This patch implements hypercalls allowing a PAPR guest to resize its own
hash page table. This will eventually allow for more flexible memory
hotplug.
The implementation is partially asynchronous, handled in a special thread
running the hpt_prepare_thread() function. The state of a pending resize
is stored in SPAPR_MACHINE->pending_hpt.
The H_RESIZE_HPT_PREPARE hypercall will kick off creation of a new HPT, or,
if one is already in progress, monitor it for completion. If there is an
existing HPT resize in progress that doesn't match the size specified in
the call, it will cancel it, replacing it with a new one matching the
given size.
The H_RESIZE_HPT_COMMIT completes transition to a resized HPT, and can only
be called successfully once H_RESIZE_HPT_PREPARE has successfully
completed initialization of a new HPT. The guest must ensure that there
are no concurrent accesses to the existing HPT while this is called (this
effectively means stop_machine() for Linux guests).
For now H_RESIZE_HPT_COMMIT goes through the whole old HPT, rehashing each
HPTE into the new HPT. This can have quite high latency, but it seems to
be of the order of typical migration downtime latencies for HPTs of size
up to ~2GiB (which would be used in a 256GiB guest).
In future we probably want to move more of the rehashing to the "prepare"
phase, by having H_ENTER and other hcalls update both current and
pending HPTs. That's a project for another day, but should be possible
without any changes to the guest interface.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
SDR_64_HTABORG, which indicates the bits of the SDR1 register to use for
the base of a 64-bit machine's hashed page table (HPT) isn't correct. It
includes the top 46 bits of the register, but in fact the top 4 bits must
be zero (according to the ISA v2.07). No actual implementation has
supported close to 2^60 bytes of physical address space, so it's kind of
irrelevant, but we might as well correct this.
In addition, although we checked for bad size values in SDR1, we never
reported an error if entirely invalid bits were set there. Add this check
to ppc_store_sdr1().
Reported-by: Suraj Jitindar Singh <sjitindarsingh@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The function ppc_hash64_set_sdr1 basically checked the htabsize and set an
error if it was too big, otherwise it just stored the value in SPR_SDR1.
Given that the only function which calls ppc_hash64_set_sdr1() is
ppc_store_sdr1(), why not handle the checking in ppc_store_sdr1() to avoid
the extra function call. Note that ppc_store_sdr1() already stores the
value in SPR_SDR1 anyway, so we were doing it twice.
Signed-off-by: Suraj Jitindar Singh <sjitindarsingh@gmail.com>
[dwg: Remove unnecessary error temporary]
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The pseries machine type implements the behaviour of a PAPR compliant
hypervisor, without actually executing such a hypervisor on the virtual
CPU. To do this we need some hooks in the CPU code to make hypervisor
facilities get redirected to the machine instead of emulated internally.
For hypercalls this is managed through the cpu->vhyp field, which points
to a QOM interface with a method implementing the hypercall.
For the hashed page table (HPT) - also a hypervisor resource - we use an
older hack. CPUPPCState has an 'external_htab' field which when non-NULL
indicates that the HPT is stored in qemu memory, rather than within the
guest's address space.
For consistency - and to make some future extensions easier - this merges
the external HPT mechanism into the vhyp mechanism. Methods are added
to vhyp for the basic operations the core hash MMU code needs: map_hptes()
and unmap_hptes() for reading the HPT, store_hpte() for updating it and
hpt_mask() to retrieve its size.
To match this, the pseries machine now sets these vhyp fields in its
existing vhyp class, rather than reaching into the cpu object to set the
external_htab field.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Suraj Jitindar Singh <sjitindarsingh@gmail.com>
CPUPPCState includes fields htab_base and htab_mask which store the base
address (GPA) and size (as a mask) of the guest's hashed page table (HPT).
These are set when the SDR1 register is updated.
Keeping these in sync with the SDR1 is actually a little bit fiddly, and
probably not useful for performance, since keeping them expands the size of
CPUPPCState. It also makes some upcoming changes harder to implement.
This patch removes these fields, in favour of calculating them directly
from the SDR1 contents when necessary.
This does make a change to the behaviour of attempting to write a bad value
(invalid HPT size) to the SDR1 with an mtspr instruction. Previously, the
bad value would be stored in SDR1 and could be retrieved with a later
mfspr, but the HPT size as used by the softmmu would be, clamped to the
allowed values. Now, writing a bad value is treated as a no-op. An error
message is printed in both new and old versions.
I'm not sure which behaviour, if either, matches real hardware. I don't
think it matters that much, since it's pretty clear that if an OS writes
a bad value to SDR1, it's not going to boot.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Accesses to the hashed page table (HPT) are complicated by the fact that
the HPT could be in one of three places:
1) Within guest memory - when we're emulating a full guest CPU at the
hardware level (e.g. powernv, mac99, g3beige)
2) Within qemu, but outside guest memory - when we're emulating user and
supervisor instructions within TCG, but instead of emulating
the CPU's hypervisor mode, we just emulate a hypervisor's behaviour
(pseries in TCG or KVM-PR)
3) Within the host kernel - a pseries machine using KVM-HV
acceleration. Mostly accesses to the HPT are handled by KVM,
but there are a few cases where qemu needs to access it via a
special fd for the purpose.
In order to batch accesses to the fd in case (3), we use a somewhat awkward
ppc_hash64_start_access() / ppc_hash64_stop_access() pair, which for case
(3) reads / releases several HPTEs from the kernel as a batch (usually a
whole PTEG). For cases (1) & (2) it just returns an address value. The
actual HPTE load helpers then need to interpret the returned token
differently in the 3 cases.
This patch keeps the same basic structure, but simplfiies the details.
First start_access() / stop_access() are renamed to map_hptes() and
unmap_hptes() to make their operation more obvious. Second, map_hptes()
now always returns a qemu pointer, which can always be used in the same way
by the load_hpte() helpers. In case (1) it comes from address_space_map()
in case (2) directly from qemu's HPT buffer and in case (3) from a
temporary buffer read from the KVM fd.
While we're at it, make things a bit more consistent in terms of types and
variable names: avoid variables named 'index' (it shadows index(3) which
can lead to confusing results), use 'hwaddr ptex' for HPTE indices and
uint64_t for each of the HPTE words, use ptex throughout the call stack
instead of pte_offset in some places (we still need that at the bottom
layer, but nowhere else).
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
When a 'pseries' guest is running with KVM-HV, the guest's hashed page
table (HPT) is stored within the host kernel, so it is not directly
accessible to qemu. Most of the time, qemu doesn't need to access it:
we're using the hardware MMU, and KVM itself implements the guest
hypercalls for manipulating the HPT.
However, qemu does need access to the in-KVM HPT to implement
get_phys_page_debug() for the benefit of the gdbstub, and maybe for
other debug operations.
To allow this, 7c43bca "target-ppc: Fix page table lookup with kvm
enabled" added kvmppc_hash64_read_pteg() to target/ppc/kvm.c to read
in a batch of HPTEs from the KVM table. Unfortunately, there are a
couple of problems with this:
First, the name of the function implies it always reads a whole PTEG
from the HPT, but in fact in some cases it's used to grab individual
HPTEs (which ends up pulling 8 HPTEs, not aligned to a PTEG from the
kernel).
Second, and more importantly, the code to read the HPTEs from KVM is
simply wrong, in general. The data from the fd that KVM provides is
designed mostly for compact migration rather than this sort of one-off
access, and so needs some decoding for this purpose. The current code
will work in some cases, but if there are invalid HPTEs then it will
not get sane results.
This patch rewrite the HPTE reading function to have a simpler
interface (just read n HPTEs into a caller provided buffer), and to
correctly decode the stream from the kernel.
For consistency we also clean up the similar function for altering
HPTEs within KVM (introduced in c138593 "target-ppc: Update
ppc_hash64_store_hpte to support updating in-kernel htab").
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
We are calculating the authority mask register key value wrong.
The pte entry contains the key value with the two upper bits and the three
lower bits stored separately. We should use these two portions to get a 5
bit value, not or them together which will only give us a 3 bit value.
Fix this.
Signed-off-by: Suraj Jitindar Singh <sjitindarsingh@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
We've currently got 18 architectures in QEMU, and thus 18 target-xxx
folders in the root folder of the QEMU source tree. More architectures
(e.g. RISC-V, AVR) are likely to be included soon, too, so the main
folder of the QEMU sources slowly gets quite overcrowded with the
target-xxx folders.
To disburden the main folder a little bit, let's move the target-xxx
folders into a dedicated target/ folder, so that target-xxx/ simply
becomes target/xxx/ instead.
Acked-by: Laurent Vivier <laurent@vivier.eu> [m68k part]
Acked-by: Bastian Koppelmann <kbastian@mail.uni-paderborn.de> [tricore part]
Acked-by: Michael Walle <michael@walle.cc> [lm32 part]
Acked-by: Cornelia Huck <cornelia.huck@de.ibm.com> [s390x part]
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com> [s390x part]
Acked-by: Eduardo Habkost <ehabkost@redhat.com> [i386 part]
Acked-by: Artyom Tarasenko <atar4qemu@gmail.com> [sparc part]
Acked-by: Richard Henderson <rth@twiddle.net> [alpha part]
Acked-by: Max Filippov <jcmvbkbc@gmail.com> [xtensa part]
Reviewed-by: David Gibson <david@gibson.dropbear.id.au> [ppc part]
Acked-by: Edgar E. Iglesias <edgar.iglesias@xilinx.com> [crisµblaze part]
Acked-by: Guan Xuetao <gxt@mprc.pku.edu.cn> [unicore32 part]
Signed-off-by: Thomas Huth <thuth@redhat.com>
