Merge commit fde5924dcc69 from llvm-project (by Serge Pavlov):
[clang] Reset FP options before template instantiation
AST nodes that may depend on FP options keep them as a difference
relative to the options outside the AST node. At the moment of
instantiation the FP options may be different from the default values,
defined by command-line option. In such case FP attributes would have
unexpected values. For example, the code:
template <class C> void func_01(int last, C) {
func_01(last, int());
}
void func_02() { func_01(0, 1); }
#pragma STDC FENV_ACCESS ON
caused compiler crash, because template instantiation takes place at the
end of translation unit, where pragma STDC FENV_ACCESS is in effect. As
a result, code in the template instantiation would use constrained
intrinsics while the function does not have StrictFP attribute.
To solve this problem, FP attributes in Sema must be set to default
values, defined by command line options.
This change resolves https://github.com/llvm/llvm-project/issues/63542.
Doug Moore [Thu, 13 Jul 2023 18:28:21 +0000 (13:28 -0500)]
vfs_bio: initialize pctries
bufobj_init depends on fields bo_dirty.bv_root and bo_clean.bv_root
being zeroed on entry and pctrie_init zeroing whatever is passed to
them, and so does not call pctrie_init for either of them. That fails
if pctrie_init ever changes to do something other that zeroing data,
so add explicit calls to them.
Reviewed by: cem
Differential Revision: https://reviews.freebsd.org/D40978
Actions applied to a processed packet come in case of stateless
firewalling from a rule or in case of statefull firewalling from a
state. The state obtains the actions from a rule when it is created by a
rule or by pfsync. The logic for deciding if actions come from a rule or
a state is spread across many places in pf.
There already is struct pf_rule_actions in struct pf_pdesc and thus it
can be used as a central place for storing actions and their parameters.
OpenBSD does something similar: they also store the actions in struct
pf_pdesc and have no variables in pf_test() but they use separate
variables instead of a structure. By using struct pf_rule_actions we can
simplify the code even further. Applying of actions is done *only* in
pf_rule_to_actions() no matter if for the legacy scrub rules or for the
normal match / pass rules. The logic of choosing if rule or state
actions are used is applied only once in pf_test() by copying the whole
struct.
The re-initialisation of idx in the later else branch is a bit hidden,
and results in deviation from the earlier two paths. Use more consistent
code instead to make it abundantly clear what's going on.
The variable storing the direction of a processed packet is passed
around to many functions. Most of those functions already have a pointer
to struct pf_pdesc which also contains the direction. By using the one
in struct pf_pdesc we can reduce the amount of arguments passed around.
amd64 pmap: Fix compilation when superpage reservations are disabled
The function pmap_pde_ept_executable() should not be conditionally
compiled based on VM_NRESERVLEVEL. It is required indirectly by
pmap_enter(..., psind=1) even when reservation-based allocation is
disabled at compile time.
Eric van Gyzen [Tue, 11 Jul 2023 17:58:29 +0000 (12:58 -0500)]
dtrace: prevent forked child from running after an error condition
The pid/killonerror test uses an invalid probe specifier to verify that
the child process is killed. It occasionally fails because the "date"
command is allowed to run long enough to print the date. This is harmless
in this case, but is clearly not ideal.
When the dt_proc_control thread is about to exit, and the dtrace command
forked the child, do not make the child runnable.
getlocalbase: Make default path actually configurable
We include paths.h in getlocalbase.c, which defines _PATH_LOCALBASE.
This will override a definition on the command line, meaning it is
impossible to override like how the manpage says you can, and it means
the code to provide a fallback default is dead as the macro is always
defined.
Instead, introduce a new LOCALBASE_PATH macro like LOCALBASE_CTL_LEN
that can be set on the command line and will default to the system's
existing _PATH_LOCALBASE to avoid duplicating the default here.
kmod.mk: Use portable printf '%s' over non-portable echo -n
Whilst /bin/echo on macOS and Linux implement -n, as do the builtin
echos in bash and zsh, the builtin echo in dash does not, causing the
first line of the output to be -n foo rather than just foo, and there to
be an extra newline in the output and thus blank line, both of which
result in "Symbol ... is not present in *.kld" warnings appearing in the
build output (once for -n foo and once for the empty string for each
module where EXPORT_SYMS is a list of symbols).
csu: Implement _start using as to satisfy unwinders on i386
The right unwinding stop indicator should be CFI-undefined PC.
https://dwarfstd.org/doc/Dwarf3.pdf - page 118:
If a Return Address register is defined in the virtual unwind table,
and its rule is undefined (for example, by DW_CFA_undefined), then
there is no return address and no call address, and the virtual
unwind of stack activations is complete.
csu: Add the prologue and epilogue to the _init and _fini on i386
Normally, modern unwinders uses Dwarf information to unwind stack,
however in case when the code is not annotated by Dwarf instructions,
unwinders fallbacks to a frame-pointer based algorithm.
That is allows libunwind to unwind stack from global constructors and
destructors. Also it makes gdb happy as it printed nonexistent frame
before.
Add a stop indicator to rtld_start to satisfy unwinders:
The right unwinding stop indicator should be CFI-undefined PC.
https://dwarfstd.org/doc/Dwarf3.pdf - page 118:
If a Return Address register is defined in the virtual unwind table,
and its rule is undefined (for example, by DW_CFA_undefined), then
there is no return address and no call address, and the virtual
unwind of stack activations is complete.
That is allows gdb and libunwind successfully stop when unwinding stack
from global constructors and destructors.
This matches the beadm behavior; generally, we need to keep promoting
until the BE is no longer a clone from a snapshot. This fixes scenarios
where the dataset associated with a BE's origin is itself a clone,
activating the BE previously would promote it to a clone of the origin's
origin.
We could keep using be_get_dataset_props here, except for two
annoyances:
1.) I couldn't find a clean way to just clear an nvlist rather than
having to re-alloc it, and I didn't want to just remove the one prop
we're inspecting out of it.
2.) That's a lot of overhead when all we want to do is fetch the origin
anyways.
Note that this is not a complete fix, but it does fix the majority of
cases; deep BE subordinates are still notably broken, pending a patch
from Christian.
Reported by: R. Christian McDonald <rcm@rcm.sh>
Reviewed by: rew
Differential Revision: https://reviews.freebsd.org/D40903
Peter Grehan [Fri, 7 Jul 2023 08:23:42 +0000 (18:23 +1000)]
igc: Increase default per-queue interrupt rate to 20000.
The default per-queue packet rate of 8000 will cause packet loss when
forwarding at 2.5G with a single stream, as is common when using e.g.
iperf3 to test a platform.
Bump this to 20000 (the "low latency" value in the Linux driver) which
avoids packet loss for this type of test.
Future work will use adaptive interrupt rate in a similar fashion
to the ixgbe driver.
Make sure MACHINE(_ARCH) are exported for depend-cleanup.sh
We run depend-cleanup.sh twice during the build. The second time is the
normal run, where we run it under WMAKEENV and thus have CROSSENV's
MACHINE(_ARCH)=${TARGET(_ARCH)} in the environment. However, the first
time is for bootstrap-tools, where it's run under BMAKEENV and we don't
have any assignments to MACHINE(_ARCH) in the environment, meaning the
script sees them as unset. In practice this doesn't matter since the
only use doesn't apply to bootstrap-tools, but it could be a future
issue. Thus, explicitly export them for depend-cleanup.sh and have the
script verify they're set.
depend-cleanup.sh: cleanup stale ffs.S dependencies for arm and i386
Now that depend-cleanup.sh handles 32-bit dependencies a bit better, get
rid of the stale ffs.S ones, otherwise an incremental build on amd64
will fail with:
cc: error: no such file or directory: '/usr/src/lib/libc/i386/string/ffs.S'
depend-cleanup.sh: handle 32-bit dependencies having no 64-bit equivalent
For example with the recent removal of ffs.S for 32-bit arm and i386,
the egrep in clean_dep() did not find any files to remove, even if you
added a "clean_dep lib/libc ffs S" line. This is because it will never
find the ffs.S filename in the 64-bit .depend files for libc.
Split the searching and removing of 32-bit dependencies and objects into
a separate part to cope with this. This can be used similarly later on,
for other bitnesses.
Alan Somers [Mon, 3 Apr 2023 21:43:17 +0000 (15:43 -0600)]
Multiple fixes to the zfsd test suite
* Wait for gnop devices to disappear after "gnop destroy".
Apparently that process is asynchronous now, or maybe it's just slower
than it used to be. Also, after removing a gnop wait for its pool to
be degraded. That isn't instant.
* Fix the zfsd_degrade_001_pos test for recent zfs versions.
ZFS now rate limits checksum errors to about 20 per second. But
zfsd's threshold for degrading a disk is 50 per minute. So we must
alternately corrupt and scrub the pool to ensure that checksum errors
are generated in multiple 1-second windows, so that zfsd will see
enough of them.
* Fix the zfsd_fault_001_pos test in VMs
And, for that matter, when using NVME or SATA disks. As originally
written, the test used the da driver to inject errors. Rewrite it to
use gnop vdevs. gnop can also inject errors. It works on top of any
disk device, and it's also faster than using da.
Alan Somers [Mon, 3 Apr 2023 21:45:55 +0000 (15:45 -0600)]
zfsd: listen for sysevent.fs.zfs instead of misc.fs.zfs
At some point the names of these devd events changed. Probably it
happened when importing OpenZFS. Before that, FreeBSD's sysevent_alloc
method didn't create a "class" nvpair in the event, which led to
log_sysevent using the event's ev_subclass field as its type.
Colin Percival [Mon, 10 Jul 2023 06:16:31 +0000 (23:16 -0700)]
release: Add code to make FreeBSD/Firecracker
This just stages the kernel and builds a stripped-down rootfs for use
with the Firecracker VMM. At some point in the future the release
engineering team might start publishing these, but initially it's
just here to simplify FreeBSD/Firecracker development and testing.
Note that the rootfs generated:
* Uses an IP address of 10.0.0.2 with a gateway of 10.0.0.1,
* Has sshd enabled,
* Has user "freebsd" with password "freebsd" and a root password
of "root", and
* Is 1 GB in size (but has growfs enabled).
All of those are subject to change without notice; anyone intending to
use FreeBSD/Firecracker in anything remotely resembling a production
environment should talk to cperciva first.
Alan Somers [Sun, 9 Jul 2023 20:48:10 +0000 (14:48 -0600)]
During F_SETFL, don't change file flags on error
Previously, even if the FIONBIO or FIOASYNC ioctl failed, the file's
f_flags variable would still be changed. Now, kern_fcntl will restore
the original flags if the ioctl fails.
Doug Moore [Sun, 9 Jul 2023 20:06:02 +0000 (15:06 -0500)]
radix_trie: avoid code duplication in insert
Two cases in the insert routine are written differently, when
they're really doing the same thing. Writing that case only once
saves 208 bytes in the compiled vm_radix_insert code and reduces
instructions executed by about 2%.
Reviewed by: alc
Tested by: pho
Differential Revision: https://reviews.freebsd.org/D40807
The examples I wrote previously do not work. We parse the flags using
explicit names, not the shortened flag suffix. Fix the list of major
flags, and add a compact list of minor flags. Fix the examples, and
tweak some wording for clarity.
libsysdecode: Migrate from COMPAT_32BIT to generic COMPAT_LIBCOMPAT
The use isn't any more generic, just the variable itself, which will
allow COMPAT_32BIT to be removed. The fact we even have to check
COMPAT_LIBCOMPAT here in order to pass the right flags to CPP points at
our libcompat infrastructure not suitably modifying the CPP variable
(which we barely use for world; this and bsd.symver.mk are the two
uses, and the latter could benefit from the right flags too), but this
change doesn't attempt to fix that.
See commit 8fad2cda93c7 ("bsd.compat.mk: Provide new CPP and sub-make
variables") for the context behind this change.
rtld-elf: Migrate from COMPAT_32BIT to generic COMPAT_libcompat/LIBCOMPAT
We still have a tiny amount of libcompat-specific code in rtld_paths.h,
but it's been deduplicated as much as possible, and in future we may
wish to just push these variables down to the few consumers of them and
make them use the double-underscore variants with a libcompat argument
rather than give them names here.
See commit 8fad2cda93c7 ("bsd.compat.mk: Provide new CPP and sub-make
variables") for the context behind this change.
bsd.compat.mk: Provide new CPP and sub-make variables
Currently the only way to detect for a libcompat build is to consult
whether COMPAT_32BIT is defined (or equivalent, for downstreams with
other libcompats or past releases with libsoft as COMPAT_SOFTFP). There
are two issues with this:
1. COMPAT_32BIT is a new naming scheme that doesn't match the libcompat
name, which is unnecessary deviation.
2. When multiple libcompats exist, everywhere that needs to detect a
libcompat must check each variable in turn, despite the fact that it
normally just wants to know if this is a libcompat build and perhaps
what ${LIBCOMPAT} and/or ${libcompat} are for it.
As a result, far too many places in the tree need to know about the set
of possible libcompats.
Instead, introduce two new CPP and sub-make variables, COMPAT_LIBCOMPAT
and COMPAT_libcompat, which give the values for ${LIBCOMPAT} and
${libcompat} respectively, so that uses can be made parameterised. For
when code really does need to know the specific libcompat, Makefiles can
perform a string comparison, but the C preprocessor cannot, so introduce
an additional CPP-only COMPAT_LIB${LIBCOMPAT} which is intended to
replace the inconsistently-named COMPAT_32BIT (which will be removed in
future). Uses of this new variable should still be kept to a minimum,
however, given the code duplication needed for new libcompats.
Makefile.libcompat: Quote CFLAGS and CXXFLAGS for sub-make
Currently none of the words in these require quoting, but a future
commit will add words that do, thus we should make sure to quote each
word so the shell doesn't mangle them before calling the sub-make.
(Note that :@var@expr@ is the bmake syntax for map, replacing each word
with expr's evaluation, with var containing the input word)
Having the symbols exported by libc differ between i386 and amd64 lib32
is questionable. Since these files build just fine today, stop guarding
them with !defined(COMPAT_32BIT). Whether or not they work at run time
is a different matter, but an i386 jail would be similarly affected if
not, so that's not a problem with lib32.
We no longer have COMPAT_32BIT hacks for libusb, instead supporting the
normal 32-bit ioctls for freebsd32 processes, so we can enable these for
the lib32 build.
elf-hints.h: Allow rtld_paths.h to be included before it
Currently rtld_paths.h will #undef _PATH_ELF_HINTS in order to override
this to the right value if included afterwards, but the other way round
does not work as elf-hints.h tries to define an already-defined macro to
a potentially different value. Thus, guard the definition here so that
rtld_paths.h's definition continues to take precedence.
Note that, although all in-tree uses of _PATH_ELF_HINTS have included
rtld_paths.h already, pax-utils wants _PATH_ELF_HINTS from elf-hints.h
and so we cannot just drop the define. In-tree uses must just continue
to make sure that they include rtld_paths.h to get the right value for
libcompat builds as is already required.
openssl: include d2i_KeyParams() and d2i_KeyParams_bio()
These functions are new, and some ports (e.g.opensc) expect to have them
available. Add the file they're defined in to the build, and add them
to Version.map.
libthr: Patch to reduce latency to acquire+release a pthread mutex.
The acquisition and release of an uncontended default/normal pthread
mutex on FreeBSD is suprisingly slow, e.g., pthread wrlocks and binary
semaphores both exhibit roughly 33% lower latency, while default/normal
mutexes on Linux exhibit roughly 67% lower latency than FreeBSD. This is
likely explained by the fact that AFAICT in the best case to acquire an
uncontended mutex on Linux one need touch only 1 page and read+modify
only 1 cacheline, whereas on FreeBSD we need to touch at least 4 pages,
read 6 cachelines, and modify at least 4 cachelines.
This patch does not address the pthread mutex architecture. Instead,
it improves performance by adding the __always_inline attribute to
mutex_lock_common() and mutex_unlock_common() to encourage constant
folding and propagation, thereby lowering the latency to acquire and
release a mutex due to a shorter code path with fewer compares, jumps,
and mispredicts.
With this patch on a stock build I see a reduction in latency of roughly
7% for default/normal mutexes, and 17% for robust mutexes. When built
without PTHREADS_ASSERTIONS enabled I see a reduction in latency of
roughly 15% and 26%, respectively. Suprisingly, I see similar reductions
in latency for heavily contended mutexes.
By default, this patch increases the size of libthr.so.3 by 2448 bytes,
but when built without PTHREAD_ASSERTIONS enabled it only increases by
448 bytes.
Brooks Davis [Fri, 7 Jul 2023 22:25:11 +0000 (23:25 +0100)]
etc/rc.d/routing: use find_system_scripts
In 3693d9140e05aba9942232df13468f51a6cde136 /etc/rc switched to using
find_system_scripts rather than directly including /etc/rc.d/* in the
list of scripts to run in order to skip .pkgsave files. Follow suit
in etc/rc.d/routing.
Graham Perrin [Fri, 7 Jul 2023 22:05:46 +0000 (23:05 +0100)]
gpart(8): slices are out of context for GPT
Correct the GPT example. Creation of a partition leaves free the rest
of the device (not the slice).
Approved-by: imp Fixes: ae1b731b5df0 Rewrite the GPT and MBR examples. For GPT, ensure that the boot partition is large enough for gptzfsboot, which has doubled in size since 10.
MFC after: 1 week
Pull-request: https://github.com/freebsd/freebsd-src/pull/795