[v4,03/53] libperf: Lazily allocate mmap event copy

  The event copy in the mmap is used to have storage to a read
event. Not all users of mmaps read the events, such as perf record, so
switch the allocation to being on first read rather than being
embedded within the perf_mmap.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/lib/perf/include/internal/mmap.h | 2 +-
 tools/lib/perf/mmap.c                  | 9 +++++++++
 2 files changed, 10 insertions(+), 1 deletion(-)
  

Hi, 

On Thu, Nov 02, 2023 at 10:56:45AM -0700, Ian Rogers wrote:
> The event copy in the mmap is used to have storage to a read
> event. Not all users of mmaps read the events, such as perf record, so
> switch the allocation to being on first read rather than being
> embedded within the perf_mmap.
> 
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/lib/perf/include/internal/mmap.h | 2 +-
>  tools/lib/perf/mmap.c                  | 9 +++++++++
>  2 files changed, 10 insertions(+), 1 deletion(-)
> 
> diff --git a/tools/lib/perf/include/internal/mmap.h b/tools/lib/perf/include/internal/mmap.h
> index 5a062af8e9d8..b11aaf5ed645 100644
> --- a/tools/lib/perf/include/internal/mmap.h
> +++ b/tools/lib/perf/include/internal/mmap.h
> @@ -33,7 +33,7 @@ struct perf_mmap {
>  	bool			 overwrite;
>  	u64			 flush;
>  	libperf_unmap_cb_t	 unmap_cb;
> -	char			 event_copy[PERF_SAMPLE_MAX_SIZE] __aligned(8);
> +	void			*event_copy;
>  	struct perf_mmap	*next;
>  };
>  
> diff --git a/tools/lib/perf/mmap.c b/tools/lib/perf/mmap.c
> index 2184814b37dd..91ae46aac378 100644
> --- a/tools/lib/perf/mmap.c
> +++ b/tools/lib/perf/mmap.c
> @@ -51,6 +51,8 @@ int perf_mmap__mmap(struct perf_mmap *map, struct perf_mmap_param *mp,
>  
>  void perf_mmap__munmap(struct perf_mmap *map)
>  {
> +	free(map->event_copy);
> +	map->event_copy = NULL;
>  	if (map && map->base != NULL) {

If map can be NULL as the if statement above suggests, then there is a
potential a null pointer dereference bug here. Suggestion:

    if (!map)
        return;

    free(map->event_copy);
    map->event_copy = NULL;
    if (map->base != NULL) {

    ...

Cheers,
-Guilherme

>  		munmap(map->base, perf_mmap__mmap_len(map));
>  		map->base = NULL;
> @@ -226,6 +228,13 @@ static union perf_event *perf_mmap__read(struct perf_mmap *map,
>  			unsigned int len = min(sizeof(*event), size), cpy;
>  			void *dst = map->event_copy;
>  
> +			if (!dst) {
> +				dst = malloc(PERF_SAMPLE_MAX_SIZE);
> +				if (!dst)
> +					return NULL;
> +				map->event_copy = dst;
> +			}
> +
>  			do {
>  				cpy = min(map->mask + 1 - (offset & map->mask), len);
>  				memcpy(dst, &data[offset & map->mask], cpy);
> -- 
> 2.42.0.869.gea05f2083d-goog
> 
>
  

On Fri, Nov 3, 2023 at 1:33 AM Guilherme Amadio <amadio@gentoo.org> wrote:
>
> Hi,
>
> On Thu, Nov 02, 2023 at 10:56:45AM -0700, Ian Rogers wrote:
> > The event copy in the mmap is used to have storage to a read
> > event. Not all users of mmaps read the events, such as perf record, so
> > switch the allocation to being on first read rather than being
> > embedded within the perf_mmap.
> >
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> >  tools/lib/perf/include/internal/mmap.h | 2 +-
> >  tools/lib/perf/mmap.c                  | 9 +++++++++
> >  2 files changed, 10 insertions(+), 1 deletion(-)
> >
> > diff --git a/tools/lib/perf/include/internal/mmap.h b/tools/lib/perf/include/internal/mmap.h
> > index 5a062af8e9d8..b11aaf5ed645 100644
> > --- a/tools/lib/perf/include/internal/mmap.h
> > +++ b/tools/lib/perf/include/internal/mmap.h
> > @@ -33,7 +33,7 @@ struct perf_mmap {
> >       bool                     overwrite;
> >       u64                      flush;
> >       libperf_unmap_cb_t       unmap_cb;
> > -     char                     event_copy[PERF_SAMPLE_MAX_SIZE] __aligned(8);
> > +     void                    *event_copy;
> >       struct perf_mmap        *next;
> >  };
> >
> > diff --git a/tools/lib/perf/mmap.c b/tools/lib/perf/mmap.c
> > index 2184814b37dd..91ae46aac378 100644
> > --- a/tools/lib/perf/mmap.c
> > +++ b/tools/lib/perf/mmap.c
> > @@ -51,6 +51,8 @@ int perf_mmap__mmap(struct perf_mmap *map, struct perf_mmap_param *mp,
> >
> >  void perf_mmap__munmap(struct perf_mmap *map)
> >  {
> > +     free(map->event_copy);
> > +     map->event_copy = NULL;
> >       if (map && map->base != NULL) {
>
> If map can be NULL as the if statement above suggests, then there is a
> potential a null pointer dereference bug here. Suggestion:
>
>     if (!map)
>         return;
>
>     free(map->event_copy);
>     map->event_copy = NULL;
>     if (map->base != NULL) {
>
>     ...

Makes sense, will fix in v5. Waiting to get additional feedback to
avoid too much email.

Thanks,
Ian

> Cheers,
> -Guilherme
>
> >               munmap(map->base, perf_mmap__mmap_len(map));
> >               map->base = NULL;
> > @@ -226,6 +228,13 @@ static union perf_event *perf_mmap__read(struct perf_mmap *map,
> >                       unsigned int len = min(sizeof(*event), size), cpy;
> >                       void *dst = map->event_copy;
> >
> > +                     if (!dst) {
> > +                             dst = malloc(PERF_SAMPLE_MAX_SIZE);
> > +                             if (!dst)
> > +                                     return NULL;
> > +                             map->event_copy = dst;
> > +                     }
> > +
> >                       do {
> >                               cpy = min(map->mask + 1 - (offset & map->mask), len);
> >                               memcpy(dst, &data[offset & map->mask], cpy);
> > --
> > 2.42.0.869.gea05f2083d-goog
> >
> >
  

On Fri, Nov 3, 2023 at 8:49 AM Ian Rogers <irogers@google.com> wrote:
>
> On Fri, Nov 3, 2023 at 1:33 AM Guilherme Amadio <amadio@gentoo.org> wrote:
> >
> > Hi,
> >
> > On Thu, Nov 02, 2023 at 10:56:45AM -0700, Ian Rogers wrote:
> > > The event copy in the mmap is used to have storage to a read
> > > event. Not all users of mmaps read the events, such as perf record, so
> > > switch the allocation to being on first read rather than being
> > > embedded within the perf_mmap.
> > >
> > > Signed-off-by: Ian Rogers <irogers@google.com>
> > > ---
> > >  tools/lib/perf/include/internal/mmap.h | 2 +-
> > >  tools/lib/perf/mmap.c                  | 9 +++++++++
> > >  2 files changed, 10 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/tools/lib/perf/include/internal/mmap.h b/tools/lib/perf/include/internal/mmap.h
> > > index 5a062af8e9d8..b11aaf5ed645 100644
> > > --- a/tools/lib/perf/include/internal/mmap.h
> > > +++ b/tools/lib/perf/include/internal/mmap.h
> > > @@ -33,7 +33,7 @@ struct perf_mmap {
> > >       bool                     overwrite;
> > >       u64                      flush;
> > >       libperf_unmap_cb_t       unmap_cb;
> > > -     char                     event_copy[PERF_SAMPLE_MAX_SIZE] __aligned(8);
> > > +     void                    *event_copy;
> > >       struct perf_mmap        *next;
> > >  };
> > >
> > > diff --git a/tools/lib/perf/mmap.c b/tools/lib/perf/mmap.c
> > > index 2184814b37dd..91ae46aac378 100644
> > > --- a/tools/lib/perf/mmap.c
> > > +++ b/tools/lib/perf/mmap.c
> > > @@ -51,6 +51,8 @@ int perf_mmap__mmap(struct perf_mmap *map, struct perf_mmap_param *mp,
> > >
> > >  void perf_mmap__munmap(struct perf_mmap *map)
> > >  {
> > > +     free(map->event_copy);
> > > +     map->event_copy = NULL;
> > >       if (map && map->base != NULL) {
> >
> > If map can be NULL as the if statement above suggests, then there is a
> > potential a null pointer dereference bug here. Suggestion:
> >
> >     if (!map)
> >         return;
> >
> >     free(map->event_copy);
> >     map->event_copy = NULL;
> >     if (map->base != NULL) {
> >
> >     ...
>
> Makes sense, will fix in v5. Waiting to get additional feedback to
> avoid too much email.

Acked-by: Namhyung Kim <namhyung@kernel.org>


But I have another concern (not related to this change).

> >
> > >               munmap(map->base, perf_mmap__mmap_len(map));
> > >               map->base = NULL;
> > > @@ -226,6 +228,13 @@ static union perf_event *perf_mmap__read(struct perf_mmap *map,
> > >                       unsigned int len = min(sizeof(*event), size), cpy;

I'm not sure if it's ok to read less than the actual size, IOW
it seems to assume 'size' is smaller than sizeof(*event).
I guess it's true for most cases as union perf_event has
perf_record_mmap2 (among others) which contains a
filename array of size PATH_MAX.

But the SAMPLE record can be larger than that when it has
PERF_SAMPLE_AUX IIRC.  It'd happen only if it crossed the mmap
boundary and I'm afraid it'd corrupt the data.

Thanks,
Namhyung


> > >                       void *dst = map->event_copy;
> > >
> > > +                     if (!dst) {
> > > +                             dst = malloc(PERF_SAMPLE_MAX_SIZE);
> > > +                             if (!dst)
> > > +                                     return NULL;
> > > +                             map->event_copy = dst;
> > > +                     }
> > > +
> > >                       do {
> > >                               cpy = min(map->mask + 1 - (offset & map->mask), len);
> > >                               memcpy(dst, &data[offset & map->mask], cpy);
> > > --
> > > 2.42.0.869.gea05f2083d-goog
> > >
> > >
  

On Sun, Nov 5, 2023 at 10:12 AM Namhyung Kim <namhyung@kernel.org> wrote:
>
> On Fri, Nov 3, 2023 at 8:49 AM Ian Rogers <irogers@google.com> wrote:
> >
> > On Fri, Nov 3, 2023 at 1:33 AM Guilherme Amadio <amadio@gentoo.org> wrote:
> > >
> > > Hi,
> > >
> > > On Thu, Nov 02, 2023 at 10:56:45AM -0700, Ian Rogers wrote:
> > > > The event copy in the mmap is used to have storage to a read
> > > > event. Not all users of mmaps read the events, such as perf record, so
> > > > switch the allocation to being on first read rather than being
> > > > embedded within the perf_mmap.
> > > >
> > > > Signed-off-by: Ian Rogers <irogers@google.com>
> > > > ---
> > > >  tools/lib/perf/include/internal/mmap.h | 2 +-
> > > >  tools/lib/perf/mmap.c                  | 9 +++++++++
> > > >  2 files changed, 10 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/tools/lib/perf/include/internal/mmap.h b/tools/lib/perf/include/internal/mmap.h
> > > > index 5a062af8e9d8..b11aaf5ed645 100644
> > > > --- a/tools/lib/perf/include/internal/mmap.h
> > > > +++ b/tools/lib/perf/include/internal/mmap.h
> > > > @@ -33,7 +33,7 @@ struct perf_mmap {
> > > >       bool                     overwrite;
> > > >       u64                      flush;
> > > >       libperf_unmap_cb_t       unmap_cb;
> > > > -     char                     event_copy[PERF_SAMPLE_MAX_SIZE] __aligned(8);
> > > > +     void                    *event_copy;
> > > >       struct perf_mmap        *next;
> > > >  };
> > > >
> > > > diff --git a/tools/lib/perf/mmap.c b/tools/lib/perf/mmap.c
> > > > index 2184814b37dd..91ae46aac378 100644
> > > > --- a/tools/lib/perf/mmap.c
> > > > +++ b/tools/lib/perf/mmap.c
> > > > @@ -51,6 +51,8 @@ int perf_mmap__mmap(struct perf_mmap *map, struct perf_mmap_param *mp,
> > > >
> > > >  void perf_mmap__munmap(struct perf_mmap *map)
> > > >  {
> > > > +     free(map->event_copy);
> > > > +     map->event_copy = NULL;
> > > >       if (map && map->base != NULL) {
> > >
> > > If map can be NULL as the if statement above suggests, then there is a
> > > potential a null pointer dereference bug here. Suggestion:
> > >
> > >     if (!map)
> > >         return;
> > >
> > >     free(map->event_copy);
> > >     map->event_copy = NULL;
> > >     if (map->base != NULL) {
> > >
> > >     ...
> >
> > Makes sense, will fix in v5. Waiting to get additional feedback to
> > avoid too much email.
>
> Acked-by: Namhyung Kim <namhyung@kernel.org>
>
>
> But I have another concern (not related to this change).
>
> > >
> > > >               munmap(map->base, perf_mmap__mmap_len(map));
> > > >               map->base = NULL;
> > > > @@ -226,6 +228,13 @@ static union perf_event *perf_mmap__read(struct perf_mmap *map,
> > > >                       unsigned int len = min(sizeof(*event), size), cpy;
>
> I'm not sure if it's ok to read less than the actual size, IOW
> it seems to assume 'size' is smaller than sizeof(*event).
> I guess it's true for most cases as union perf_event has
> perf_record_mmap2 (among others) which contains a
> filename array of size PATH_MAX.
>
> But the SAMPLE record can be larger than that when it has
> PERF_SAMPLE_AUX IIRC.  It'd happen only if it crossed the mmap
> boundary and I'm afraid it'd corrupt the data.

Thanks, I was thinking this would just be a drop in change but I think
given this feedback it would be better to switch from allocating once
a PERF_SAMPLE_MAX_SIZE buffer to allocating or reallocating one based
on size. This potentially saves memory when size is less than
PERF_SAMPLE_MAX_SIZE and by removing the min calculation for the
amount copied (len) we can potentially exceed it and fix a potential
bug. I'll add this in v5.

Thanks,
Ian

> Thanks,
> Namhyung
>
>
> > > >                       void *dst = map->event_copy;
> > > >
> > > > +                     if (!dst) {
> > > > +                             dst = malloc(PERF_SAMPLE_MAX_SIZE);
> > > > +                             if (!dst)
> > > > +                                     return NULL;
> > > > +                             map->event_copy = dst;
> > > > +                     }
> > > > +
> > > >                       do {
> > > >                               cpy = min(map->mask + 1 - (offset & map->mask), len);
> > > >                               memcpy(dst, &data[offset & map->mask], cpy);
> > > > --
> > > > 2.42.0.869.gea05f2083d-goog
> > > >
> > > >