Re: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
From: Ian Rogers
Date: Tue Jun 14 2022 - 19:51:48 EST
On Tue, Jun 14, 2022 at 3:44 PM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
>
> Hi Ian,
>
> On Tue, Jun 14, 2022 at 7:34 AM Ian Rogers <irogers@xxxxxxxxxx> wrote:
> >
> > A mask encoding of a cpu map is laid out as:
> > u16 nr
> > u16 long_size
> > unsigned long mask[];
> > However, the mask may be 8-byte aligned meaning there is a 4-byte pad
> > after long_size. This means 32-bit and 64-bit builds see the mask as
> > being at different offsets. On top of this the structure is in the byte
> > data[] encoded as:
> > u16 type
> > char data[]
> > This means the mask's struct isn't the required 4 or 8 byte aligned, but
> > is offset by 2. Consequently the long reads and writes are causing
> > undefined behavior as the alignment is broken.
> >
> > Fix the mask struct by creating explicit 32 and 64-bit variants, use a
> > union to avoid data[] and casts; the struct must be packed so the
> > layout matches the existing perf.data layout. Taking an address of a
> > member of a packed struct breaks alignment so pass the packed
> > perf_record_cpu_map_data to functions, so they can access variables with
> > the right alignment.
> >
> > As the 64-bit version has 4 bytes of padding, optimizing writing to only
> > write the 32-bit version.
> >
> > Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
> > ---
> > tools/lib/perf/include/perf/event.h | 36 +++++++++++--
> > tools/perf/tests/cpumap.c | 19 ++++---
> > tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------
> > tools/perf/util/cpumap.h | 4 +-
> > tools/perf/util/session.c | 30 +++++------
> > tools/perf/util/synthetic-events.c | 34 +++++++-----
> > 6 files changed, 143 insertions(+), 60 deletions(-)
> >
> > diff --git a/tools/lib/perf/include/perf/event.h b/tools/lib/perf/include/perf/event.h
> > index e7758707cadd..d2d32589758a 100644
> > --- a/tools/lib/perf/include/perf/event.h
> > +++ b/tools/lib/perf/include/perf/event.h
> > @@ -6,6 +6,7 @@
> > #include <linux/types.h>
> > #include <linux/limits.h>
> > #include <linux/bpf.h>
> > +#include <linux/compiler.h>
> > #include <sys/types.h> /* pid_t */
> >
> > #define event_contains(obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))
> > @@ -153,20 +154,47 @@ enum {
> > PERF_CPU_MAP__MASK = 1,
> > };
> >
> > +/*
> > + * Array encoding of a perf_cpu_map where nr is the number of entries in cpu[]
> > + * and each entry is a value for a CPU in the map.
> > + */
> > struct cpu_map_entries {
> > __u16 nr;
> > __u16 cpu[];
> > };
> >
> > -struct perf_record_record_cpu_map {
> > +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */
> > +struct perf_record_mask_cpu_map32 {
> > + /* Number of mask values. */
> > __u16 nr;
> > + /* Constant 4. */
> > __u16 long_size;
> > - unsigned long mask[];
> > + /* Bitmap data. */
> > + __u32 mask[];
> > };
> >
> > -struct perf_record_cpu_map_data {
> > +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */
> > +struct perf_record_mask_cpu_map64 {
> > + /* Number of mask values. */
> > + __u16 nr;
> > + /* Constant 8. */
> > + __u16 long_size;
> > + /* Legacy padding. */
> > + char __pad[4];
> > + /* Bitmap data. */
> > + __u64 mask[];
> > +};
> > +
> > +struct __packed perf_record_cpu_map_data {
> > __u16 type;
> > - char data[];
> > + union {
> > + /* Used when type == PERF_CPU_MAP__CPUS. */
> > + struct cpu_map_entries cpus_data;
> > + /* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */
> > + struct perf_record_mask_cpu_map32 mask32_data;
> > + /* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */
> > + struct perf_record_mask_cpu_map64 mask64_data;
> > + };
> > };
>
> How about moving the 'type' to the union as well?
> This way we don't need to pack the entire struct
> and can have a common struct for 32 and 64 bit..
>
> struct cpu_map_entries {
> __u16 type;
> __u16 nr;
> __u16 cpu[];
> };
>
> struct perf_record_mask_cpu_map {
> __u16 type;
> __u16 nr;
> __u16 long_size; // still needed?
> __u16 pad;
> unsigned long mask[];
> };
>
> // changed it to union
> union perf_record_cpu_map_data {
> __u16 type;
> struct cpu_map_entries cpus_data;
> struct perf_record_mask_cpu_map mask_data;
> };
Thanks Namhyung,
Unfortunately this doesn't quite work as I want to make it so that the
existing cpu map encodings work with this change - ie, an old
perf.data should be readable in a newer perf with this change (the
range encoding will require that new perf.data files are read by
versions of perf with these changes). For this to work I need the
layout to match the existing unaligned code, I either need to make
mask bytes and memcpy or use an attribute like packed. Fwiw, this is a
little more efficient than the layout above as with long_size == 4 the
pad isn't necessary saving 2 bytes. I think with the packed approach
we can also add new unpacked variants like above, although I'd be keen
not to use a type that varies in size like long. I guess at some
future date we could remove the legacy supporting packed versions so
that packed or byte copying is unnecessary.
I could use a union as you show above, unfortunately that will need
the 'struct perf_record_mask_cpu_map32' and 'struct
perf_record_mask_cpu_map64' to be packed or to use bytes. We'd lose
one use of packed just to introduce two others. Potentially it is more
of a breaking change for users of this code via libperf.
These changes are something of a bug report along with fixes. If there
is a consensus that the right way to fix the bug is to break legacy
perf.data files then I'm happy to update the code accordingly (as you
show above).
Thanks,
Ian
> Thanks,
> Namhyung