On 22/06/2021 17:10, Jan Beulich wrote: > On 22.06.2021 17:39, Andrew Cooper wrote: >> This reverts commit aad7b5c11d51d57659978e04702ac970906894e8. >> >> The change from OvmfX64 to OvmfXen causes a change in behaviour, whereby >> OvmfXen maps its shared info page at the top of address space. When trying >> to >> migrate such a domain, XENMEM_maximum_gpfn returns a very large value. This >> has uncovered multiple issues: >> >> 1) The userspace hypercall wrappers truncate all return values to int on >> Linux and Solaris. This needs fixing in Xen. >> 2) 32bit toolstacks can't migrate any domain with RAM above the 2^40 mark, >> because of virtual address constraints. This needs fixing in OVMF. > And I suspect even that presently enforce boundary of 2^40 is actually > too high, and things still wouldn't work when getting close. At the > very least the tool stack then depends on a fairly big chunk of memory > (2^30 bytes) to be available in one single, virtually contiguous piece. > Iirc 32-bit Linux can be configured to not even leave this much space > for user mode.
I tested it once during Migration v2 development, and it worked for me, but I do expect that that is as much testing as it has had since... A 3G/1G split is the default under multiple 32bit kernels, and the allocation is made right at the start, so there is a reasonable chance of finding space. After all, it only needs 4k alignment. Whether ASLR has changed the chances in the meantime remains to be seen, but honestly - 32bit toolstacks on x86 really don't exist in production any more, and Arm still hasn't implemented logdirty support, so the limit has little practical consequence. ~Andrew
