mirror of https://github.com/xemu-project/xemu.git
75d08a4072
The description in the Arm ARM of the requirements of FEAT_BBM is
admirably clear on the guarantees it provides software, but slightly
more obscure on what that means for implementations. The description
of the equivalent SMMU feature in the SMMU specification (IHI0070D.b
section 3.21.1) is perhaps a bit more detailed and includes some
example valid implementation choices. (The SMMU version of this
feature is slightly tighter than the CPU version: the CPU is permitted
to raise TLB Conflict aborts in some situations that the SMMU may
not. This doesn't matter for QEMU because we don't want to do TLB
Conflict aborts anyway.)
The informal summary of FEAT_BBM is that it is about permitting an OS
to switch a range of memory between "covered by a huge page" and
"covered by a sequence of normal pages" without having to engage in
the 'break-before-make' dance that has traditionally been
necessary. The 'break-before-make' sequence is:
* replace the old translation table entry with an invalid entry
* execute a DSB insn
* execute a broadcast TLB invalidate insn
* execute a DSB insn
* write the new translation table entry
* execute a DSB insn
The point of this is to ensure that no TLB can simultaneously contain
TLB entries for the old and the new entry, which would traditionally
be UNPREDICTABLE (allowing the CPU to generate a TLB Conflict fault
or to use a random mishmash of values from the old and the new
entry). FEAT_BBM level 2 says "for the specific case where the only
thing that changed is the size of the block, the TLB is guaranteed
not to do weird things even if there are multiple entries for an
address", which means that software can now do:
* replace old translation table entry with new entry
* DSB
* broadcast TLB invalidate
* DSB
As the SMMU spec notes, valid ways to do this include:
* if there are multiple entries in the TLB for an address,
choose one of them and use it, ignoring the others
* if there are multiple entries in the TLB for an address,
throw them all out and do a page table walk to get a new one
QEMU's page table walk implementation for Arm CPUs already meets the
requirements for FEAT_BBM level 2. When we cache an entry in our TCG
TLB, we do so only for the specific (non-huge) page that the address
is in, and there is no way for the TLB data structure to ever have
more than one TLB entry for that page. (We handle huge pages only in
that we track what part of the address space is covered by huge pages
so that a TLB invalidate operation for an address in a huge page
results in an invalidation of the whole TLB.) We ignore the Contiguous
bit in page table entries, so we don't have to do anything for the
parts of FEAT_BBM that deal with changis to the Contiguous bit.
FEAT_BBM level 2 also requires that the nT bit in block descriptors
must be ignored; since commit
|
||
|---|---|---|
| .. | ||
| _templates | ||
| about | ||
| config | ||
| devel | ||
| interop | ||
| specs | ||
| sphinx | ||
| sphinx-static | ||
| spin | ||
| system | ||
| tools | ||
| user | ||
| COLO-FT.txt | ||
| block-replication.txt | ||
| bypass-iommu.txt | ||
| colo-proxy.txt | ||
| conf.py | ||
| defs.rst.inc | ||
| hyperv.txt | ||
| igd-assign.txt | ||
| image-fuzzer.txt | ||
| index.rst | ||
| memory-hotplug.txt | ||
| meson.build | ||
| multi-thread-compression.txt | ||
| multiseat.txt | ||
| nvdimm.txt | ||
| pci_expander_bridge.txt | ||
| pcie.txt | ||
| pcie_pci_bridge.txt | ||
| pcie_sriov.txt | ||
| pvrdma.txt | ||
| qcow2-cache.txt | ||
| qdev-device-use.txt | ||
| qemu-option-trace.rst.inc | ||
| qemupciserial.inf | ||
| rdma.txt | ||
| replay.txt | ||
| spice-port-fqdn.txt | ||
| throttle.txt | ||
| u2f.txt | ||
| xbzrle.txt | ||
| xen-save-devices-state.txt | ||