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>
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>
Currently we don't support board configurations that put an IOMMU
in the path of the CPU's memory transactions, and instead just
assert() if the memory region fonud in address_space_translate_for_iotlb()
is an IOMMUMemoryRegion.
Remove this limitation by having the function handle IOMMUs.
This is mostly straightforward, but we must make sure we have
a notifier registered for every IOMMU that a transaction has
passed through, so that we can flush the TLB appropriately
when any of the IOMMUs change their mappings.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-id: 20180604152941.20374-5-peter.maydell@linaro.org
Add support for multiple IOMMU indexes to the IOMMU notifier APIs.
When initializing a notifier with iommu_notifier_init(), the caller
must pass the IOMMU index that it is interested in. When a change
happens, the IOMMU implementation must pass
memory_region_notify_iommu() the IOMMU index that has changed and
that notifiers must be called for.
IOMMUs which support only a single index don't need to change.
Callers which only really support working with IOMMUs with a single
index can use the result of passing MEMTXATTRS_UNSPECIFIED to
memory_region_iommu_attrs_to_index().
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-id: 20180604152941.20374-3-peter.maydell@linaro.org
There's a common pattern in QEMU where a function needs to perform
a data load or store of an N byte integer in a particular endianness.
At the moment this is handled by doing a switch() on the size and
calling the appropriate ld*_p or st*_p function for each size.
Provide a new family of functions ldn_*_p() and stn_*_p() which
take the size as an argument and do the switch() themselves.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20180611171007.4165-2-peter.maydell@linaro.org
The API for cpu_transaction_failed() says that it takes the physical
address for the failed transaction. However we were actually passing
it the offset within the target MemoryRegion. We don't currently
have any target CPU implementations of this hook that require the
physical address; fix this bug so we don't get confused if we ever
do add one.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20180611125633.32755-3-peter.maydell@linaro.org
The migration code should be using the
RAMBLOCK_FOREACH_MIGRATABLE and qemu_ram_foreach_block_migratable
not the all-block versions; poison them so that we can't accidentally
use them.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Message-Id: <20180605162545.80778-3-dgilbert@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
On the POWER9 processor, the XIVE interrupt controller can control
interrupt sources using MMIO to trigger events, to EOI or to turn off
the sources. Priority management and interrupt acknowledgment is also
controlled by MMIO in the presenter sub-engine.
These MMIO regions are exposed to guests in QEMU with a set of 'ram
device' memory mappings, similarly to VFIO, and the VMAs are populated
dynamically with the appropriate pages using a fault handler.
But, these regions are an issue for migration. We need to discard the
associated RAMBlocks from the RAM state on the source VM and let the
destination VM rebuild the memory mappings on the new host in the
post_load() operation just before resuming the system.
To achieve this goal, the following introduces a new RAMBlock flag
RAM_MIGRATABLE which is updated in the vmstate_register_ram() and
vmstate_unregister_ram() routines. This flag is then used by the
migration to identify RAMBlocks to discard on the source. Some checks
are also performed on the destination to make sure nothing invalid was
sent.
This change impacts the boston, malta and jazz mips boards for which
migration compatibility is broken.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
Do the cast to uintptr_t within the helper, so that the compiler
can type check the pointer argument. We can also do some more
sanity checking of the index argument.
Reviewed-by: Laurent Vivier <laurent@vivier.eu>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The function has been deprecated for 2.5 years, and there are just a handful
of users. Convert them to memory_region_init_io with NULL callbacks,
and while at it pass the right device as the owner.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Peter Maydell
Emilio G. CotaandRichard Henderson
Dr. David Alan Gilbert
Cédric Le GoaterandJuan Quintela
Paolo Bonzini