This patch defines a new ECAM region located after the 256GB limit.
The virt machine state is augmented with a new highmem_ecam field
which guards the usage of this new ECAM region instead of the legacy
16MB one. With the highmem ECAM region, up to 256 PCIe buses can be
used.
Signed-off-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Andrew Jones <drjones@redhat.com>
Message-id: 1529072910-16156-9-git-send-email-eric.auger@redhat.com
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
This patch allows the creation of a GICv3 node with 1 or 2
redistributor regions depending on the number of smu_cpus.
The second redistributor region is located just after the
existing RAM region, at 256GB and contains up to up to 512 vcpus.
Please refer to kernel documentation for further node details:
Documentation/devicetree/bindings/interrupt-controller/arm,gic-v3.txt
Signed-off-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Andrew Jones <drjones@redhat.com>
Message-id: 1529072910-16156-6-git-send-email-eric.auger@redhat.com
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
To prepare for multiple redistributor regions, we introduce
an array of uint32_t properties that stores the redistributor
count of each redistributor region.
Non accelerated VGICv3 only supports a single redistributor region.
The capacity of all redist regions is checked against the number of
vcpus.
Machvirt is updated to set those properties, ie. a single
redistributor region with count set to the number of vcpus
capped by 123.
Signed-off-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Andrew Jones <drjones@redhat.com>
Message-id: 1529072910-16156-4-git-send-email-eric.auger@redhat.com
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Add experimental x-nbd-server-add-bitmap to expose a disabled
bitmap over NBD, in preparation for a pull model incremental
backup scheme. Also fix a corner case protocol issue with
NBD_CMD_BLOCK_STATUS, and add new NBD_CMD_CACHE.
- Eric Blake: tests: Simplify .gitignore
- Eric Blake: nbd/server: Reject 0-length block status request
- Vladimir Sementsov-Ogievskiy: 0/6 NBD export bitmaps
- Vladimir Sementsov-Ogievskiy: nbd/server: introduce NBD_CMD_CACHE
-----BEGIN PGP SIGNATURE-----
Comment: Public key at http://people.redhat.com/eblake/eblake.gpg
iQEcBAABCAAGBQJbK7wDAAoJEKeha0olJ0NqJuMH/0o7wzP3HWIO1pJZvhRA81AP
ZZqZy/8xO0B++keCxzgTbdr9yf6RDtYGjDrNuOcVzqfA1k9E1Cl2pPjS93TzPlzZ
PEgtG6bLXL2jCUi/8/EbZ/TLuoQx5YJc7eXTNREgXDRMLiTEg5Kcmg1n0efhr8Sl
5/bRMDmW1+atO4dJS8bEW5WLuy4IoB823cSHjWoPFRHeNCwmFREwTx2xVULVje84
JZzHYxyaBRquYLKIAew1WidhKVC4OiYtm4F5PxGBB/ZCVkeoDBX5uqZSjOUVt8y7
uOhGYQGW/qSj/s3x74NEL6IPo57Iay/Wq065DbolDx7s0y5t38TocrBnakzOBFU=
=VK6J
-----END PGP SIGNATURE-----
Merge remote-tracking branch 'remotes/ericb/tags/pull-nbd-2018-06-20-v2' into staging
nbd patches for 2018-06-20
Add experimental x-nbd-server-add-bitmap to expose a disabled
bitmap over NBD, in preparation for a pull model incremental
backup scheme. Also fix a corner case protocol issue with
NBD_CMD_BLOCK_STATUS, and add new NBD_CMD_CACHE.
- Eric Blake: tests: Simplify .gitignore
- Eric Blake: nbd/server: Reject 0-length block status request
- Vladimir Sementsov-Ogievskiy: 0/6 NBD export bitmaps
- Vladimir Sementsov-Ogievskiy: nbd/server: introduce NBD_CMD_CACHE
# gpg: Signature made Thu 21 Jun 2018 15:53:55 BST
# gpg: using RSA key A7A16B4A2527436A
# gpg: Good signature from "Eric Blake <eblake@redhat.com>"
# gpg: aka "Eric Blake (Free Software Programmer) <ebb9@byu.net>"
# gpg: aka "[jpeg image of size 6874]"
# Primary key fingerprint: 71C2 CC22 B1C4 6029 27D2 F3AA A7A1 6B4A 2527 436A
* remotes/ericb/tags/pull-nbd-2018-06-20-v2:
nbd/server: introduce NBD_CMD_CACHE
docs/interop: add nbd.txt
qapi: new qmp command nbd-server-add-bitmap
nbd/server: implement dirty bitmap export
nbd/server: add nbd_meta_empty_or_pattern helper
nbd/server: refactor NBDExportMetaContexts
nbd/server: fix trace
nbd/server: Reject 0-length block status request
tests: Simplify .gitignore
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The way we used to handle KVM allowable guest pagesizes for PAPR guests
required some convoluted checking of memory attached to the guest.
The allowable pagesizes advertised to the guest cpus depended on the memory
which was attached at boot, but then we needed to ensure that any memory
later hotplugged didn't change which pagesizes were allowed.
Now that we have an explicit machine option to control the allowable
maximum pagesize we can simplify this. We just check all memory backends
against that declared pagesize. We check base and cold-plugged memory at
reset time, and hotplugged memory at pre_plug() time.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
The way the POWER Hash Page Table (HPT) MMU is virtualized by KVM HV means
that every page that the guest puts in the pagetables must be truly
physically contiguous, not just GPA-contiguous. In effect this means that
an HPT guest can't use any pagesizes greater than the host page size used
to back its memory.
At present we handle this by changing what we advertise to the guest based
on the backing pagesizes. This is pretty bad, because it means the guest
sees a different environment depending on what should be host configuration
details.
As a start on fixing this, we add a new capability parameter to the
pseries machine type which gives the maximum allowed pagesizes for an
HPT guest. For now we just create and validate the parameter without
making it do anything.
For backwards compatibility, on older machine types we set it to the max
available page size for the host. For the 3.0 machine type, we fix it to
16, the intention being to only allow HPT pagesizes up to 64kiB by default
in future.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Handle nbd CACHE command. Just do read, without sending read data back.
Cache mechanism should be done by exported node driver chain.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20180413143156.11409-1-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
[eblake: fix two missing case labels in switch statements]
Signed-off-by: Eric Blake <eblake@redhat.com>
Handle a new NBD meta namespace: "qemu", and corresponding queries:
"qemu:dirty-bitmap:<export bitmap name>".
With the new metadata context negotiated, BLOCK_STATUS query will reply
with dirty-bitmap data, converted to extents. The new public function
nbd_export_bitmap selects which bitmap to export. For now, only one bitmap
may be exported.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20180609151758.17343-5-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
[eblake: wording tweaks, minor cleanups, additional tracing]
Signed-off-by: Eric Blake <eblake@redhat.com>
As well as being able to generate its own i2c transactions, the ppc4xx
i2c controller has a DIRECTCNTL register which allows explicit control
of the i2c lines.
Using this register an OS can directly bitbang i2c operations. In
order to let emulated i2c devices respond to this, we need to wire up
the DIRECTCNTL register to qemu's bitbanged i2c handling code.
Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
We don't emulate slave mode so related registers are not needed.
[lh]sadr are only retained to avoid too many warnings and simplify
debugging but sdata is not even correct because device has a 4 byte
FIFO instead so just remove this unimplemented register for now.
The intr register is also not implemented correctly, it is for
diagnostics and normally not even visible on device without explicitly
enabling it. As no guests are known to need this remove it as well.
Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
spapr_irq_alloc_block and spapr_irq_alloc() are now deprecated.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Today, when a device requests for IRQ number in a sPAPR machine, the
spapr_irq_alloc() routine first scans the ICSState status array to
find an empty slot and then performs the assignement of the selected
numbers. Split this sequence in two distinct routines : spapr_irq_find()
for lookups and spapr_irq_claim() for claiming the IRQ numbers.
This will ease the introduction of a static layout of IRQ numbers.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
spapr capabilities have an apply hook to actually activate (or deactivate)
the feature in the system at reset time. However, a number of capabilities
affect the setup of cpus, and need to be applied to each of them -
including hotplugged cpus for extra complication. To make this simpler,
add an optional cpu_apply hook that is called from spapr_cpu_reset().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Previously, the effective values of the various spapr capability flags
were only determined at machine reset time. That was a lazy way of making
sure it was after cpu initialization so it could use the cpu object to
inform the defaults.
But we've now improved the compat checking code so that we don't need to
instantiate the cpus to use it. That lets us move the resolution of the
capability defaults much earlier.
This is going to be necessary for some future capabilities.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
It introduces a base PnvChip class from which the specific processor
chip classes, Pnv8Chip and Pnv9Chip, inherit. Each of them needs to
define an init and a realize routine which will create the controllers
of the target processor. For the moment, the base PnvChip class
handles the XSCOM bus and the cores.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
A per-CPU machine data pointer was recently added to PowerPCCPU. The
motivation is to to hide platform specific details from the core CPU
code. This per-CPU data can hold state which is relevant to the guest
though, eg, Virtual Processor Areas, and we should migrate this state.
This patch adds the plumbing so that we can migrate the per-CPU data
for PAPR guests. We only do this for newer machine types for the sake
of backward compatibility. No state is migrated for the moment: the
vmstate_spapr_cpu_state structure will be populated by subsequent
patches.
Signed-off-by: Greg Kurz <groug@kaod.org>
[dwg: Fix some trivial spelling and spacing errors]
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This moves the details of the ISA bus creation under the LPC model but
more important, the new PnvChip operation will let us choose the chip
class to use when we introduce the different chip classes for Power9
and Power8. It hides away the processor chip controllers from the
machine.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
On Power9, the thread interrupt presenter has a different type and is
linked to the chip owning the cores.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This patch allows the user to specify whether to use active or only
background mode for mirror block jobs. Currently, this setting will
remain constant for the duration of the entire block job.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Message-id: 20180613181823.13618-14-mreitz@redhat.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Message-id: 20180613181823.13618-12-mreitz@redhat.com
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
This new function allows to look for a consecutively dirty area in a
dirty bitmap.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
Message-id: 20180613181823.13618-10-mreitz@redhat.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This new parameter allows the caller to just query the next dirty
position without moving the iterator.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
Message-id: 20180613181823.13618-8-mreitz@redhat.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
bdrv_drain_all_*() used bdrv_next() to iterate over all root nodes and
did a subtree drain for each of them. This works fine as long as the
graph is static, but sadly, reality looks different.
If the graph changes so that root nodes are added or removed, we would
have to compensate for this. bdrv_next() returns each root node only
once even if it's the root node for multiple BlockBackends or for a
monitor-owned block driver tree, which would only complicate things.
The much easier and more obviously correct way is to fundamentally
change the way the functions work: Iterate over all BlockDriverStates,
no matter who owns them, and drain them individually. Compensation is
only necessary when a new BDS is created inside a drain_all section.
Removal of a BDS doesn't require any action because it's gone afterwards
anyway.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
In the future, bdrv_drained_all_begin/end() will drain all invidiual
nodes separately rather than whole subtrees. This means that we don't
want to propagate the drain to all parents any more: If the parent is a
BDS, it will already be drained separately. Recursing to all parents is
unnecessary work and would make it an O(n²) operation.
Prepare the drain function for the changed drain_all by adding an
ignore_bds_parents parameter to the internal implementation that
prevents the propagation of the drain to BDS parents. We still (have to)
propagate it to non-BDS parents like BlockBackends or Jobs because those
are not drained separately.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
bdrv_drain_all() wants to have a single polling loop for draining the
in-flight requests of all nodes. This means that the AIO_WAIT_WHILE()
condition relies on activity in multiple AioContexts, which is polled
from the mainloop context. We must therefore call AIO_WAIT_WHILE() from
the mainloop thread and use the AioWait notification mechanism.
Just randomly picking the AioContext of any non-mainloop thread would
work, but instead of bothering to find such a context in the caller, we
can just as well accept NULL for ctx.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
bdrv_do_drained_begin() is only safe if we have a single
BDRV_POLL_WHILE() after quiescing all affected nodes. We cannot allow
that parent callbacks introduce a nested polling loop that could cause
graph changes while we're traversing the graph.
Split off bdrv_do_drained_begin_quiesce(), which only quiesces a single
node without waiting for its requests to complete. These requests will
be waited for in the BDRV_POLL_WHILE() call down the call chain.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Anything can happen inside BDRV_POLL_WHILE(), including graph
changes that may interfere with its callers (e.g. child list iteration
in recursive callers of bdrv_do_drained_begin).
Switch to a single BDRV_POLL_WHILE() call for the whole subtree at the
end of bdrv_do_drained_begin() to avoid such effects. The recursion
happens now inside the loop condition. As the graph can only change
between bdrv_drain_poll() calls, but not inside of it, doing the
recursion here is safe.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
We already requested that block jobs be paused in .bdrv_drained_begin,
but no guarantee was made that the job was actually inactive at the
point where bdrv_drained_begin() returned.
This introduces a new callback BdrvChildRole.bdrv_drained_poll() and
uses it to make bdrv_drain_poll() consider block jobs using the node to
be drained.
For the test case to work as expected, we have to switch from
block_job_sleep_ns() to qemu_co_sleep_ns() so that the test job is even
considered active and must be waited for when draining the node.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Commit 91af091f923 added an additional aio_poll() to BDRV_POLL_WHILE()
in order to make sure that all pending BHs are executed on drain. This
was the wrong place to make the fix, as it is useless overhead for all
other users of the macro and unnecessarily complicates the mechanism.
This patch effectively reverts said commit (the context has changed a
bit and the code has moved to AIO_WAIT_WHILE()) and instead polls in the
loop condition for drain.
The effect is probably hard to measure in any real-world use case
because actual I/O will dominate, but if I run only the initialisation
part of 'qemu-img convert' where it calls bdrv_block_status() for the
whole image to find out how much data there is copy, this phase actually
needs only roughly half the time after this patch.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The boot framebuffer is expected to be configured by the firmware, so it
uses fw_cfg as interface. Initialization goes as follows:
(1) Check whenever etc/ramfb is present.
(2) Allocate framebuffer from RAM.
(3) Fill struct RAMFBCfg, write it to etc/ramfb.
Done. You can write stuff to the framebuffer now, and it should appear
automagically on the screen.
Note that this isn't very efficient because it does a full display
update on each refresh. No dirty tracking. Dirty tracking would have
to be active for the whole ram slot, so that wouldn't be very efficient
either. For a boot display which is active for a short time only this
isn't a big deal. As permanent guest display something better should be
used (if possible).
This is the ramfb core code. Some windup is needed for display devices
which want have a ramfb boot display.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
Message-id: 20180613122948.18149-2-kraxel@redhat.com
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
CPUPPCState currently contains a number of fields containing the state of
the VPA. The VPA is a PAPR specific concept covering several guest/host
shared memory areas used to communicate some information with the
hypervisor.
As a PAPR concept this is really machine specific information, although it
is per-cpu, so it doesn't really belong in the core CPU state structure.
There's also other information that's per-cpu, but platform/machine
specific. So create a (void *)machine_data in PowerPCCPU which can be
used by the machine to locate per-cpu data. Intialization, lifetime and
cleanup of machine_data is entirely up to the machine type.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Greg Kurz <groug@kaod.org>
Currently, we allocate space for all the cpu objects within a single core
in one big block. This was copied from an older version of the spapr code
and requires some ugly pointer manipulation to extract the individual
objects.
This design was due to a misunderstanding of qemu lifetime conventions and
has already been changed in spapr (in 94ad93bd "spapr_cpu_core: instantiate
CPUs separately".
Make an equivalent change in pnv_core to get rid of the nasty pointer
arithmetic.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
In the case where we have an interrupt generated externally from inputs to
bits 1 and 2 of port A and/or port B, it is necessary to expose
mos6522_update_irq() so it can be called by the interrupt source.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The PMU device supercedes the CUDA device found on older New World Macs and
is supported by a larger number of guest OSs from OS 9 to OS X 10.5.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
MacOS 9 has a bug in its PMU driver whereby after configuring the ADB bus
devices it sends another write to reg 3 on both devices resetting them
both back to the same address.
Add a new disable_direct_reg3_writes property to ADBDevice to disable these
direct writes which can enabled just for the upcoming pmu-adb support.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
PMU-enabled New World Macs expose their GPIOs via a separate memory region
within the macio device.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This option allows the VIA configuration to be controlled between 3
different possible setups: cuda, pmu-adb and pmu with USB rather than ADB
keyboard/mouse.
For the moment we don't do anything with the configuration except to pass
it to the macio device (the via-cuda parent) and also to the firmware via
the fw_cfg interface so that it can present the correct device tree.
The default is cuda which is the current default and so will have no
change in behaviour.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Use mmap_lock in user-mode to protect TCG state and the page descriptors.
In !user-mode, each vCPU has its own TCG state, so no locks needed.
Per-page locks are used to protect the page descriptors.
Per-TB locks are used in both modes to protect TB jumps.
Some notes:
- tb_lock is removed from notdirty_mem_write by passing a
locked page_collection to tb_invalidate_phys_page_fast.
- tcg_tb_lookup/remove/insert/etc have their own internal lock(s),
so there is no need to further serialize access to them.
- do_tb_flush is run in a safe async context, meaning no other
vCPU threads are running. Therefore acquiring mmap_lock there
is just to please tools such as thread sanitizer.
- Not visible in the diff, but tb_invalidate_phys_page already
has an assert_memory_lock.
- cpu_io_recompile is !user-only, so no mmap_lock there.
- Added mmap_unlock()'s before all siglongjmp's that could
be called in user-mode while mmap_lock is held.
+ Added an assert for !have_mmap_lock() after returning from
the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic.
Performance numbers before/after:
Host: AMD Opteron(tm) Processor 6376
ubuntu 17.04 ppc64 bootup+shutdown time
700 +-+--+----+------+------------+-----------+------------*--+-+
| + + + + + *B |
| before ***B*** ** * |
|tb lock removal ###D### *** |
600 +-+ *** +-+
| ** # |
| *B* #D |
| *** * ## |
500 +-+ *** ### +-+
| * *** ### |
| *B* # ## |
| ** * #D# |
400 +-+ ** ## +-+
| ** ### |
| ** ## |
| ** # ## |
300 +-+ * B* #D# +-+
| B *** ### |
| * ** #### |
| * *** ### |
200 +-+ B *B #D# +-+
| #B* * ## # |
| #* ## |
| + D##D# + + + + |
100 +-+--+----+------+------------+-----------+------------+--+-+
1 8 16 Guest CPUs 48 64
png: https://imgur.com/HwmBHXe
debian jessie aarch64 bootup+shutdown time
90 +-+--+-----+-----+------------+------------+------------+--+-+
| + + + + + + |
| before ***B*** B |
80 +tb lock removal ###D### **D +-+
| **### |
| **## |
70 +-+ ** # +-+
| ** ## |
| ** # |
60 +-+ *B ## +-+
| ** ## |
| *** #D |
50 +-+ *** ## +-+
| * ** ### |
| **B* ### |
40 +-+ **** # ## +-+
| **** #D# |
| ***B** ### |
30 +-+ B***B** #### +-+
| B * * # ### |
| B ###D# |
20 +-+ D ##D## +-+
| D# |
| + + + + + + |
10 +-+--+-----+-----+------------+------------+------------+--+-+
1 8 16 Guest CPUs 48 64
png: https://imgur.com/iGpGFtv
The gains are high for 4-8 CPUs. Beyond that point, however, unrelated
lock contention significantly hurts scalability.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This applies to both user-mode and !user-mode emulation.
Instead of relying on a global lock, protect the list of incoming
jumps with tb->jmp_lock. This lock also protects tb->cflags,
so update all tb->cflags readers outside tb->jmp_lock to use
atomic reads via tb_cflags().
In order to find the destination TB (and therefore its jmp_lock)
from the origin TB, we introduce tb->jmp_dest[].
I considered not using a linked list of jumps, which simplifies
code and makes the struct smaller. However, it unnecessarily increases
memory usage, which results in a performance decrease. See for
instance these numbers booting+shutting down debian-arm:
Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%)
------------------------------------------------------------------------------
before 20.88 0.74 0.154512 0.
after 20.81 0.38 0.079078 -0.33524904
GTree 21.02 0.28 0.058856 0.67049808
GHashTable + xxhash 21.63 1.08 0.233604 3.5919540
Using a hash table or a binary tree to keep track of the jumps
doesn't really pay off, not only due to the increased memory usage,
but also because most TBs have only 0 or 1 jumps to them. The maximum
number of jumps when booting debian-arm that I measured is 35, but
as we can see in the histogram below a TB with that many incoming jumps
is extremely rare; the average TB has 0.80 incoming jumps.
n_jumps: 379208; avg jumps/tb: 0.801099
dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0]
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The appended adds assertions to make sure we do not longjmp with page
locks held. Note that user-mode has nothing to check, since page_locks
are !user-mode only.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Groundwork for supporting parallel TCG generation.
Instead of using a global lock (tb_lock) to protect changes
to pages, use fine-grained, per-page locks in !user-mode.
User-mode stays with mmap_lock.
Sometimes changes need to happen atomically on more than one
page (e.g. when a TB that spans across two pages is
added/invalidated, or when a range of pages is invalidated).
We therefore introduce struct page_collection, which helps
us keep track of a set of pages that have been locked in
the appropriate locking order (i.e. by ascending page index).
This commit first introduces the structs and the function helpers,
to then convert the calling code to use per-page locking. Note
that tb_lock is not removed yet.
While at it, rename tb_alloc_page to tb_page_add, which pairs with
tb_page_remove.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This commit does several things, but to avoid churn I merged them all
into the same commit. To wit:
- Use uintptr_t instead of TranslationBlock * for the list of TBs in a page.
Just like we did in (c37e6d7e "tcg: Use uintptr_t type for
jmp_list_{next|first} fields of TB"), the rationale is the same: these
are tagged pointers, not pointers. So use a more appropriate type.
- Only check the least significant bit of the tagged pointers. Masking
with 3/~3 is unnecessary and confusing.
- Introduce the TB_FOR_EACH_TAGGED macro, and use it to define
PAGE_FOR_EACH_TB, which improves readability. Note that
TB_FOR_EACH_TAGGED will gain another user in a subsequent patch.
- Update tb_page_remove to use PAGE_FOR_EACH_TB. In case there
is a bug and we attempt to remove a TB that is not in the list, instead
of segfaulting (since the list is NULL-terminated) we will reach
g_assert_not_reached().
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Thereby making it per-TCGContext. Once we remove tb_lock, this will
avoid an atomic increment every time a TB is invalidated.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This paves the way for enabling scalable parallel generation of TCG code.
Instead of tracking TBs with a single binary search tree (BST), use a
BST for each TCG region, protecting it with a lock. This is as scalable
as it gets, since each TCG thread operates on a separate region.
The core of this change is the introduction of struct tcg_region_tree,
which contains a pointer to a GTree and an associated lock to serialize
accesses to it. We then allocate an array of tcg_region_tree's, adding
the appropriate padding to avoid false sharing based on
qemu_dcache_linesize.
Given a tc_ptr, we first find the corresponding region_tree. This
is done by special-casing the first and last regions first, since they
might be of size != region.size; otherwise we just divide the offset
by region.stride. I was worried about this division (several dozen
cycles of latency), but profiling shows that this is not a fast path.
Note that region.stride is not required to be a power of two; it
is only required to be a multiple of the host's page size.
Note that with this design we can also provide consistent snapshots
about all region trees at once; for instance, tcg_tb_foreach
acquires/releases all region_tree locks before/after iterating over them.
For this reason we now drop tb_lock in dump_exec_info().
As an alternative I considered implementing a concurrent BST, but this
can be tricky to get right, offers no consistent snapshots of the BST,
and performance and scalability-wise I don't think it could ever beat
having separate GTrees, given that our workload is insert-mostly (all
concurrent BST designs I've seen focus, understandably, on making
lookups fast, which comes at the expense of convoluted, non-wait-free
insertions/removals).
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The meaning of "existing" is now changed to "matches in hash and
ht->cmp result". This is saner than just checking the pointer value.
Suggested-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
