The three vector shift by vector operations are all implemented via
expansion. Therefore do not actually set TCG_TARGET_HAS_shv_vec,
as none of shlv_vec, shrv_vec, sarv_vec may actually appear in the
instruction stream, and therefore also do not appear in tcg_target_op_def.
Change the return value to bool, because that's what is should
have been from the start. Pass the ct mask instead of the whole
TCGArgConstraint, as that's the only part that's relevant.
Change the value argument to int64_t. We will need the extra
width for 32-bit hosts wanting to match vector constants.
Peter Maydell [Fri, 4 Jun 2021 16:27:29 +0000 (17:27 +0100)]
Merge remote-tracking branch 'remotes/bonzini-gitlab/tags/for-upstream' into staging
* OpenBSD cleanup (Brad)
* fixes for the i386 accel/cpu refactoring (Claudio)
* unmap test for emulated SCSI (Kit)
* fix for iscsi module (myself)
* fix for -readconfig of objects (myself)
* fixes for x86 16-bit task switching (myself)
* fix for x86 MOV from/to CR8 (Richard)
* remotes/bonzini-gitlab/tags/for-upstream:
vl: plug -object back into -readconfig
vl: plumb keyval-based options into -readconfig
qemu-config: parse configuration files to a QDict
i386: run accel_cpu_instance_init as post_init
i386: reorder call to cpu_exec_realizefn
tests/qtest/virtio-scsi-test: add unmap large LBA with 4k blocks test
target/i386: Fix decode of cr8
target/i386: tcg: fix switching from 16-bit to 32-bit tasks or vice versa
target/i386: tcg: fix loading of registers from 16-bit TSS
target/i386: tcg: fix segment register offsets for 16-bit TSS
oslib-posix: Remove OpenBSD workaround for fcntl("/dev/null", F_SETFL, O_NONBLOCK) failure
iscsi: link libm into the module
meson: allow optional dependencies for block modules
Peter Maydell [Fri, 4 Jun 2021 12:38:48 +0000 (13:38 +0100)]
Merge remote-tracking branch 'remotes/jasowang/tags/net-pull-request' into staging
# gpg: Signature made Fri 04 Jun 2021 08:26:16 BST
# gpg: using RSA key EF04965B398D6211
# gpg: Good signature from "Jason Wang (Jason Wang on RedHat) <[email protected]>" [marginal]
# gpg: WARNING: This key is not certified with sufficiently trusted signatures!
# gpg: It is not certain that the signature belongs to the owner.
# Primary key fingerprint: 215D 46F4 8246 689E C77F 3562 EF04 965B 398D 6211
Paolo Bonzini [Mon, 24 May 2021 10:57:52 +0000 (06:57 -0400)]
vl: plug -object back into -readconfig
Commit bc2f4fcb1d ("qom: move user_creatable_add_opts logic to vl.c
and QAPIfy it", 2021-03-19) switched the creation of objects from
qemu_opts_foreach to a bespoke QTAILQ in preparation for supporting JSON
syntax in -object.
Unfortunately in doing so it lost support for [object] stanzas in
configuration files and also for "-set object.ID.KEY=VAL". The latter
is hard to re-establish and probably best solved by deprecating -set.
This patch uses the infrastructure introduced by the previous two
patches in order to parse QOM objects correctly from configuration
files.
Paolo Bonzini [Mon, 24 May 2021 10:57:50 +0000 (06:57 -0400)]
qemu-config: parse configuration files to a QDict
Change the parser to put the values into a QDict and pass them
to a callback. qemu_config_parse's QemuOpts creation is
itself turned into a callback function.
This is useful for -readconfig to support keyval-based options;
getting a QDict from the parser removes a roundtrip from
QDict to QemuOpts and then back to QDict.
Unfortunately there is a disadvantage in that semantic errors will
point to the last line of the group, because the entries of the QDict
do not have a location attached.
Claudio Fontana [Thu, 3 Jun 2021 12:30:01 +0000 (14:30 +0200)]
i386: run accel_cpu_instance_init as post_init
This fixes host and max cpu initialization, by running the accel cpu
initialization only after all instance init functions are called for all
X86 cpu subclasses.
The bug this is fixing is related to the "max" and "host" i386 cpu
subclasses, which set cpu->max_features, which is then used at cpu
realization time.
In order to properly split the accel-specific max features code that
needs to be executed at cpu instance initialization time,
we cannot call the accel cpu initialization at the end of the x86 base
class initialization, or we will have no way to specialize
"max features" cpu behavior, overriding the "max" cpu class defaults,
and checking for the "max features" flag itself.
This patch moves the accel-specific cpu instance initialization to after
all x86 cpu instance code has been executed, including subclasses,
so that proper initialization of cpu "host" and "max" can be restored.
Claudio Fontana [Thu, 3 Jun 2021 12:30:00 +0000 (14:30 +0200)]
i386: reorder call to cpu_exec_realizefn
i386 realizefn code is sensitive to ordering, and recent commits
aimed at refactoring it, splitting accelerator-specific code,
broke assumptions which need to be fixed.
We need to:
* process hyper-v enlightements first, as they assume features
not to be expanded
* only then, expand features
* after expanding features, attempt to check them and modify them in the
accel-specific realizefn code called by cpu_exec_realizefn().
* after the framework has been called via cpu_exec_realizefn,
the code can check for what has or hasn't been set by accel-specific
code, or extend its results, ie:
- check and evenually set code_urev default
- modify cpu->mwait after potentially being set from host CPUID.
- finally check for phys_bits assuming all user and accel-specific
adjustments have already been taken into account.
Paolo Bonzini [Tue, 1 Jun 2021 13:36:00 +0000 (15:36 +0200)]
target/i386: tcg: fix switching from 16-bit to 32-bit tasks or vice versa
The format of the task state segment is governed by bit 3 in the
descriptor type field. On a task switch, the format for saving
is given by the current value of TR's type field, while the
format for loading is given by the new descriptor.
Paolo Bonzini [Tue, 1 Jun 2021 13:31:38 +0000 (15:31 +0200)]
target/i386: tcg: fix segment register offsets for 16-bit TSS
The TSS offsets in the manuals have only 2-byte slots for the
segment registers. QEMU incorrectly uses 4-byte slots, so
that SS overlaps the LDT selector.
Paolo Bonzini [Tue, 1 Jun 2021 08:00:48 +0000 (10:00 +0200)]
iscsi: link libm into the module
Depending on the configuration of QEMU, some binaries might not need libm
at all. In that case libiscsi, which uses exp(), will fail to load.
Link it in the module explicitly.
Paolo Bonzini [Tue, 1 Jun 2021 07:54:10 +0000 (09:54 +0200)]
meson: allow optional dependencies for block modules
Right now all dependencies for block modules are passed to
module_ss.add(when: ...), so they are mandatory. In the next patch we
will need to add a libm dependency to a module, but libm does not exist
on all systems. So, modify the creation of module_ss and modsrc so that
dependencies can also be passed to module_ss.add(if_true: ...).
While touching the array, remove the useless dependency of the curl
module on glib. glib is always linked in QEMU and in fact all other
block modules also need it, but they don't have to specify it.
Peter Maydell [Fri, 4 Jun 2021 09:04:11 +0000 (10:04 +0100)]
Merge remote-tracking branch 'remotes/rth-gitlab/tags/pull-fpu-20210603' into staging
Finish conversion of float128 and floatx80 to FloatParts.
Implement float128_muladd and float128_{min,max}*.
Optimize int-to-float conversion with hard-float.
# gpg: Signature made Thu 03 Jun 2021 22:13:10 BST
# gpg: using RSA key 7A481E78868B4DB6A85A05C064DF38E8AF7E215F
# gpg: issuer "[email protected]"
# gpg: Good signature from "Richard Henderson <[email protected]>" [full]
# Primary key fingerprint: 7A48 1E78 868B 4DB6 A85A 05C0 64DF 38E8 AF7E 215F
* remotes/rth-gitlab/tags/pull-fpu-20210603: (29 commits)
softfloat: Use hard-float for {u}int64_to_float{32,64}
tests/fp: Enable more tests
softfloat: Convert modrem operations to FloatParts
softfloat: Move floatN_log2 to softfloat-parts.c.inc
softfloat: Convert float32_exp2 to FloatParts
softfloat: Convert floatx80 compare to FloatParts
softfloat: Convert floatx80_scalbn to FloatParts
softfloat: Convert floatx80 to integer to FloatParts
softfloat: Convert floatx80 float conversions to FloatParts
softfloat: Convert integer to floatx80 to FloatParts
softfloat: Convert floatx80_round_to_int to FloatParts
softfloat: Convert floatx80_round to FloatParts
softfloat: Convert floatx80_sqrt to FloatParts
softfloat: Convert floatx80_div to FloatParts
softfloat: Convert floatx80_mul to FloatParts
softfloat: Convert floatx80_add/sub to FloatParts
tests/fp/fp-test: Reverse order of floatx80 precision tests
softfloat: Adjust parts_uncanon_normal for floatx80
softfloat: Introduce Floatx80RoundPrec
softfloat: Reduce FloatFmt
...
When RSS is enabled the device tries to load the eBPF program
to select RX virtqueue in the TUN. If eBPF can be loaded
the RSS will function also with vhost (works with kernel 5.8 and later).
Software RSS is used as a fallback with vhost=off when eBPF can't be loaded
or when hash population requested by the guest.
RSS program and Makefile to build it.
The bpftool used to generate '.h' file.
The data in that file may be loaded by libbpf.
EBPF compilation is not required for building qemu.
You can use Makefile if you need to regenerate rss.bpf.skeleton.h.
softfloat: Use hard-float for {u}int64_to_float{32,64}
For the normal case of no additional scaling, this reduces the
profile contribution of int64_to_float64 to the testcase in the
linked issue from 0.81% to 0.04%.
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/134 Reviewed-by: Alex Bennée <[email protected]> Signed-off-by: Richard Henderson <[email protected]>
softfloat: Move floatN_log2 to softfloat-parts.c.inc
Rename to parts$N_log2. Though this is partly a ruse, since I do not
believe the code will succeed for float128 without work. Which is ok
for now, because we do not need this for more than float32 and float64.
Since berkeley-testfloat-3 doesn't support log2, compare float64_log2
vs the system log2. Fix the errors for inputs near 1.0:
tests/fp/fp-test: Reverse order of floatx80 precision tests
Many qemu softfloat will check floatx80_rounding_precision
even when berkeley testfloat will not. So begin with
floatx80_precision_x, so that's the one we use
when !FUNC_EFF_ROUNDINGPRECISION.
softfloat: Adjust parts_uncanon_normal for floatx80
With floatx80_precision_x, the rounding happens across
the break between words. Notice this case with
frac_lsb = round_mask + 1 -> 0
and check the bits in frac_hi as needed.
In addition, since frac_shift == 0, we won't implicitly clear
round_mask via the right-shift, so explicitly clear those bits.
This fixes rounding for floatx80_precision_[sd].
softfloat: Move sqrt_float to softfloat-parts.c.inc
Rename to parts$N_sqrt.
Reimplement float128_sqrt with FloatParts128.
Reimplement with the inverse sqrt newton-raphson algorithm from musl.
This is significantly faster than even the berkeley sqrt n-r algorithm,
because it does not use division instructions, only multiplication.
Ordinarily, changing algorithms at the same time as migrating code is
a bad idea, but this is the only way I found that didn't break one of
the routines at the same time.
softfloat: Move compare_floats to softfloat-parts.c.inc
Rename to parts$N_compare. Rename all of the intermediate
functions to ftype_do_compare. Rename the hard-float functions
to ftype_hs_compare. Convert float128 to FloatParts128.
softfloat: Move minmax_flags to softfloat-parts.c.inc
Rename to parts$N_minmax. Combine 3 bool arguments to a bitmask.
Introduce ftype_minmax functions as a common optimization point.
Fold bfloat16 expansions into the same macro as the other types.
Peter Maydell [Thu, 3 Jun 2021 15:59:46 +0000 (16:59 +0100)]
Merge remote-tracking branch 'remotes/pmaydell/tags/pull-target-arm-20210603' into staging
target-arm queue:
* Some not-yet-enabled preliminaries for M-profile MVE support
* Consistently use "Cortex-Axx", not "Cortex Axx" in docs, comments
* docs: Fix installation of man pages with Sphinx 4.x
* Mark LDS{MIN,MAX} as signed operations
* Fix missing syndrome value for DAIF and PAC check exceptions
* Implement BFloat16 extensions
* Refactoring of hvf accelerator code in preparation for aarch64 support
* Fix some coverity nits in test code
* remotes/pmaydell/tags/pull-target-arm-20210603: (45 commits)
tests/unit/test-vmstate: Assert that dup() and mkstemp() succeed
tests/qtest/tpm-tests: Remove unnecessary NULL checks
tests/qtest/pflash-cfi02-test: Avoid potential integer overflow
tests/qtest/hd-geo-test: Fix checks on mkstemp() return value
tests/qtest/e1000e-test: Check qemu_recv() succeeded
tests/qtest/bios-tables-test: Check for dup2() failure
hvf: Simplify post reset/init/loadvm hooks
hvf: Introduce hvf vcpu struct
hvf: Remove hvf-accel-ops.h
hvf: Make synchronize functions static
hvf: Use cpu_synchronize_state()
hvf: Split out common code on vcpu init and destroy
hvf: Remove use of hv_uvaddr_t and hv_gpaddr_t
hvf: Make hvf_set_phys_mem() static
hvf: Move hvf internal definitions into common header
hvf: Move cpu functions into common directory
hvf: Move vcpu thread functions into common directory
hvf: Move assert_hvf_ok() into common directory
target/arm: Enable BFloat16 extensions
linux-user/aarch64: Enable hwcap bits for bfloat16
...
Coverity points out that in tpm_test_swtpm_migration_test() we
assume that src_tpm_addr and dst_tpm_addr are non-NULL (we
pass them to tpm_util_migration_start_qemu() which will
unconditionally dereference them) but then later explicitly
check them for NULL. Remove the pointless checks.
Coverity points out that we calculate a 64-bit value using 32-bit
arithmetic; add the cast to force the multiply to be done as 64-bits.
(The overflow will never happen with the current test data.)
Peter Maydell [Tue, 25 May 2021 13:44:55 +0000 (14:44 +0100)]
tests/qtest/hd-geo-test: Fix checks on mkstemp() return value
Coverity notices that the checks against mkstemp() failing in
create_qcow2_with_mbr() are wrong: mkstemp returns -1 on failure but
the check is just "g_assert(fd)". Fix to use "g_assert(fd >= 0)",
matching the correct check in create_test_img().
The e1000e_send_verify() test calls qemu_recv() but doesn't
check that the call succeeded, which annoys Coverity. Add
an explicit test check for the length of the data.
(This is a test check, not a "we assume this syscall always
succeeds", so we use g_assert_cmpint() rather than g_assert().)
Peter Maydell [Tue, 25 May 2021 13:44:53 +0000 (14:44 +0100)]
tests/qtest/bios-tables-test: Check for dup2() failure
Coverity notes that we don't check for dup2() failing. Add some
assertions so that if it does ever happen we get some indication.
(This is similar to how we handle other "don't expect this syscall to
fail" checks in this test code.)
Alexander Graf [Thu, 3 Jun 2021 13:42:41 +0000 (14:42 +0100)]
hvf: Simplify post reset/init/loadvm hooks
The hooks we have that call us after reset, init and loadvm really all
just want to say "The reference of all register state is in the QEMU
vcpu struct, please push it".
We already have a working pushing mechanism though called cpu->vcpu_dirty,
so we can just reuse that for all of the above, syncing state properly the
next time we actually execute a vCPU.
This fixes PSCI resets on ARM, as they modify CPU state even after the
post init call has completed, but before we execute the vCPU again.
To also make the scheme work for x86, we have to make sure we don't
move stale eflags into our env when the vcpu state is dirty.
Alexander Graf [Thu, 3 Jun 2021 13:09:34 +0000 (14:09 +0100)]
hvf: Introduce hvf vcpu struct
We will need more than a single field for hvf going forward. To keep
the global vcpu struct uncluttered, let's allocate a special hvf vcpu
struct, similar to how hax does it.
Alexander Graf [Thu, 3 Jun 2021 13:09:33 +0000 (14:09 +0100)]
hvf: Use cpu_synchronize_state()
There is no reason to call the hvf specific hvf_cpu_synchronize_state()
when we can just use the generic cpu_synchronize_state() instead. This
allows us to have less dependency on internal function definitions and
allows us to make hvf_cpu_synchronize_state() static.
Alexander Graf [Thu, 3 Jun 2021 13:09:32 +0000 (14:09 +0100)]
hvf: Split out common code on vcpu init and destroy
Until now, Hypervisor.framework has only been available on x86_64 systems.
With Apple Silicon shipping now, it extends its reach to aarch64. To
prepare for support for multiple architectures, let's start moving common
code out into its own accel directory.
This patch splits the vcpu init and destroy functions into a generic and
an architecture specific portion. This also allows us to move the generic
functions into the generic hvf code, removing exported functions.
Alexander Graf [Thu, 3 Jun 2021 13:09:31 +0000 (14:09 +0100)]
hvf: Move hvf internal definitions into common header
Until now, Hypervisor.framework has only been available on x86_64 systems.
With Apple Silicon shipping now, it extends its reach to aarch64. To
prepare for support for multiple architectures, let's start moving common
code out into its own accel directory.
This patch moves a few internal struct and constant defines over.
Alexander Graf [Thu, 3 Jun 2021 13:09:31 +0000 (14:09 +0100)]
hvf: Move cpu functions into common directory
Until now, Hypervisor.framework has only been available on x86_64 systems.
With Apple Silicon shipping now, it extends its reach to aarch64. To
prepare for support for multiple architectures, let's start moving common
code out into its own accel directory.
This patch moves CPU and memory operations over. While at it, make sure
the code is consumable on non-i386 systems.
Alexander Graf [Thu, 3 Jun 2021 13:09:31 +0000 (14:09 +0100)]
hvf: Move vcpu thread functions into common directory
Until now, Hypervisor.framework has only been available on x86_64 systems.
With Apple Silicon shipping now, it extends its reach to aarch64. To
prepare for support for multiple architectures, let's start moving common
code out into its own accel directory.
Alexander Graf [Thu, 3 Jun 2021 13:09:30 +0000 (14:09 +0100)]
hvf: Move assert_hvf_ok() into common directory
Until now, Hypervisor.framework has only been available on x86_64 systems.
With Apple Silicon shipping now, it extends its reach to aarch64. To
prepare for support for multiple architectures, let's start moving common
code out into its own accel directory.
This patch moves assert_hvf_ok() and introduces generic build infrastructure.
Jamie Iles [Wed, 26 May 2021 12:18:47 +0000 (13:18 +0100)]
target/arm: use raise_exception_ra for stack limit exception
The sequence cpu_restore_state() + raise_exception() is equivalent to
raise_exception_ra(), so use that instead. (In this case we never
cared about the syndrome value, because M-profile doesn't use the
syndrome; the old code was just written unnecessarily awkwardly.)
projects / qemu.git / log