Re: [PATCH] perf tools: Add struct ordered_events_buffer layer

From: Stephane Eranian
Date: Tue Aug 14 2018 - 09:27:27 EST


Jiri,
On Mon, Aug 13, 2018 at 6:04 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
>
> On Fri, Aug 10, 2018 at 01:54:31PM +0200, Jiri Olsa wrote:
> > On Fri, Aug 10, 2018 at 01:21:18AM -0700, Stephane Eranian wrote:
> > > On Thu, Aug 9, 2018 at 1:07 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> > > >
> > > > On Wed, Aug 08, 2018 at 03:33:20PM -0700, Stephane Eranian wrote:
> > > > > This patch fixes a bug in ordered_event.c:alloc_event().
> > > > > An ordered_event struct was not initialized properly potentially
> > > > > causing crashes later on in free_dup_event() depending on the
> > > > > content of the memory. If it was NULL, then it would work fine,
> > > > > otherwise, it could cause crashes such as:
> > > >
> > > > I'm now little puzzled what do we use this first event for..
> > > > I can't see anything special about it, other than it's added
> > > > on the list uninitialized ;-)
> > > >
> > > > it seems to work properly when we ditch it.. might be some
> > > > prehistoric leftover or I'm terribly missing something
> > > >
> > > You need to keep track of the buffers to free. You do not free the
> > > ordered_event structs
> > > individually. For each oe->buffer, you need one free(). Each buffer is
> > > put in the to_free
> > > list. But to link it into the list it needs a list_head. This is what
> > > buffer[0] is used for.
> > > But the logic is broken in ordered_events__free(). It does not free individual
> > > ordered_event structs, but a buffer with many. Yet, it is missing
> > > freeing all of the duped
> > > events.
> > >
> > > void ordered_events__free(struct ordered_events *oe)
> > > {
> > > while (!list_empty(&oe->to_free)) {
> > > struct ordered_event *buffer;
> > >
> > > buffer = list_entry(oe->to_free.next, struct
> > > ordered_event, list);
> > > list_del(&buffer->list);
> > > ----> free_dup_event(oe, event->event);
> > > free(buffer);
> > > }
> > > }
> > > This only frees the dup_event of buffer[0] which we know is NULL (well, now).
> > > It needs to walk all the entries in buffer[] to free buffer[x].event.
> >
> > yes.. if there's copy_on_queue set, we need to do that,
> > otherwise we're leaking all the events
> >
> > >
> > > I think the goal was likely to avoid adding another list_head field to
> > > each ordered_event
> > > and instead use one per allocated buffer.
> > > This is very convoluted and prone to errors and we are seeing right
> > > now. This should
> > > be cleaned. So either you add a list_head to ordered_event or you
> > > would buffer[x] in
> > > ordered_events_free().
> > >
> > > At this point, this is my understanding.
> > > Do you agree?
> >
> > yea, I see it now.. thanks for pointing this out
> >
> > how about something like below? haven't tested properly yet
>
> attaching full patch
>
> thanks,
> jirka
>
>
> ---
> When ordering events, we use preallocated buffers to store separated
> events. Those buffers currently don't have their own struct, but since
> they are basically array of 'struct ordered_event' objects, we use the
> first event to hold buffers data - list head, that holds all buffers
> together:
>
> struct ordered_events {
> ...
> struct ordered_event *buffer;
> ...
> };
>
> struct ordered_event {
> u64 timestamp;
> u64 file_offset;
> union perf_event *event;
> struct list_head list;
> };
>
> This is quite convoluted and error prone as demonstrated by
> free-ing issue discovered and fixed by Stephane in here [1].
>
> This patch adds the 'struct ordered_events_buffer' object,
> that holds the buffer data and frees it up properly.
>
> [1] - https://marc.info/?l=linux-kernel&m=153376761329335&w=2
>
> Reported-by: Stephane Eranian <eranian@xxxxxxxxxx>
> Link: http://lkml.kernel.org/n/tip-qrkcqm5m1sugy4q83pfn5a1r@xxxxxxxxxxxxxx
> Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> ---
> tools/perf/util/ordered-events.c | 44 ++++++++++++++++++++++----------
> tools/perf/util/ordered-events.h | 37 +++++++++++++++------------
> 2 files changed, 52 insertions(+), 29 deletions(-)
>
> diff --git a/tools/perf/util/ordered-events.c b/tools/perf/util/ordered-events.c
> index bad9e0296e9a..038515a52e2c 100644
> --- a/tools/perf/util/ordered-events.c
> +++ b/tools/perf/util/ordered-events.c
> @@ -80,14 +80,20 @@ static union perf_event *dup_event(struct ordered_events *oe,
> return oe->copy_on_queue ? __dup_event(oe, event) : event;
> }
>
> -static void free_dup_event(struct ordered_events *oe, union perf_event *event)
> +static void __free_dup_event(struct ordered_events *oe, union perf_event *event)
> {
> - if (event && oe->copy_on_queue) {
> + if (event) {
> oe->cur_alloc_size -= event->header.size;
> free(event);
> }
> }
>
> +static void free_dup_event(struct ordered_events *oe, union perf_event *event)
> +{
> + if (oe->copy_on_queue)
> + __free_dup_event(oe, event);
> +}
> +
> #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct ordered_event))
> static struct ordered_event *alloc_event(struct ordered_events *oe,
> union perf_event *event)
> @@ -104,11 +110,12 @@ static struct ordered_event *alloc_event(struct ordered_events *oe,
> new = list_entry(cache->next, struct ordered_event, list);
> list_del(&new->list);
> } else if (oe->buffer) {
> - new = oe->buffer + oe->buffer_idx;
> + new = &oe->buffer->event[oe->buffer_idx];
> if (++oe->buffer_idx == MAX_SAMPLE_BUFFER)
> oe->buffer = NULL;
> } else if (oe->cur_alloc_size < oe->max_alloc_size) {
> - size_t size = MAX_SAMPLE_BUFFER * sizeof(*new);
> + size_t size = sizeof(*oe->buffer) +
> + MAX_SAMPLE_BUFFER * sizeof(*new);
>
> oe->buffer = malloc(size);
> if (!oe->buffer) {
> @@ -122,9 +129,8 @@ static struct ordered_event *alloc_event(struct ordered_events *oe,
> oe->cur_alloc_size += size;
> list_add(&oe->buffer->list, &oe->to_free);
>
> - /* First entry is abused to maintain the to_free list. */
> - oe->buffer_idx = 2;
> - new = oe->buffer + 1;
> + oe->buffer_idx = 1;
> + new = &oe->buffer->event[0];

Ok, but I think this section between the malloc() and the line above
needs some comments to clarify what is going on.
It is still hard to read.

> } else {
> pr("allocation limit reached %" PRIu64 "B\n", oe->max_alloc_size);
> }
> @@ -300,15 +306,27 @@ void ordered_events__init(struct ordered_events *oe, ordered_events__deliver_t d
> oe->deliver = deliver;
> }
>
> +static void
> +ordered_events_buffer__free(struct ordered_events_buffer *buffer,
> + struct ordered_events *oe)
> +{
> + if (oe->copy_on_queue) {
> + unsigned int i;
> +
> + for (i = 0; i < MAX_SAMPLE_BUFFER; i++)
> + __free_dup_event(oe, buffer->event[i].event);
> + }
> +
I have a problem with this one, given that the buffer->event[] is
never actually zeroed.
So what happens if you do not use all the entries by the time you have to free?
I think one way to avoid this is by iterating only all the way to
oe->buffer_idx.

> + free(buffer);
> +}
> +
> void ordered_events__free(struct ordered_events *oe)
> {
> - while (!list_empty(&oe->to_free)) {
> - struct ordered_event *event;
> + struct ordered_events_buffer *buffer, *tmp;
>
> - event = list_entry(oe->to_free.next, struct ordered_event, list);
> - list_del(&event->list);
> - free_dup_event(oe, event->event);
> - free(event);
> + list_for_each_entry_safe(buffer, tmp, &oe->to_free, list) {
> + list_del(&buffer->list);
> + ordered_events_buffer__free(buffer, oe);
> }
> }
>
> diff --git a/tools/perf/util/ordered-events.h b/tools/perf/util/ordered-events.h
> index 8c7a2948593e..1338d5c345dc 100644
> --- a/tools/perf/util/ordered-events.h
> +++ b/tools/perf/util/ordered-events.h
> @@ -25,23 +25,28 @@ struct ordered_events;
> typedef int (*ordered_events__deliver_t)(struct ordered_events *oe,
> struct ordered_event *event);
>
> +struct ordered_events_buffer {
> + struct list_head list;
> + struct ordered_event event[0];
> +};
> +
> struct ordered_events {
> - u64 last_flush;
> - u64 next_flush;
> - u64 max_timestamp;
> - u64 max_alloc_size;
> - u64 cur_alloc_size;
> - struct list_head events;
> - struct list_head cache;
> - struct list_head to_free;
> - struct ordered_event *buffer;
> - struct ordered_event *last;
> - ordered_events__deliver_t deliver;
> - int buffer_idx;
> - unsigned int nr_events;
> - enum oe_flush last_flush_type;
> - u32 nr_unordered_events;
> - bool copy_on_queue;
> + u64 last_flush;
> + u64 next_flush;
> + u64 max_timestamp;
> + u64 max_alloc_size;
> + u64 cur_alloc_size;
> + struct list_head events;
> + struct list_head cache;
> + struct list_head to_free;
> + struct ordered_events_buffer *buffer;
> + struct ordered_event *last;
> + ordered_events__deliver_t deliver;
> + int buffer_idx;
> + unsigned int nr_events;
> + enum oe_flush last_flush_type;
> + u32 nr_unordered_events;
> + bool copy_on_queue;
> };
>
> int ordered_events__queue(struct ordered_events *oe, union perf_event *event,
> --
> 2.17.1
>