTLBEntry like

	hwaddr offset = notifier->start > entry->iova ? notifier->start -
entry->iova : 0;
	IOMMUTLBEntry partial = {
	    .target_as = entry->target_as,
	    .iova = entry->iova + offset,
	    .translated_addr = entry->translated_addr + offset,
	    .addr_mask = MIN(entry->addr_mask, notifier->end - notifier->start),
	    .perm = entry->perm
	};

The addr_mask however would not be a mask if the notifier is not
naturally aligned

2) pass the offset/size pair (computed as above) as extra arguments to
the IOMMUNotify function

3) add a function to compute the offset/size and call it in the notifier

You choose. :)

Paolo