On Tue, Oct 20, 2020 at 7:50 AM Vlastimil Babka <vbabka@xxxxxxx> wrote:
>
> On 10/10/20 12:05 AM, Axel Rasmussen wrote:
> > The goal of these tracepoints is to be able to debug lock contention
> > issues. This lock is acquired on most (all?) mmap / munmap / page fault
> > operations, so a multi-threaded process which does a lot of these can
> > experience significant contention.
> >
> > We trace just before we start acquisition, when the acquisition returns
> > (whether it succeeded or not), and when the lock is released (or
> > downgraded). The events are broken out by lock type (read / write).
> >
> > The events are also broken out by memcg path. For container-based
> > workloads, users often think of several processes in a memcg as a single
> > logical "task", so collecting statistics at this level is useful.
> >
> > The end goal is to get latency information. This isn't directly included
> > in the trace events. Instead, users are expected to compute the time
> > between "start locking" and "acquire returned", using e.g. synthetic
> > events or BPF. The benefit we get from this is simpler code.
> >
> > Because we use tracepoint_enabled() to decide whether or not to trace,
> > this patch has effectively no overhead unless tracepoints are enabled at
> > runtime. If tracepoints are enabled, there is a performance impact, but
> > how much depends on exactly what e.g. the BPF program does.
> >
> > Signed-off-by: Axel Rasmussen <axelrasmussen@xxxxxxxxxx>
>
> Yeah I agree with this approach that follows the page ref one.
>
> ...
>
> > diff --git a/mm/mmap_lock.c b/mm/mmap_lock.c
> > new file mode 100644
> > index 000000000000..b849287bd12a
> > --- /dev/null
> > +++ b/mm/mmap_lock.c
> > @@ -0,0 +1,87 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +#define CREATE_TRACE_POINTS
> > +#include <trace/events/mmap_lock.h>
> > +
> > +#include <linux/mm.h>
> > +#include <linux/cgroup.h>
> > +#include <linux/memcontrol.h>
> > +#include <linux/mmap_lock.h>
> > +#include <linux/percpu.h>
> > +#include <linux/smp.h>
> > +#include <linux/trace_events.h>
> > +
> > +/*
> > + * We have to export these, as drivers use mmap_lock, and our inline functions
> > + * in the header check if the tracepoint is enabled. They can't be GPL, as e.g.
> > + * the nvidia driver is an existing caller of this code.
>
> I don't think this argument works in the kernel community. I would just remove
> this comment.
>
> > + */
> > +EXPORT_SYMBOL(__tracepoint_mmap_lock_start_locking);
> > +EXPORT_SYMBOL(__tracepoint_mmap_lock_acquire_returned);
> > +EXPORT_SYMBOL(__tracepoint_mmap_lock_released);
>
> You can use EXPORT_TRACEPOINT_SYMBOL() here.
This is simpler, thanks for the pointer!
Agree the comment isn't needed in this case. I added it mainly since
checkpatch.pl doesn't like the EXPORT_SYMBOL-ing things not defined
just above, but EXPORT_TRACEPOINT_SYMBOL won't raise the same
concerns.
>
> > +#ifdef CONFIG_MEMCG
> > +
> > +DEFINE_PER_CPU(char[MAX_FILTER_STR_VAL], trace_memcg_path);
> > +
> > +/*
> > + * Write the given mm_struct's memcg path to a percpu buffer, and return a
> > + * pointer to it. If the path cannot be determined, the buffer will contain the
> > + * empty string.
> > + *
> > + * Note: buffers are allocated per-cpu to avoid locking, so preemption must be
> > + * disabled by the caller before calling us, and re-enabled only after the
> > + * caller is done with the pointer.
> > + */
> > +static const char *get_mm_memcg_path(struct mm_struct *mm)
> > +{
> > + struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
> > +
> > + if (memcg != NULL && likely(memcg->css.cgroup != NULL)) {
> > + char *buf = this_cpu_ptr(trace_memcg_path);
> > +
> > + cgroup_path(memcg->css.cgroup, buf, MAX_FILTER_STR_VAL);
> > + return buf;
> > + }
> > + return "";
> > +}
> > +
> > +#define TRACE_MMAP_LOCK_EVENT(type, mm, ...) \
> > + do { \
> > + if (trace_mmap_lock_##type##_enabled()) { \
>
> Is this check really needed? We only got called from the functions inlined in
> the .h file because tracepoint_enabled() was true in the first place, so this
> seems redundant.
Right, now that we've moved the check into the header, this isn't needed.
>
> > + get_cpu(); \
> > + trace_mmap_lock_##type(mm, get_mm_memcg_path(mm), \
> > + ##__VA_ARGS__); \
> > + put_cpu(); \
> > + } \
> > + } while (0)
> > +
> > +#else /* !CONFIG_MEMCG */
> > +
> > +#define TRACE_MMAP_LOCK_EVENT(type, mm, ...) \
> > + trace_mmap_lock_##type(mm, "", ##__VA_ARGS__)
> > +
> > +#endif /* CONFIG_MEMCG */
> > +
> > +/*
> > + * Trace calls must be in a separate file, as otherwise there's a circular
> > + * dependency between linux/mmap_lock.h and trace/events/mmap_lock.h.
> > + */
> > +
> > +void __mmap_lock_do_trace_start_locking(struct mm_struct *mm, bool write)
> > +{
> > + TRACE_MMAP_LOCK_EVENT(start_locking, mm, write, true);
>
> Seems wasteful to have an always-true success field here. Yeah, not reusing the
> same event class for all three tracepoints means more code, but for tracing
> efficiency it's worth it, IMHO.
Right, originally I was worried about code size. But, I switched to
not re-using an event class, and I only measure an increase of 524
bytes in .text, which seems trivial.
I'll send a v4 with all of the above changes.
>
> > +}
> > +EXPORT_SYMBOL(__mmap_lock_do_trace_start_locking);
> > +
> > +void __mmap_lock_do_trace_acquire_returned(struct mm_struct *mm, bool write,
> > + bool success)
> > +{
> > + TRACE_MMAP_LOCK_EVENT(acquire_returned, mm, write, success);
> > +}
> > +EXPORT_SYMBOL(__mmap_lock_do_trace_acquire_returned);
> > +
> > +void __mmap_lock_do_trace_released(struct mm_struct *mm, bool write)
> > +{
> > + TRACE_MMAP_LOCK_EVENT(released, mm, write, true);
>
> Ditto.
>
> > +}
> > +EXPORT_SYMBOL(__mmap_lock_do_trace_released);
> >
>
