When using bit-wise operations that exploit the power-of-two
nature of the second argument of ROUND_UP(), we still need to
ensure that the mask is as wide as the first argument (done
by using a ternary to force proper arithmetic promotion).
Unpatched, ROUND_UP(2ULL*1024*1024*1024*1024, 512U) produces 0,
instead of the intended 2TiB, because negation of an unsigned
32-bit quantity followed by widening to 64-bits does not
sign-extend the mask.
Broken since its introduction in commit 292c8e50 (v1.5.0).
Callers that passed the same width type to both macro parameters,
or that had other code to ensure the first parameter's maximum
runtime value did not exceed the second parameter's width, are
unaffected, but I did not audit to see which (if any) existing
clients of the macro could trigger incorrect behavior (I found
the bug while adding a new use of the macro).
While preparing the patch, checkpatch complained about poor
spacing, so I also fixed that here and in the nearby DIV_ROUND_UP.
Backports commit 33a599667a9e70588483a31286dfff8cfc27d513 from qemu
According to the ARM ARM exclusive loads require the same alignment as
exclusive stores. Let's update the memops used for the load to match
that of the store. This adds the alignment requirement to the memops.
Backports commit 4a2fdb78e794c1ad93aa9e160235d6a61a2125de from qemu
We are not providing the required single-copy atomic semantics for
the 64-bit operation that is the 32-bit paired load.
At the same time, leave the entire 64-bit value in cpu_exclusive_val
and stop writing to cpu_exclusive_high. This means that we do not
have to re-assemble the 64-bit quantity when it comes time to store.
At the same time, drop a redundant temporary and perform all loads
directly into the cpu_exclusive_* globals.
Backports commit 19514cde3b92938df750acaecf2caaa85e1d36a6 from qemu
When we perform the atomic_cmpxchg operation we want to perform the
operation on a pair of 32-bit registers. Previously we were just passing
the register size in which was set to MO_32. This would result in the
high register to be ignored. To fix this issue we hardcode the size to
be 64-bits long when operating on 32-bit pairs.
Backports commit 955fd0ad5d610f62ba2f4ce46a872bf50434dcf8 from qemu
This reverts commit e2a7f28693aea7e194ec1435697ec4feb24f8a6f.
This was not supposed to go upstream yet. Reverting.
Backports commit cde0a63ad721dbb538419a00f9405587680be436 from qemu
When emulating various SSE4.1 instructions such as pinsrd, the address
of a memory operand is computed without allowing for the 8-bit
immediate operand located after the memory operand, meaning that the
memory operand uses the wrong address in the case where it is
rip-relative. This patch adds the required rip_offset setting for
those instructions, so fixing some GCC test failures (13 in the gcc
testsuite in my GCC 6-based testing) when testing with a default CPU
setting enabling those instructions.
Backports commit ab6ab3e9972a49a359f59895a88bed311472ca97 from qemu
For a 64-bit ILP32 host, aligning to sizeof(long) is not enough.
Guess the minimum for any host is 8, as that covers uint64_t.
Qemu doesn't use a host long double or host vectors, except in
extremely limited circumstances.
Fixes a bus error for a sparc v8plus host.
Backports commit 13aaef678ed377b12b76dc7fb9e615b2f2f9047b from qemu
Patch 85aa80813dd changed the IF emitting the TST instruction,
but failed to change the ?: converting CMP to CMPEQ, so the
result of the TST is ignored.
Backports commit ca671de8af96798e0f493378240034620a3a04ee from qemu
RDHWR CC reads the CPU timer like MFC0 CP0_Count, so with icount enabled
it must set can_do_io while it calls the helper to avoid the "Bad icount
read" error. It should also break out of the translation loop to ensure
that timer interrupts are immediately handled.
Backports commit d673a68db6963e86536b125af464bb6ed03eba33 from qemu
DMTC0 CP0_Cause does a redundant gen_io_start() and gen_io_end() pair,
even though this is done for all DMTC0 operations outside of the switch
statement. Remove these redundant calls.
Backports commit 51ca717b079dccae5b6cc9f45153f5044abd34f0 from qemu
Commit e350d8ca3ac7 ("target/mips: optimize indirect branches") made
indirect branches able to directly find the next TB and jump straight to
it without breaking out of translated code and going around the main
execution loop. This breaks the assumption in target/mips/translate.c
that BS_STOP is sufficient to cause pending interrupts to be handled,
since interrupts are only checked in the main loop.
Fix a few of these assumptions by using gen_save_pc to update the saved
PC and using BS_EXCP instead of BS_STOP:
- [D]MFC0 CP0_Count may trigger a timer interrupt which should be
immediately handled.
- [D]MTC0 CP0_Cause may trigger an interrupt (but in fact translation
was only even being stopped in the DMTC0 case).
- [D]MTC0 CP0_<any> when icount is used is assumed could potentially
cause interrupts.
- EI may trigger an interrupt which was pending. I specifically hit
this case when running KVM nested in mipsel-softmmu. A timer
interrupt while the 2nd guest was executing is caught by KVM which
switches back to the normal Linux exception base and re-enables
interrupts with EI. Since the above commit QEMU doesn't leave
translated code until the nested KVM has already restored the KVM
exception base and returned to the 2nd guest, at which point it is
too late to check for pending interrupts and it gets stuck in an
infinite loop of unhandled interrupts.
Something similar was needed for ARM in commit b29fd33db578
("target/arm: use DISAS_EXIT for eret handling").
Backports commit b74cddcbf6063f684725e3f8bca49a68e30cba71 from qemu
Improve the segment definitions used by get_physical_address() to yield
target_ulong types, e.g. 0xffffffff80000000 instead of 0x80000000. This
is in preparation for enabling emulation of MIPS KVM T&E segments in TCG
MIPS targets, which unlike KVM could potentially have 64-bit
target_ulong. In such a case the offset guest KSEG0 address ends up at
e.g. 0x000000008xxxxxxx instead of 0xffffffff8xxxxxxx.
This also allows the casts to int32_t that force sign extension to be
removed, which removes any confusion due to relational comparison of
unsigned (target_ulong) and signed (int32_t) types.
Backports commit 6743334568933199927af4992a04bfb3c30610f5 from qemu
Writing to the MIPS DESAVE register (and now the KScratch registers)
will stop translation, supposedly due to risk of execution mode
switches. However these registers are basically RW scratch registers
with no side effects so there is no risk of them triggering execution
mode changes.
Drop the bstate = BS_STOP for these registers for both mtc0 and dmtc0.
Backports commit cb539fd241900f51de7d21244f7a55422ad0d40a from qemu
Commit 04bf2526ce87f21b32c9acba1c5518708c243ad0 (exec: use
qemu_ram_ptr_length to access guest ram) start using qemu_ram_ptr_length
instead of qemu_map_ram_ptr, but when used with Xen, the behavior of
both function is different. They both call xen_map_cache, but one with
"lock", meaning the mapping of guest memory is never released
implicitly, and the second one without, which means, mapping can be
release later, when needed.
In the context of address_space_{read,write}_continue, the ptr to those
mapping should not be locked because it is used immediatly and never
used again.
The lock parameter make it explicit in which context qemu_ram_ptr_length
is called.
Backports commit f5aa69bdc3418773f26747ca282c291519626ece from qemu
When the PMSAv7 implementation was originally added it was for R profile
CPUs only, and reset was handled using the cpreg .resetfn hooks.
Unfortunately for M profile cores this doesn't work, because they do
not register any cpregs. Move the reset handling into arm_cpu_reset(),
where it will work for both R profile and M profile cores.
Backports commit 69ceea64bf565559a2b865ffb2a097d2caab805b from qemu
Almost all of the PMSAv7 state is in the pmsav7 substruct of
the ARM CPU state structure. The exception is the region
number register, which is in cp15.c6_rgnr. This exception
is a bit odd for M profile, which otherwise generally does
not store state in the cp15 substruct.
Rename cp15.c6_rgnr to pmsav7.rnr accordingly.
Backports commit 8531eb4f614a60e6582d4832b15eee09f7d27874 from qemu
For an M profile v7PMSA, the system space (0xe0000000 - 0xffffffff) can
never be executable, even if the guest tries to set the MPU registers
up that way. Enforce this restriction.
Backports commit bf446a11dfb17ae7d8ed2b61a2444804eb458075 from qemu
The M profile PMSAv7 specification says that if the address being looked
up is in the PPB region (0xe0000000 - 0xe00fffff) then we do not use
the MPU regions but always use the default memory map. Implement this
(we were previously behaving like an R profile PMSAv7, which does not
special case this).
Backports commit 38aaa60ca464b48e6feef346709e97335d01b289 from qemu
Correct off-by-one bug in the PSMAv7 MPU tracing where it would print
a write access as "reading", an insn fetch as "writing", and a read
access as "execute".
Since we have an MMUAccessType enum now, we can make the code clearer
in the process by using that rather than the raw 0/1/2 values.
Backports commit 709e4407add7acacc593cb6cdac026558c9a8fb6 from qemu
Enable the CP0_EBase.WG (write gate) on the I6400 and MIPS64R2-generic
CPUs. This allows 64-bit guests to run KVM itself, which uses
CP0_EBase.WG to point CP0_EBase at XKPhys.
Backports commit bad63a8008a0aaefcd00542c89bee01623d7c9de from qemu
Add the Enhanced Virtual Addressing (EVA) feature to the P5600 core
configuration, along with the related Segmentation Control (SC) feature
and writable CP0_EBase.WG bit.
This allows it to run Malta EVA kernels.
Backports commit 574da58e4678b3c09048f268821295422d8cde6d from qemu
Implement the optional segmentation control feature in the virtual to
physical address translation code.
The fixed legacy segment and xkphys handling is replaced with a dynamic
layout based on the segmentation control registers (which should be set
up even when the feature is not exposed to the guest).
Backports commit 480e79aedd322fcfac17052caff21626ea7c78e2 from qemu
The optional segmentation control registers CP0_SegCtl0, CP0_SegCtl1 &
CP0_SegCtl2 control the behaviour and required privilege of the legacy
virtual memory segments.
Add them to the CP0 interface so they can be read and written when
CP0_Config3.SC=1, and initialise them to describe the standard legacy
layout so they can be used in future patches regardless of whether they
are exposed to the guest.
Backports commit cec56a733dd2c3fa81dbedbecf03922258747f7d from qemu
The segmentation control feature allows a legacy memory segment to
become unmapped uncached at error level (according to CP0_Status.ERL),
and in fact the user segment is already treated in this way by QEMU.
Add a new MMU mode for this state so that QEMU's mappings don't persist
between ERL=0 and ERL=1.
Backports commit 42c86612d507c2a8789f2b8d920a244693c4ef7b from qemu
The MIPS mmu_idx is sometimes calculated from hflags without an env
pointer available as cpu_mmu_index() requires.
Create a common hflags_mmu_index() for the purpose of this calculation
which can operate on any hflags, not just with an env pointer, and
update cpu_mmu_index() itself and gen_intermediate_code() to use it.
Also update debug_post_eret() and helper_mtc0_status() to log the MMU
mode with the status change (SM, UM, or nothing for kernel mode) based
on cpu_mmu_index() rather than directly testing hflags.
This will also allow the logic to be more easily updated when a new MMU
mode is added.
Backports commit b0fc6003224543d2bdb172eca752656a6223e4a1 from qemu
When performing virtual to physical address translation, check the
required privilege level based on the mem_idx rather than the mode in
the hflags. This will allow EVA loads & stores to operate safely only on
user memory from kernel mode.
For the cases where the mmu_idx doesn't need to be overridden
(mips_cpu_get_phys_page_debug() and cpu_mips_translate_address()), we
calculate the required mmu_idx using cpu_mmu_index(). Note that this
only tests the MIPS_HFLAG_KSU bits rather than MIPS_HFLAG_MODE, so we
don't test the debug mode hflag MIPS_HFLAG_DM any longer. This should be
fine as get_physical_address() only compares against MIPS_HFLAG_UM and
MIPS_HFLAG_SM, neither of which should get set by compute_hflags() when
MIPS_HFLAG_DM is set.
Backports commit 9fbf4a58c90183b30bb2c8ad971ccce7e6716a16 from qemu
Implement decoding of microMIPS EVA load and store instruction groups in
the POOL31C pool. These use the same gen_ld(), gen_st(), gen_st_cond()
helpers as the MIPS32 decoding, passing the equivalent MIPS32 opcodes as
opc.
Backports commit 8fffc64696783b1ff1d17262d098976479895660 from qemu
Add CP0.ErrCtl register with WST, SPR and ITC bits. In 34K and interAptiv
processors these bits are used to enable CACHE instruction access to
different arrays. When WST=0, SPR=0 and ITC=1 the CACHE instruction will
access ITC tag values.
Generally we do not model caches and we have been treating the CACHE
instruction as NOP. But since CACHE can operate on ITC Tags new
MIPS_HFLAG_ITC_CACHE hflag is introduced to generate the helper only when
CACHE is in the ITC Access mode.
Backports commit 0d74a222c27e26fc40f4f6120c61c3f9ceaa3776 from qemu
Implement decoding of MIPS32 EVA loads and stores. These access the user
address space from kernel mode when implemented, so for each instruction
we need to check that EVA is available from Config5.EVA & check for
sufficient COP0 privilege (with the new check_eva()), and then override
the mem_idx used for the operation.
Unfortunately some Loongson 2E instructions use overlapping encodings,
so we must be careful not to prevent those from being decoded when EVA
is absent.
Backports commit 7696414729b2d0f870c80ad1dd637d854bc78847 from qemu
EVA load and store instructions access the user mode address map, so
they need to use mem_idx of MIPS_HFLAG_UM. Update the various utility
functions to allow mem_idx to be more easily overridden from the
decoding logic.
Specifically we add a mem_idx argument to the op_ld/st_* helpers used
for atomics, and a mem_idx local variable to gen_ld(), gen_st(), and
gen_st_cond().
Backports commit dd4096cd2ccc19384770f336c930259da7a54980 from qemu
Add support for the CP0_EBase.WG bit, which allows upper bits to be
written (bits 31:30 on MIPS32, or bits 63:30 on MIPS64), along with the
CP0_Config5.CV bit to control whether the exception vector for Cache
Error exceptions is forced into KSeg1.
This is necessary on MIPS32 to support Segmentation Control and Enhanced
Virtual Addressing (EVA) extensions (where KSeg1 addresses may not
represent an unmapped uncached segment).
It is also useful on MIPS64 to allow the exception base to reside in
XKPhys, and possibly out of range of KSEG0 and KSEG1.
Backports commit 74dbf824a1313b6064bbebb981a7440951d70896 from qemu
There is no need to invalidate any shadow TLB entries when the ASID
changes or when access to one of the 64-bit segments has been disabled,
since doing so doesn't reveal to software whether any TLB entries have
been evicted into the shadow half of the TLB.
Therefore weaken the tlb flushes in these cases to only flush the QEMU
TLB.
Backports commit 9658e4c342e6ae0d775101f8f6bb6efb16789af1 from qemu
Writing specific TLB entries with TLBWI flushes shadow TLB entries
unless an existing entry is having its access permissions upgraded. This
is necessary as software would from then on expect the previous mapping
in that entry to no longer be in effect (even if QEMU has quietly
evicted it to the shadow TLB on a TLBWR).
However it won't do this if only EHINV, XI, or RI bits have been set,
even if that results in a reduction of permissions, so add the necessary
checks to invoke the flush when these bits are set.
Backports commit eff6ff9431aa9776062a5f4a08d1f6503ca9995a from qemu
Using MFC0 to read CP0_UserLocal uses tcg_gen_ld32s_tl, however
CP0_UserLocal is a target_ulong. On a big endian host with a MIPS64
target this reads and sign extends the more significant half of the
64-bit register.
Fix this by using ld_tl to load the whole target_ulong and ext32s_tl to
sign extend it, as done for various other target_ulong COP0 registers.
Backports commit e40df9a80bb7cdb0a4ca650985fa9fe572097fa7 from qemu
This include was forgotten when splitting cacheinfo.c out of
tcg/ppc/tcg-target.inc.c (see commit b255b2c8).
For a Centos7 host, the include path
<signal.h>
<bits/sigcontext.h>
<asm/sigcontext.h>
<asm/elf.h>
<asm/auxvec.h>
implicitly pulls in the desired AT_* defines.
Not so for Debian Jessie.
Backports commit 810d5cad4087236236e00fd3046a16adf26e9060 from qemu
Reserve a register for the guest_base using ppc code for reference.
By doing so, we do not have to recompute it for every memory load.
Backports commit 4df9cac57f5220c17d856292e90fce455f708421 from qemu
Introduce Skylake-Server cpu mode which inherits the features from
Skylake-Client and supports some additional features that are: AVX512,
CLWB and PGPE1GB.
Backports commit 53f9a6f45fb214540cb40af45efc11ac40ac454c from qemu
Currently when running KVM, we expose "KVMKVMKVM\0\0\0" in
the 0x40000000 CPUID leaf. Other hypervisors (VMWare,
HyperV, Xen, BHyve) all do the same thing, which leaves
TCG as the odd one out.
The CPUID signature is used by software to detect which
virtual environment they are running in and (potentially)
change behaviour in certain ways. For example, systemd
supports a ConditionVirtualization= setting in unit files.
The virt-what command can also report the virt type it is
running on
Currently both these apps have to resort to custom hacks
like looking for 'fw-cfg' entry in the /proc/device-tree
file to identify TCG.
This change thus proposes a signature "TCGTCGTCGTCG" to be
reported when running under TCG.
To hide this, the -cpu option tcg-cpuid=off can be used.
Backports commits 4ed3d478c63dc65a02eba774c35116618ea5ff10 and 1ce36bfe6424243082d3d7c2330e1a0a4ff72a43 from qemu
object_resolve_path*() ambiguous path detection breaks when
ambiguous==NULL and the object tree have 3 objects of the same type and
only 2 of them are under the same parent. e.g.:
/container/obj1 (TYPE_FOO)
/container/obj2 (TYPE_FOO)
/obj2 (TYPE_FOO)
With the above tree, object_resolve_path_type("", TYPE_FOO, NULL) will
incorrectly return /obj2, because the search inside "/container" will
return NULL, and the match at "/obj2" won't be detected as ambiguous.
Fix that by always calling object_resolve_partial_path() with a non-NULL
ambiguous parameter.
Backports commit ebcc479eee740937e70a94a468effcf2126a572b from qemu
This patch fixes setting DExcCode field of CP0 Debug register
when SDBBP instruction is executed. According to EJTAG specification,
this field must be set to the value 9 (Bp).
Backports commit c6c2c0fc32362ba234ae3bdad1a55c2d6aefaa12 from qemu
