Allocating an arbitrarily-sized array of tbs results in either
(a) a lot of memory wasted or (b) unnecessary flushes of the code
cache when we run out of TB structs in the array.
An obvious solution would be to just malloc a TB struct when needed,
and keep the TB array as an array of pointers (recall that tb_find_pc()
needs the TB array to run in O(log n)).
Perhaps a better solution, which is implemented in this patch, is to
allocate TB's right before the translated code they describe. This
results in some memory waste due to padding to have code and TBs in
separate cache lines--for instance, I measured 4.7% of padding in the
used portion of code_gen_buffer when booting aarch64 Linux on a
host with 64-byte cache lines. However, it can allow for optimizations
in some host architectures, since TCG backends could safely assume that
the TB and the corresponding translated code are very close to each
other in memory. See this message by rth for a detailed explanation:
https://lists.gnu.org/archive/html/qemu-devel/2017-03/msg05172.html
Subject: Re: GSoC 2017 Proposal: TCG performance enhancements
Backports commit 6e3b2bfd6af488a896f7936e99ef160f8f37e6f2 from qemu
Instead of exporting goto_ptr directly to TCG frontends, export
tcg_gen_lookup_and_goto_ptr(), which calls goto_ptr with the pointer
returned by the lookup_tb_ptr() helper. This is the only use case
we have for goto_ptr and lookup_tb_ptr, so having this function is
very convenient. Furthermore, it trivially allows us to avoid calling
the lookup helper if goto_ptr is not implemented by the backend.
Backports commit cedbcb01529cb6cf9a2289cdbebbc63f6149fc18 from qemu
This patch converts the old "is_write" bool into IOMMUAccessFlags. The
difference is that "is_write" can only express either read/write, but
sometimes what we really want is "none" here (neither read nor write).
Replay is an good example - during replay, we should not check any RW
permission bits since thats not an actual IO at all.
Backports commit bf55b7afce53718ef96f4e6616da62c0ccac37dd from qemu
MemoryRegionCache did not know about virtio support for IOMMUs (because the
two features were developed at the same time). Revert MemoryRegionCache
to "normal" address_space_* operations for 2.9, as it is simpler than
undoing the virtio patches.
Backports commit 90c4fe5fc517a045e7a7cf2f23472e114042ca29 from qemu
The 'name' parameter to memory_region_init_* had been marked as debug
only, however vmstate_region_ram uses it as a parameter to
qemu_ram_set_idstr to set RAMBlock names and these form part of the
migration stream.
Backports commit e8f5fe2de125a0bfbefbaa6a69af81f4817cb7a0 from qemu
At the moment ram device's memory regions are DEVICE_NATIVE_ENDIAN. It's
incorrect. This memory region is backed by a MMIO area in host, so the
uint64_t data that MemoryRegionOps read from/write to this area should be
host-endian rather than target-endian. Hence, current code does not work
when target and host endianness are different which is the most common case
on PPC64. To fix it, this introduces DEVICE_HOST_ENDIAN for the ram device.
This has been tested on PPC64 BE/LE host/guest in all possible combinations
including TCG.
Backports commit c99a29e702528698c0ce2590f06ca7ff239f7c39 from qemu
While the vargs approach was flexible the original MTTCG ended up
having munge the bits to a bitmap so the data could be used in
deferred work helpers. Instead of hiding that in cputlb we push the
change to the API to make it take a bitmap of MMU indexes instead.
For ARM some the resulting flushes end up being quite long so to aid
readability I've tended to move the index shifting to a new line so
all the bits being or-ed together line up nicely, for example:
tlb_flush_page_by_mmuidx(other_cs, pageaddr,
(1 << ARMMMUIdx_S1SE1) |
(1 << ARMMMUIdx_S1SE0));
Backports commit 0336cbf8532935d8e23c2aabf3e2ce2c0697b6ac from qemu
We have never has the concept of global TLB entries which would avoid
the flush so we never actually use this flag. Drop it and make clear
that tlb_flush is the sledge-hammer it has always been.
Backports commit d10eb08f5d8389c814b554d01aa2882ac58221bf from qemu
This patch introduces a helper to query the iotlb entry for a
possible iova. This will be used by later device IOTLB API to enable
the capability for a dataplane (e.g vhost) to query the IOTLB.
Backports commit 052c8fa9983f553fdfa0d61034774070dd639c2b from qemu
Device models often have to perform multiple access to a single
memory region that is known in advance, but would to use "DMA-style"
functions instead of address_space_map/unmap. This can happen
for example when the data has to undergo endianness conversion.
Introduce a new data structure to cache the result of
address_space_translate without forcing usage of a host address
like address_space_map does.
Backports commit 1f4e496e1fc2eb6c8bf377a0f9695930c380bfd3 from qemu
Templatize the address_space_* and *_phys functions, so that we can add
similar functions in the next patch that work with a lightweight,
cache-like version of address_space_map/unmap.
Backports commit 0ce265ffef87f19f4dd1ff0663e09a63d66ae408 from qemu
When icount is active, tb_add_jump is surprisingly called with an
out of bounds basic block index. I have no idea how that can work,
but it does not seem like a good idea. Clear *last_tb for all
TB_EXIT_ICOUNT_EXPIRED cases, even when all you have to do is
refill icount_extra.
Backports commit d8dea6fbcbed177ca5d23ab77b3834a9437f0e88 from qemu
In the user emulation code path, tlb_vaddr_to_host erronesously passed
vaddr as the guest address to be translated, instead of addr, the parameter
which actually contained the guest address.
This resulted in incorrect addresses being used when emulating block copy
(mvc/mvpg) and block clear (xc) instructions for the s390x target.
Backports commit c2a85316902e67530da9d6548139fcce73c0cac6 from qemu
softmmu requires more functions to be thread-safe, because translation
blocks can be invalidated from e.g. notdirty callbacks. Probably the
same holds for user-mode emulation, it's just that no one has ever
tried to produce a coherent locking there.
This patch will guide the introduction of more tb_lock and tb_unlock
calls for system emulation.
Note that after this patch some (most) of the mentioned functions are
still called outside tb_lock/tb_unlock. The next one will rectify this.
Backports commit 7d7500d99895f888f97397ef32bb536bb0df3b74 from qemu
This adds asserts to check the locking on the various translation
engines structures. There are two sets of structures that are protected
by locks.
The first the l1map and PageDesc structures used to track which
translation blocks are associated with which physical addresses. In
user-mode this is covered by the mmap_lock.
The second case are TB context related structures which are protected by
tb_lock which is also user-mode only.
Currently the asserts do nothing in SoftMMU mode but this will change
for MTTCG.
Backports commit 301e40ed8005306c009978be295ed9a4b725178b from qemu
Probe for whether the specified guest write access is permitted.
If it is not permitted then an exception will be taken in the same
way as if this were a real write access (and we will not return).
Otherwise the function will return, and there will be a valid
entry in the TLB for this access.
Backports commit 3b4afc9e75ab1a95f33e41f462921093f8a109c4 from qemu
When we cannot emulate an atomic operation within a parallel
context, this exception allows us to stop the world and try
again in a serial context.
Backports commit fdbc2b5722f6092e47181a947c90fd4bdcc1c121 from qemu
Also backports parts of commit 02d57ea115b7669f588371c86484a2e8ebc369be
Support target CPUs having a page size which isn't knownn
at compile time. To use this, the CPU implementation should:
* define TARGET_PAGE_BITS_VARY
* not define TARGET_PAGE_BITS
* define TARGET_PAGE_BITS_MIN to the smallest value it
might possibly want for TARGET_PAGE_BITS
* call set_preferred_target_page_bits() in its realize
function to indicate the actual preferred target page
size for the CPU (and report any error from it)
In CONFIG_USER_ONLY, the CPU implementation should continue
to define TARGET_PAGE_BITS appropriately for the guest
OS page size.
Machines which want to take advantage of having the page
size something larger than TARGET_PAGE_BITS_MIN must
set the MachineClass minimum_page_bits field to a value
which they guarantee will be no greater than the preferred
page size for any CPU they create.
Note that changing the target page size by setting
minimum_page_bits is a migration compatibility break
for that machine.
For debugging purposes, attempts to use TARGET_PAGE_SIZE
before it has been finally confirmed will assert.
Backports commit 20bccb82ff3ea09bcb7c4ee226d3160cab15f7da from qemu
This speeds up MEMORY_LISTENER_CALL noticeably. Right now,
with many PCI devices you have N regions added to M AddressSpaces
(M = # PCI devices with bus-master enabled) and each call looks
up the whole listener list, with at least M listeners in it.
Because most of the regions in N are BARs, which are also roughly
proportional to M, the whole thing is O(M^3). This changes it
to O(M^2), which is the best we can do without rewriting the
whole thing.
Backports commit 9a54635dcb51a3fcf7507af630168f514a8cd4e7 from qemu
The return address argument to the softmmu template helpers was
confused. In the legacy case, we wanted to indicate that there
is no return address, and so passed in NULL. However, we then
immediately subtracted GETPC_ADJ from NULL, resulting in a non-zero
value, indicating the presence of an (invalid) return address.
Push the GETPC_ADJ subtraction down to the only point it's required:
immediately before use within cpu_restore_state_from_tb, after all
NULL pointer checks have been completed.
This makes GETPC and GETRA identical. Remove GETRA as the lesser
used macro, replacing all uses with GETPC.
Backports commit 01ecaf438b1eb46abe23392c8ce5b7628b0c8cf5 from qemu
When invalidating a translation block, set an invalid flag into the
TranslationBlock structure first. It is also necessary to check whether
the target TB is still valid after acquiring 'tb_lock' but before calling
tb_add_jump() since TB lookup is to be performed out of 'tb_lock' in
future. Note that we don't have to check 'last_tb'; an already invalidated
TB will not be executed anyway and it is thus safe to patch it.
Backports commit 6d21e4208f382dd8ca1f7995a6dd9ea7ca281163 from qemu
With a vfio assigned device we lay down a base MemoryRegion registered
as an IO region, giving us read & write accessors. If the region
supports mmap, we lay down a higher priority sub-region MemoryRegion
on top of the base layer initialized as a RAM device pointer to the
mmap. Finally, if we have any quirks for the device (ie. address
ranges that need additional virtualization support), we put another IO
sub-region on top of the mmap MemoryRegion. When this is flattened,
we now potentially have sub-page mmap MemoryRegions exposed which
cannot be directly mapped through KVM.
This is as expected, but a subtle detail of this is that we end up
with two different access mechanisms through QEMU. If we disable the
mmap MemoryRegion, we make use of the IO MemoryRegion and service
accesses using pread and pwrite to the vfio device file descriptor.
If the mmap MemoryRegion is enabled and results in one of these
sub-page gaps, QEMU handles the access as RAM, using memcpy to the
mmap. Using either pread/pwrite or the mmap directly should be
correct, but using memcpy causes us problems. I expect that not only
does memcpy not necessarily honor the original width and alignment in
performing a copy, but it potentially also uses processor instructions
not intended for MMIO spaces. It turns out that this has been a
problem for Realtek NIC assignment, which has such a quirk that
creates a sub-page mmap MemoryRegion access.
To resolve this, we disable memory_access_is_direct() for ram_device
regions since QEMU assumes that it can use memcpy for those regions.
Instead we access through MemoryRegionOps, which replaces the memcpy
with simple de-references of standard sizes to the host memory.
With this patch we attempt to provide unrestricted access to the RAM
device, allowing byte through qword access as well as unaligned
access. The assumption here is that accesses initiated by the VM are
driven by a device specific driver, which knows the device
capabilities. If unaligned accesses are not supported by the device,
we don't want them to work in a VM by performing multiple aligned
accesses to compose the unaligned access. A down-side of this
philosophy is that the xp command from the monitor attempts to use
the largest available access weidth, unaware of the underlying
device. Using memcpy had this same restriction, but at least now an
operator can dump individual registers, even if blocks of device
memory may result in access widths beyond the capabilities of a
given device (RTL NICs only support up to dword).
Backports commit 1b16ded6a512809f99c133a97f19026fe612b2de from qemu
Setting skip_dump on a MemoryRegion allows us to modify one specific
code path, but the restriction we're trying to address encompasses
more than that. If we have a RAM MemoryRegion backed by a physical
device, it not only restricts our ability to dump that region, but
also affects how we should manipulate it. Here we recognize that
MemoryRegions do not change to sometimes allow dumps and other times
not, so we replace setting the skip_dump flag with a new initializer
so that we know exactly the type of region to which we're applying
this behavior.
Backports commit ca83f87a66d19fdaabf23d4f5ebb49396fe232c1 from qemu
Instead of using -1 as end of chain, use 0, and link through the 0
entry as a fully circular double-linked list.
Backports commit dcb8e75870e2de199db853697f8839cb603beefe from qemu
Header guard symbols should match their file name to make guard
collisions less likely. Offenders found with
scripts/clean-header-guards.pl -vn.
Cleaned up with scripts/clean-header-guards.pl, followed by some
renaming of new guard symbols picked by the script to better ones.
Backports commit 121d07125bb6d7079c7ebafdd3efe8c3a01cc440 from qemu
Tracked down with an ugly, brittle and probably buggy Perl script.
Also move includes converted to <...> up so they get included before
ours where that's obviously okay.
Backports commit a9c94277f07d19d3eb14f199c3e93491aa3eae0e from qemu
There are functions tlb_fill(), cpu_unaligned_access() and
do_unaligned_access() that are called with access type and mmu index
arguments. But these arguments are named 'is_write' and 'is_user' in their
declarations. The patches fix the arguments to avoid a confusion.
Backports commit b35399bb4e9968296a12303b00f9f2066470e987 from qemu
Some architectures (e.g. ARMv8) need the address which is aligned
to a size more than the size of the memory access.
To support such check it's enough the current costless alignment
check implementation in QEMU, but we need to support
an alignment size specifying.
Backports commit 1f00b27f17518a1bcb4cedca49eaec96a4d560bd from qemu
It doesn't make sense to pass a NULL ops argument to
memory_region_init_rom_device(), because the effect will
be that if the guest tries to write to the memory region
then QEMU will segfault. Catch the bug earlier by sanity
checking the arguments to this function, and remove the
misleading documentation that suggests that passing NULL
might be sensible.
Backports commit 39e0b03dec518254fabd2acff29548d3f1d2b754 from qemu
Provide a new helper function memory_region_init_rom() for memory
regions which are read-only (and unlike those created by
memory_region_init_rom_device() don't have special behaviour
for writes). This has the same behaviour as calling
memory_region_init_ram() and then memory_region_set_readonly()
(which is what we do today in boards with pure ROMs) but is a
more easily discoverable API for the purpose.
Backports commit a1777f7f6462c66e1ee6e98f0d5c431bfe988aa5 from qemu
The IOMMU driver may change behavior depending on whether a notifier
client is present. In the case of POWER, this represents a change in
the visibility of the IOTLB, for other drivers such as intel-iommu and
future AMD-Vi emulation, notifier support is not yet enabled and this
provides the opportunity to flag that incompatibility.
Backports commit d22d8956b185c002b50a4d0883aff61f857347ef from qemu
Every IOMMU has some granularity which MemoryRegionIOMMUOps::translate
uses when translating, however this information is not available outside
the translate context for various checks.
This adds a get_min_page_size callback to MemoryRegionIOMMUOps and
a wrapper for it so IOMMU users (such as VFIO) can know the minimum
actual page size supported by an IOMMU.
As IOMMU MR represents a guest IOMMU, this uses TARGET_PAGE_SIZE
as fallback.
This removes vfio_container_granularity() and uses new helper in
memory_region_iommu_replay() when replaying IOMMU mappings on added
IOMMU memory region.
Backports the relevant parts of commit f682e9c244af7166225f4a50cc18ff296bb9d43e from qemu
For some workloads such as arm bootup, tb_phys_hash is performance-critical.
The is due to the high frequency of accesses to the hash table, originated
by (frequent) TLB flushes that wipe out the cpu-private tb_jmp_cache's.
More info:
https://lists.nongnu.org/archive/html/qemu-devel/2016-03/msg05098.html
To dig further into this I modified an arm image booting debian jessie to
immediately shut down after boot. Analysis revealed that quite a bit of time
is unnecessarily spent in tb_phys_hash: the cause is poor hashing that
results in very uneven loading of chains in the hash table's buckets;
the longest observed chain had ~550 elements.
The appended addresses this with two changes:
1) Use xxhash as the hash table's hash function. xxhash is a fast,
high-quality hashing function.
2) Feed the hashing function with not just tb_phys, but also pc and flags.
This improves performance over using just tb_phys for hashing, since that
resulted in some hash buckets having many TB's, while others getting very few;
with these changes, the longest observed chain on a single hash bucket is
brought down from ~550 to ~40.
Tests show that the other element checked for in tb_find_physical,
cs_base, is always a match when tb_phys+pc+flags are a match,
so hashing cs_base is wasteful. It could be that this is an ARM-only
thing, though. UPDATE:
On Tue, Apr 05, 2016 at 08:41:43 -0700, Richard Henderson wrote:
> The cs_base field is only used by i386 (in 16-bit modes), and sparc (for a TB
> consisting of only a delay slot).
> It may well still turn out to be reasonable to ignore cs_base for hashing.
BTW, after this change the hash table should not be called "tb_hash_phys"
anymore; this is addressed later in this series.
This change gives consistent bootup time improvements. I tested two
host machines:
- Intel Xeon E5-2690: 11.6% less time
- Intel i7-4790K: 19.2% less time
Increasing the number of hash buckets yields further improvements. However,
using a larger, fixed number of buckets can degrade performance for other
workloads that do not translate as many blocks (600K+ for debian-jessie arm
bootup). This is dealt with later in this series.
Backports commit 42bd32287f3a18d823f2258b813824a39ed7c6d9 from qemu
This will be used by upcoming changes for hashing the tb hash.
Add this into a separate file to include the copyright notice from
xxhash.
Backports commit dc8b295d05ec35a8c032f9abca421772347ba5d4 from qemu
The function cpu_resume_from_signal() is now always called with a
NULL puc argument, and is rather misnamed since it is never called
from a signal handler. It is essentially forcing an exit to the
top level cpu loop but without raising any exception, so rename
it to cpu_loop_exit_noexc() and drop the useless unused argument.
Backports commit 6886b98036a8f8f5bce8b10756ce080084cef11b from qemu
The WORDS_ALIGNED #define is not used anywhere, and hasn't been since
2013 when commit 612d590ebc6cef rewrote the various ld<type>_<endian>_p
functions to not use it. Remove the #define and the comment describing it.
Also remove the line in the comment about TARGET_WORDS_ALIGNED, since
it has never actually existed.
Backports commit 0d5c21f2b3bf1e0b562a2c74e353d2e03f2f50ef from qemu
Let users of qemu_get_ram_ptr and qemu_ram_ptr_length pass in an
address that is relative to the MemoryRegion. This basically means
what address_space_translate returns.
Because the semantics of the second parameter change, rename the
function to qemu_map_ram_ptr.
Backports commit 0878d0e11ba8013dd759c6921cbf05ba6a41bd71 from qemu
Move the old qemu_ram_addr_from_host to memory_region_from_host and
make it return an offset within the region. For qemu_ram_addr_from_host
return the ram_addr_t directly, similar to what it was before
commit 1b5ec23 ("memory: return MemoryRegion from qemu_ram_addr_from_host",
2013-07-04).
Backports commit 07bdaa4196b51bc7ffa7c3f74e9e4a9dc8a7966a from qemu
Of the two callers, one does not use it, and the other can compute
it itself based on the other output argument (offset) and the RAMBlock.
Backports commit f615f39616c4fd1a3a3b078af8d75bb4be6390de from qemu
Remove direct uses of ram_addr_t and optimize memory_region_{get,set}_fd
now that a MemoryRegion knows its RAMBlock directly.
Backports commit 4ff87573df3606856a92c14eef3393a63d736d11 from qemu
The collision check does nothing and hasn't been used. Remove the
variable together with related code.
Backports commit b61359781958759317ee6fd1a45b59be0b7dbbe1 from qemu
On the one hand, we have already qemu_get_ram_block() whose function
is similar. On the other hand, we can directly use mr->ram_block but
searching RAMblock by ram_addr which is a kind of waste.
Backports commit fa53a0e53efdc7002497ea4a76aacf6cceb170ef from qemu
