The recent commit:
c6e7bd7afa ("sched/core: Optimize ttwu() spinning on p->on_cpu")
moved these lines in ttwu():
p->sched_contributes_to_load = !!task_contributes_to_load(p);
p->state = TASK_WAKING;
up before:
smp_cond_load_acquire(&p->on_cpu, !VAL);
into the 'p->on_rq == 0' block, with the thinking that once we hit
schedule() the current task cannot change it's ->state anymore. And
while this is true, it is both incorrect and flawed.
It is incorrect in that we need at least an ACQUIRE on 'p->on_rq == 0'
to avoid weak hardware from re-ordering things for us. This can fairly
easily be achieved by relying on the control-dependency already in
place.
The second problem, which makes the flaw in the original argument, is
that while schedule() will not change prev->state, it will read it a
number of times (arguably too many times since it's marked volatile).
The previous condition 'p->on_cpu == 0' was sufficient because that
indicates schedule() has completed, and will no longer read
prev->state. So now the trick is to make this same true for the (much)
earlier 'prev->on_rq == 0' case.
Furthermore, in order to make the ordering stick, the 'prev->on_rq = 0'
assignment needs to he a RELEASE, but adding additional ordering to
schedule() is an unwelcome proposition at the best of times, doubly so
for mere accounting.
Luckily we can push the prev->state load up before rq->lock, with the
only caveat that we then have to re-read the state after. However, we
know that if it changed, we no longer have to worry about the blocking
path. This gives us the required ordering, if we block, we did the
prev->state load before an (effective) smp_mb() and the p->on_rq store
needs not change.
With this we end up with the effective ordering:
LOAD p->state LOAD-ACQUIRE p->on_rq == 0
MB
STORE p->on_rq, 0 STORE p->state, TASK_WAKING
which ensures the TASK_WAKING store happens after the prev->state
load, and all is well again.
Fixes: c6e7bd7afa ("sched/core: Optimize ttwu() spinning on p->on_cpu")
Reported-by: Dave Jones <davej@codemonkey.org.uk>
Reported-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Dave Jones <davej@codemonkey.org.uk>
Tested-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Link: https://lkml.kernel.org/r/20200707102957.GN117543@hirez.programming.kicks-ass.net
As discussed on-list (cf. [1]), in order to make setns() support time
namespaces when attaching to multiple namespaces at once properly we
need to tweak vdso_join_timens() to always succeed. So switch
vdso_join_timens() to using a read lock and replacing
mmap_write_lock_killable() to mmap_read_lock() as we discussed.
Last cycle setns() was changed to support attaching to multiple namespaces
atomically. This requires all namespaces to have a point of no return where
they can't fail anymore. Specifically, <namespace-type>_install() is
allowed to perform permission checks and install the namespace into the new
struct nsset that it has been given but it is not allowed to make visible
changes to the affected task. Once <namespace-type>_install() returns
anything that the given namespace type requires to be setup in addition
needs to ideally be done in a function that can't fail or if it fails the
failure is not fatal. For time namespaces the relevant functions that fall
into this category are timens_set_vvar_page() and vdso_join_timens().
Currently the latter can fail but doesn't need to. With this we can go on
to implement a timens_commit() helper in a follow up patch to be used by
setns().
[1]: https://lore.kernel.org/lkml/20200611110221.pgd3r5qkjrjmfqa2@wittgenstein
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Andrei Vagin <avagin@gmail.com>
Cc: Will Deacon <will@kernel.org>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Dmitry Safonov <dima@arista.com>
Cc: linux-arm-kernel@lists.infradead.org
Link: https://lore.kernel.org/r/20200706154912.3248030-2-christian.brauner@ubuntu.com
Sometimes it's handy to know when the socket gets freed. In
particular, we'd like to try to use a smarter allocation of
ports for bpf_bind and explore the possibility of limiting
the number of SOCK_DGRAM sockets the process can have.
Implement BPF_CGROUP_INET_SOCK_RELEASE hook that triggers on
inet socket release. It triggers only for userspace sockets
(not in-kernel ones) and therefore has the same semantics as
the existing BPF_CGROUP_INET_SOCK_CREATE.
Signed-off-by: Stanislav Fomichev <sdf@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Link: https://lore.kernel.org/bpf/20200706230128.4073544-2-sdf@google.com
When we clone a socket in sk_clone_lock(), its sk_cgrp_data is
copied, so the cgroup refcnt must be taken too. And, unlike the
sk_alloc() path, sock_update_netprioidx() is not called here.
Therefore, it is safe and necessary to grab the cgroup refcnt
even when cgroup_sk_alloc is disabled.
sk_clone_lock() is in BH context anyway, the in_interrupt()
would terminate this function if called there. And for sk_alloc()
skcd->val is always zero. So it's safe to factor out the code
to make it more readable.
The global variable 'cgroup_sk_alloc_disabled' is used to determine
whether to take these reference counts. It is impossible to make
the reference counting correct unless we save this bit of information
in skcd->val. So, add a new bit there to record whether the socket
has already taken the reference counts. This obviously relies on
kmalloc() to align cgroup pointers to at least 4 bytes,
ARCH_KMALLOC_MINALIGN is certainly larger than that.
This bug seems to be introduced since the beginning, commit
d979a39d72 ("cgroup: duplicate cgroup reference when cloning sockets")
tried to fix it but not compeletely. It seems not easy to trigger until
the recent commit 090e28b229
("netprio_cgroup: Fix unlimited memory leak of v2 cgroups") was merged.
Fixes: bd1060a1d6 ("sock, cgroup: add sock->sk_cgroup")
Reported-by: Cameron Berkenpas <cam@neo-zeon.de>
Reported-by: Peter Geis <pgwipeout@gmail.com>
Reported-by: Lu Fengqi <lufq.fnst@cn.fujitsu.com>
Reported-by: Daniël Sonck <dsonck92@gmail.com>
Reported-by: Zhang Qiang <qiang.zhang@windriver.com>
Tested-by: Cameron Berkenpas <cam@neo-zeon.de>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Tested-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Zefan Li <lizefan@huawei.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Create an independent helper thread_group_exited which returns true
when all threads have passed exit_notify in do_exit. AKA all of the
threads are at least zombies and might be dead or completely gone.
Create this helper by taking the logic out of pidfd_poll where it is
already tested, and adding a READ_ONCE on the read of
task->exit_state.
I will be changing the user mode driver code to use this same logic
to know when a user mode driver needs to be restarted.
Place the new helper thread_group_exited in kernel/exit.c and
EXPORT it so it can be used by modules.
Link: https://lkml.kernel.org/r/20200702164140.4468-13-ebiederm@xmission.com
Acked-by: Christian Brauner <christian.brauner@ubuntu.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Tested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Pull rcu fixlet from Thomas Gleixner:
"A single fix for a printk format warning in RCU"
* tag 'core-urgent-2020-07-05' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
rcuperf: Fix printk format warning
Daniel Borkmann says:
====================
pull-request: bpf-next 2020-07-04
The following pull-request contains BPF updates for your *net-next* tree.
We've added 73 non-merge commits during the last 17 day(s) which contain
a total of 106 files changed, 5233 insertions(+), 1283 deletions(-).
The main changes are:
1) bpftool ability to show PIDs of processes having open file descriptors
for BPF map/program/link/BTF objects, relying on BPF iterator progs
to extract this info efficiently, from Andrii Nakryiko.
2) Addition of BPF iterator progs for dumping TCP and UDP sockets to
seq_files, from Yonghong Song.
3) Support access to BPF map fields in struct bpf_map from programs
through BTF struct access, from Andrey Ignatov.
4) Add a bpf_get_task_stack() helper to be able to dump /proc/*/stack
via seq_file from BPF iterator progs, from Song Liu.
5) Make SO_KEEPALIVE and related options available to bpf_setsockopt()
helper, from Dmitry Yakunin.
6) Optimize BPF sk_storage selection of its caching index, from Martin
KaFai Lau.
7) Removal of redundant synchronize_rcu()s from BPF map destruction which
has been a historic leftover, from Alexei Starovoitov.
8) Several improvements to test_progs to make it easier to create a shell
loop that invokes each test individually which is useful for some CIs,
from Jesper Dangaard Brouer.
9) Fix bpftool prog dump segfault when compiled without skeleton code on
older clang versions, from John Fastabend.
10) Bunch of cleanups and minor improvements, from various others.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
All architectures support copy_thread_tls() now, so remove the legacy
copy_thread() function and the HAVE_COPY_THREAD_TLS config option. Everyone
uses the same process creation calling convention based on
copy_thread_tls() and struct kernel_clone_args. This will make it easier to
maintain the core process creation code under kernel/, simplifies the
callpaths and makes the identical for all architectures.
Cc: linux-arch@vger.kernel.org
Acked-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Acked-by: Greentime Hu <green.hu@gmail.com>
Acked-by: Geert Uytterhoeven <geert@linux-m68k.org>
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Now that all architectures have been switched to use _do_fork() and the new
struct kernel_clone_args calling convention we can remove the legacy
do_fork() helper completely. The calling convention used to be brittle and
do_fork() didn't buy us anything. The only calling convention accepted
should be based on struct kernel_clone_args going forward. It's cleaner and
uniform.
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: "Peter Zijlstra (Intel)" <peterz@infradead.org>
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Use struct pid instead of user space pid values that are prone to wrap
araound.
In addition track the entire thread group instead of just the first
thread that is started by exec. There are no multi-threaded user mode
drivers today but there is nothing preclucing user drivers from being
multi-threaded, so it is just a good idea to track the entire process.
Take a reference count on the tgid's in question to make it possible
to remove exit_umh in a future change.
As a struct pid is available directly use kill_pid_info.
The prior process signalling code was iffy in using a userspace pid
known to be in the initial pid namespace and then looking up it's task
in whatever the current pid namespace is. It worked only because
kernel threads always run in the initial pid namespace.
As the tgid is now refcounted verify the tgid is NULL at the start of
fork_usermode_driver to avoid the possibility of silent pid leaks.
v1: https://lkml.kernel.org/r/87mu4qdlv2.fsf_-_@x220.int.ebiederm.org
v2: https://lkml.kernel.org/r/a70l4oy8.fsf_-_@x220.int.ebiederm.org
Link: https://lkml.kernel.org/r/20200702164140.4468-12-ebiederm@xmission.com
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Tested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Instead of loading a binary blob into a temporary file with
shmem_kernel_file_setup load a binary blob into a temporary tmpfs
filesystem. This means that the blob can be stored in an init section
and discared, and it means the binary blob will have a filename so can
be executed normally.
The only tricky thing about this code is that in the helper function
blob_to_mnt __fput_sync is used. That is because a file can not be
executed if it is still open for write, and the ordinary delayed close
for kernel threads does not happen soon enough, which causes the
following exec to fail. The function umd_load_blob is not called with
any locks so this should be safe.
Executing the blob normally winds up correcting several problems with
the user mode driver code discovered by Tetsuo Handa[1]. By passing
an ordinary filename into the exec, it is no longer necessary to
figure out how to turn a O_RDWR file descriptor into a properly
referende counted O_EXEC file descriptor that forbids all writes. For
path based LSMs there are no new special cases.
[1] https://lore.kernel.org/linux-fsdevel/2a8775b4-1dd5-9d5c-aa42-9872445e0942@i-love.sakura.ne.jp/
v1: https://lkml.kernel.org/r/87d05mf0j9.fsf_-_@x220.int.ebiederm.org
v2: https://lkml.kernel.org/r/87wo3p4p35.fsf_-_@x220.int.ebiederm.org
Link: https://lkml.kernel.org/r/20200702164140.4468-8-ebiederm@xmission.com
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Tested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Pull data race annotation from Christian Brauner:
"This contains an annotation patch for a data race in copy_process()
reported by KCSAN when reading and writing nr_threads.
The data race is intentional and benign. This is obvious from the
comment above the relevant code and based on general consensus when
discussing this issue. So simply using data_race() to annotate this as
an intentional race seems the best option"
* tag 'for-linus-2020-07-02' of git://git.kernel.org/pub/scm/linux/kernel/git/brauner/linux:
fork: annotate data race in copy_process()
Pull io_uring fixes from Jens Axboe:
"One fix in here, for a regression in 5.7 where a task is waiting in
the kernel for a condition, but that condition won't become true until
task_work is run. And the task_work can't be run exactly because the
task is waiting in the kernel, so we'll never make any progress.
One example of that is registering an eventfd and queueing io_uring
work, and then the task goes and waits in eventfd read with the
expectation that it'll get woken (and read an event) when the io_uring
request completes. The io_uring request is finished through task_work,
which won't get run while the task is looping in eventfd read"
* tag 'io_uring-5.8-2020-07-01' of git://git.kernel.dk/linux-block:
io_uring: use signal based task_work running
task_work: teach task_work_add() to do signal_wake_up()
Right now user-space tools like 'makedumpfile' and 'crash' need to rely
on a best-guess method of determining value of 'MAX_PHYSMEM_BITS'
supported by underlying kernel.
This value is used in user-space code to calculate the bit-space
required to store a section for SPARESMEM (similar to the existing
calculation method used in the kernel implementation):
#define SECTIONS_SHIFT (MAX_PHYSMEM_BITS - SECTION_SIZE_BITS)
Now, regressions have been reported in user-space utilities
like 'makedumpfile' and 'crash' on arm64, with the recently added
kernel support for 52-bit physical address space, as there is
no clear method of determining this value in user-space
(other than reading kernel CONFIG flags).
As per suggestion from makedumpfile maintainer (Kazu), it makes more
sense to append 'MAX_PHYSMEM_BITS' to vmcoreinfo in the core code itself
rather than in arch-specific code, so that the user-space code for other
archs can also benefit from this addition to the vmcoreinfo and use it
as a standard way of determining 'SECTIONS_SHIFT' value in user-land.
A reference 'makedumpfile' implementation which reads the
'MAX_PHYSMEM_BITS' value from vmcoreinfo in a arch-independent fashion
is available here:
While at it also update vmcoreinfo documentation for 'MAX_PHYSMEM_BITS'
variable being added to vmcoreinfo.
'MAX_PHYSMEM_BITS' defines the maximum supported physical address
space memory.
Signed-off-by: Bhupesh Sharma <bhsharma@redhat.com>
Tested-by: John Donnelly <john.p.donnelly@oracle.com>
Acked-by: Dave Young <dyoung@redhat.com>
Cc: Boris Petkov <bp@alien8.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: James Morse <james.morse@arm.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Kazuhito Hagio <k-hagio@ab.jp.nec.com>
Cc: x86@kernel.org
Cc: linuxppc-dev@lists.ozlabs.org
Cc: linux-arm-kernel@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Cc: kexec@lists.infradead.org
Link: https://lore.kernel.org/r/1589395957-24628-2-git-send-email-bhsharma@redhat.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Currently, most CPUFreq governors are registered at the core_initcall
time when the given governor is the default one, and the module_init
time otherwise.
In preparation for letting users specify the default governor on the
kernel command line, change all of them to be registered at the
core_initcall unconditionally, as it is already the case for the
schedutil and performance governors. This will allow us to assume
that builtin governors have been registered before the built-in
CPUFreq drivers probe.
And since all governors have similar init/exit patterns now, introduce
two new macros, cpufreq_governor_{init,exit}(), to factorize the code.
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Quentin Perret <qperret@google.com>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
[ rjw: Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
After tweaking the ring buffer to be a bit faster, a warning is triggering
on one of my machines, and causing my tests to fail. This warning is caused
when the delta (current time stamp minus previous time stamp), is larger
than the max time held by the ring buffer (59 bits).
If the clock were to go backwards slightly, this would then easily trigger
this warning. The machine that it triggered on, the clock did go backwards
by around 450 nanoseconds, and this happened after a recalibration of the
TSC clock. Now that the ring buffer is faster, it detects this, and the
delta that is used larger than the max, the warning is triggered and my test
fails.
To handle the clock going backwards, look at the saved before and after time
stamps. If they are the same, it means that the current event did not
interrupt another event, and that those timestamp are of a previous event
that was recorded. If the max delta is triggered, look at those time stamps,
make sure they are the same, then use them to compare with the current
timestamp. If the current timestamp is less than the before/after time
stamps, then that means the clock being used went backward.
Print out a message that this has happened, but do not warn about it (and
only print the message once).
Still do the warning if the delta is indeed larger than what can be used.
Also remove the unneeded KERN_WARNING from the WARN_ONCE() print.
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
After doing some benchmarks and examining the code, I found that the ring
buffer clock calls were quite expensive, and noticed that it uses
retpolines. This is because the ring buffer clock is programmable, and can
be set. But in most cases it simply uses the fastest ns unit clock which is
the trace_clock_local(). For RETPOLINE builds, checking if the ring buffer
clock is set to trace_clock_local() and then calling it directly has brought
the time of an event on my i7 box from an average of 93 nanoseconds an event
down to 83 nanoseconds an event, and the minimum time from 81 nanoseconds to
68 nanoseconds!
Suggested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Make a helper function rb_add_timestamp() that moves the adding of the
extended time stamps into its own function. Also, remove the noinline and
inline for the functions it calls, as recent benchmarks appear they do not
make a difference (just let gcc decide).
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Reorganize a little the logic to handle adding the absolute time stamp,
extended and forced time stamps, in such a way to remove a branch or two.
This is just a micro optimization.
Also add before and after time stamps to the rb_event_info structure to
display those values in the rb_check_timestamps() code, if something were to
go wrong.
Suggested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Introduce helper bpf_get_task_stack(), which dumps stack trace of given
task. This is different to bpf_get_stack(), which gets stack track of
current task. One potential use case of bpf_get_task_stack() is to call
it from bpf_iter__task and dump all /proc/<pid>/stack to a seq_file.
bpf_get_task_stack() uses stack_trace_save_tsk() instead of
get_perf_callchain() for kernel stack. The benefit of this choice is that
stack_trace_save_tsk() doesn't require changes in arch/. The downside of
using stack_trace_save_tsk() is that stack_trace_save_tsk() dumps the
stack trace to unsigned long array. For 32-bit systems, we need to
translate it to u64 array.
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Link: https://lore.kernel.org/bpf/20200630062846.664389-3-songliubraving@fb.com
bpf_free_used_maps() or close(map_fd) will trigger map_free callback.
bpf_free_used_maps() is called after bpf prog is no longer executing:
bpf_prog_put->call_rcu->bpf_prog_free->bpf_free_used_maps.
Hence there is no need to call synchronize_rcu() to protect map elements.
Note that hash_of_maps and array_of_maps update/delete inner maps via
sys_bpf() that calls maybe_wait_bpf_programs() and synchronize_rcu().
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Acked-by: Paul E. McKenney <paulmck@kernel.org>
Link: https://lore.kernel.org/bpf/20200630043343.53195-2-alexei.starovoitov@gmail.com
The unicore32 port do not seem maintained for a long time now, there is no
upstream toolchain that can create unicore32 binaries and all the links to
prebuilt toolchains for unicore32 are dead. Even compilers that were
available are not supported by the kernel anymore.
Guenter Roeck says:
I have stopped building unicore32 images since v4.19 since there is no
available compiler that is still supported by the kernel. I am surprised
that support for it has not been removed from the kernel.
Remove unicore32 port.
Signed-off-by: Mike Rapoport <rppt@linux.ibm.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Guenter Roeck <linux@roeck-us.net>
Daniel Borkmann says:
====================
pull-request: bpf 2020-06-30
The following pull-request contains BPF updates for your *net* tree.
We've added 28 non-merge commits during the last 9 day(s) which contain
a total of 35 files changed, 486 insertions(+), 232 deletions(-).
The main changes are:
1) Fix an incorrect verifier branch elimination for PTR_TO_BTF_ID pointer
types, from Yonghong Song.
2) Fix UAPI for sockmap and flow_dissector progs that were ignoring various
arguments passed to BPF_PROG_{ATTACH,DETACH}, from Lorenz Bauer & Jakub Sitnicki.
3) Fix broken AF_XDP DMA hacks that are poking into dma-direct and swiotlb
internals and integrate it properly into DMA core, from Christoph Hellwig.
4) Fix RCU splat from recent changes to avoid skipping ingress policy when
kTLS is enabled, from John Fastabend.
5) Fix BPF ringbuf map to enforce size to be the power of 2 in order for its
position masking to work, from Andrii Nakryiko.
6) Fix regression from CAP_BPF work to re-allow CAP_SYS_ADMIN for loading
of network programs, from Maciej Żenczykowski.
7) Fix libbpf section name prefix for devmap progs, from Jesper Dangaard Brouer.
8) Fix formatting in UAPI documentation for BPF helpers, from Quentin Monnet.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
It is the uncommon case where an event crosses a sub buffer boundary (page)
mark that check at the end of reserving an event as unlikely.
Suggested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
On a 144 thread system, `perf ftrace` takes about 20 seconds to start
up, due to calling synchronize_rcu() for each CPU.
cat /proc/108560/stack
0xc0003e7eb336f470
__switch_to+0x2e0/0x480
__wait_rcu_gp+0x20c/0x220
synchronize_rcu+0x9c/0xc0
ring_buffer_reset_cpu+0x88/0x2e0
tracing_reset_online_cpus+0x84/0xe0
tracing_open+0x1d4/0x1f0
On a system with 10x more threads, it starts to become an annoyance.
Batch these up so we disable all the per-cpu buffers first, then
synchronize_rcu() once, then reset each of the buffers. This brings
the time down to about 0.5s.
Link: https://lkml.kernel.org/r/20200625053403.2386972-1-npiggin@gmail.com
Tested-by: Anton Blanchard <anton@ozlabs.org>
Acked-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
After a discussion with the new time algorithm to have nested events still
have proper time keeping but required using local64_t atomic operations.
Mathieu was concerned about the performance this would have on 32 bit
machines, as in most cases, atomic 64 bit operations on them can be
expensive.
As the ring buffer's timing needs do not require full features of local64_t,
a wrapper is made to implement a new rb_time_t operation that uses two longs
on 32 bit machines but still uses the local64_t operations on 64 bit
machines. There's a switch that can be made in the file to force 64 bit to
use the 32 bit version just for testing purposes.
All reads do not need to succeed if a read happened while the stamp being
read is in the process of being updated. The requirement is that all reads
must succed that were done by an interrupting event (where this event was
interrupted by another event that did the write). Or if the event itself did
the write first. That is: rb_time_set(t, x) followed by rb_time_read(t) will
always succeed (even if it gets interrupted by another event that writes to
t. The result of the read will be either the previous set, or a set
performed by an interrupting event.
If the read is done by an event that interrupted another event that was in
the process of setting the time stamp, and no other event came along to
write to that time stamp, it will fail and the rb_time_read() will return
that it failed (the value to read will be undefined).
A set will always write to the time stamp and return with a valid time
stamp, such that any read after it will be valid.
A cmpxchg may fail if it interrupted an event that was in the process of
updating the time stamp just like the reads do. Other than that, it will act
like a normal cmpxchg.
The way this works is that the rb_time_t is made of of three fields. A cnt,
that gets updated atomically everyting a modification is made. A top that
represents the most significant 30 bits of the time, and a bottom to
represent the least significant 30 bits of the time. Notice, that the time
values is only 60 bits long (where the ring buffer only uses 59 bits, which
gives us 18 years of nanoseconds!).
The top two bits of both the top and bottom is a 2 bit counter that gets set
by the value of the least two significant bits of the cnt. A read of the top
and the bottom where both the top and bottom have the same most significant
top 2 bits, are considered a match and a valid 60 bit number can be created
from it. If they do not match, then the number is considered invalid, and
this must only happen if an event interrupted another event in the midst of
updating the time stamp.
This is only used for 32 bits machines as 64 bit machines can get better
performance out of the local64_t. This has been tested heavily by forcing 64
bit to use this logic.
Link: https://lore.kernel.org/r/20200625225345.18cf5881@oasis.local.home
Link: http://lkml.kernel.org/r/20200629025259.309232719@goodmis.org
Inspired-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Wenbo reported an issue in [1] where a checking of null
pointer is evaluated as always false. In this particular
case, the program type is tp_btf and the pointer to
compare is a PTR_TO_BTF_ID.
The current verifier considers PTR_TO_BTF_ID always
reprents a non-null pointer, hence all PTR_TO_BTF_ID compares
to 0 will be evaluated as always not-equal, which resulted
in the branch elimination.
For example,
struct bpf_fentry_test_t {
struct bpf_fentry_test_t *a;
};
int BPF_PROG(test7, struct bpf_fentry_test_t *arg)
{
if (arg == 0)
test7_result = 1;
return 0;
}
int BPF_PROG(test8, struct bpf_fentry_test_t *arg)
{
if (arg->a == 0)
test8_result = 1;
return 0;
}
In above bpf programs, both branch arg == 0 and arg->a == 0
are removed. This may not be what developer expected.
The bug is introduced by Commit cac616db39 ("bpf: Verifier
track null pointer branch_taken with JNE and JEQ"),
where PTR_TO_BTF_ID is considered to be non-null when evaluting
pointer vs. scalar comparison. This may be added
considering we have PTR_TO_BTF_ID_OR_NULL in the verifier
as well.
PTR_TO_BTF_ID_OR_NULL is added to explicitly requires
a non-NULL testing in selective cases. The current generic
pointer tracing framework in verifier always
assigns PTR_TO_BTF_ID so users does not need to
check NULL pointer at every pointer level like a->b->c->d.
We may not want to assign every PTR_TO_BTF_ID as
PTR_TO_BTF_ID_OR_NULL as this will require a null test
before pointer dereference which may cause inconvenience
for developers. But we could avoid branch elimination
to preserve original code intention.
This patch simply removed PTR_TO_BTD_ID from reg_type_not_null()
in verifier, which prevented the above branches from being eliminated.
[1]: https://lore.kernel.org/bpf/79dbb7c0-449d-83eb-5f4f-7af0cc269168@fb.com/T/
Fixes: cac616db39 ("bpf: Verifier track null pointer branch_taken with JNE and JEQ")
Reported-by: Wenbo Zhang <ethercflow@gmail.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Link: https://lore.kernel.org/bpf/20200630171240.2523722-1-yhs@fb.com
Instead of calling out the absolute test for each time to check if the
ring buffer wants absolute time stamps for all its recording, incorporate it
with the add_timestamp field and turn it into flags for faster processing
between wanting a absolute tag and needing to force one.
Link: http://lkml.kernel.org/r/20200629025259.154892368@goodmis.org
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Up until now, if an event is interrupted while it is recorded by an
interrupt, and that interrupt records events, the time of those events will
all be the same. This is because events only record the delta of the time
since the previous event (or beginning of a page), and to handle updating
the time keeping for that of nested events is extremely racy. After years of
thinking about this and several failed attempts, I finally have a solution
to solve this puzzle.
The problem is that you need to atomically calculate the delta and then
update the time stamp you made the delta from, as well as then record it
into the buffer, all this while at any time an interrupt can come in and
do the same thing. This is easy to solve with heavy weight atomics, but that
would be detrimental to the performance of the ring buffer. The current
state of affairs sacrificed the time deltas for nested events for
performance.
The reason for previous failed attempts at solving this puzzle was because I
was trying to completely avoid slow atomic operations like cmpxchg. I final
came to the conclusion to always avoid cmpxchg is not possible, which is why
those previous attempts always failed. But it is possible to pick one path
(the most common case) and avoid cmpxchg in that path, which is the "fast
path". The most common case is that an event will not be interrupted and
have other events added into it. An event can detect if it has
interrupted another event, and for these cases we can make it the slow
path and use the heavy operations like cmpxchg.
One more player was added to the game that made this possible, and that is
the "absolute timestamp" (by Tom Zanussi) that allows us to inject a full 59
bit time stamp. (Of course this breaks if a machine is running for more than
18 years without a reboot!).
There's barrier() placements around for being paranoid, even when they
are not needed because of other atomic functions near by. But those
should not hurt, as if they are not needed, they basically become a nop.
Note, this also makes the race window much smaller, which means there
are less slow paths to slow down the performance.
The basic idea is that there's two main paths taken.
1) Not being interrupted between time stamps and reserving buffer space.
In this case, the time stamps taken are true to the location in the
buffer.
2) Was interrupted by another path between taking time stamps and reserving
buffer space.
The objective is to know what the delta is from the last reserved location
in the buffer.
As it is possible to detect if an event is interrupting another event before
reserving data, space is added to the length to be reserved to inject a full
time stamp along with the event being reserved.
When an event is not interrupted, the write stamp is always the time of the
last event written to the buffer.
In path 1, there's two sub paths we care about:
a) The event did not interrupt another event.
b) The event interrupted another event.
In case a, as the write stamp was read and known to be correct, the delta
between the current time stamp and the write stamp is the delta between the
current event and the previously recorded event.
In case b, extra space was reserved to just put the full time stamp into the
buffer. Which is done, as stated, in this path the time stamp taken is known
to match the location in the buffer.
In path 2, there's also two sub paths we care about:
a) The event was not interrupted by another event since it reserved space
on the buffer and re-reading the write stamp.
b) The event was interrupted by another event.
In case a, the write stamp is that of the last event that interrupted this
event between taking the time stamps and reserving. As no event came in
after re-reading the write stamp, that event is known to be the time of the
event directly before this event and the delta can be the new time stamp and
the write stamp.
In case b, one or more events came in between reserving the event and
re-reading he write stamp. Since this event's buffer reservation is between
other events at this path, there's no way to know what the delta is. But
because an event interrupted this event after it started, its fine to just
give a zero delta, and take the same time stamp as the events that happened
within the event being recorded.
Here's the implementation of the design of this solution:
All this is per cpu, and only needs to worry about nested events (not
parallel events).
The players:
write_tail: The index in the buffer where new events can be written to.
It is incremented via local_add() to reserve space for a new event.
before_stamp: A time stamp set by all events before reserving space.
write_stamp: A time stamp updated by events after it has successfully
reserved space.
/* Save the current position of write */
[A] w = local_read(write_tail);
barrier();
/* Read both before and write stamps before touching anything */
before = local_read(before_stamp);
after = local_read(write_stamp);
barrier();
/*
* If before and after are the same, then this event is not
* interrupting a time update. If it is, then reserve space for adding
* a full time stamp (this can turn into a time extend which is
* just an extended time delta but fill up the extra space).
*/
if (after != before)
abs = true;
ts = clock();
/* Now update the before_stamp (everyone does this!) */
[B] local_set(before_stamp, ts);
/* Now reserve space on the buffer */
[C] write = local_add_return(len, write_tail);
/* Set tail to be were this event's data is */
tail = write - len;
if (w == tail) {
/* Nothing interrupted this between A and C */
[D] local_set(write_stamp, ts);
barrier();
[E] save_before = local_read(before_stamp);
if (!abs) {
/* This did not interrupt a time update */
delta = ts - after;
} else {
delta = ts; /* The full time stamp will be in use */
}
if (ts != save_before) {
/* slow path - Was interrupted between C and E */
/* The update to write_stamp could have overwritten the update to
* it by the interrupting event, but before and after should be
* the same for all completed top events */
after = local_read(write_stamp);
if (save_before > after)
local_cmpxchg(write_stamp, after, save_before);
}
} else {
/* slow path - Interrupted between A and C */
after = local_read(write_stamp);
temp_ts = clock();
barrier();
[F] if (write == local_read(write_tail) && after < temp_ts) {
/* This was not interrupted since C and F
* The last write_stamp is still valid for the previous event
* in the buffer. */
delta = temp_ts - after;
/* OK to keep this new time stamp */
ts = temp_ts;
} else {
/* Interrupted between C and F
* Well, there's no use to try to know what the time stamp
* is for the previous event. Just set delta to zero and
* be the same time as that event that interrupted us before
* the reservation of the buffer. */
delta = 0;
}
/* No need to use full timestamps here */
abs = 0;
}
Link: https://lkml.kernel.org/r/20200625094454.732790f7@oasis.local.home
Link: https://lore.kernel.org/r/20200627010041.517736087@goodmis.org
Link: http://lkml.kernel.org/r/20200629025258.957440797@goodmis.org
Reviewed-by: Masami Hiramatsu <mhiramat@kernel.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
If a process has the trace_pipe open on a trace_array, the current tracer
for that trace array should not be changed. This was original enforced by a
global lock, but when instances were introduced, it was moved to the
current_trace. But this structure is shared by all instances, and a
trace_pipe is for a single instance. There's no reason that a process that
has trace_pipe open on one instance should prevent another instance from
changing its current tracer. Move the reference counter to the trace_array
instead.
This is marked as "Fixes" but is more of a clean up than a true fix.
Backport if you want, but its not critical.
Fixes: cf6ab6d914 ("tracing: Add ref count to tracer for when they are being read by pipe")
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
So that the target task will exit the wait_event_interruptible-like
loop and call task_work_run() asap.
The patch turns "bool notify" into 0,TWA_RESUME,TWA_SIGNAL enum, the
new TWA_SIGNAL flag implies signal_wake_up(). However, it needs to
avoid the race with recalc_sigpending(), so the patch also adds the
new JOBCTL_TASK_WORK bit included in JOBCTL_PENDING_MASK.
TODO: once this patch is merged we need to change all current users
of task_work_add(notify = true) to use TWA_RESUME.
Cc: stable@vger.kernel.org # v5.7
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
