On Tue, Jul 27, 2021 at 1:47 PM <sdf@xxxxxxxxxx> wrote:
>
> On 07/27, Andrii Nakryiko wrote:
> > On Mon, Jul 26, 2021 at 4:00 PM Stanislav Fomichev <sdf@xxxxxxxxxx> wrote:
> > >
> > > Current max cgroup storage value size is 4k (PAGE_SIZE). The other local
> > > storages accept up to 64k (BPF_LOCAL_STORAGE_MAX_VALUE_SIZE). Let's
> > align
> > > max cgroup value size with the other storages.
> > >
> > > For percpu, the max is 32k (PCPU_MIN_UNIT_SIZE) because percpu
> > > allocator is not happy about larger values.
> > >
> > > netcnt test is extended to exercise those maximum values
> > > (non-percpu max size is close to, but not real max).
> > >
> > > v3:
> > > * refine SIZEOF_BPF_LOCAL_STORAGE_ELEM comment (Yonghong Song)
> > > * anonymous struct in percpu_net_cnt & net_cnt (Yonghong Song)
> > > * reorder free (Yonghong Song)
> > >
> > > v2:
> > > * cap max_value_size instead of BUILD_BUG_ON (Martin KaFai Lau)
> > >
> > > Cc: Martin KaFai Lau <kafai@xxxxxx>
> > > Cc: Yonghong Song <yhs@xxxxxx>
> > > Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> > > ---
> > > kernel/bpf/local_storage.c | 11 +++++-
> > > tools/testing/selftests/bpf/netcnt_common.h | 38 +++++++++++++++++----
> > > tools/testing/selftests/bpf/test_netcnt.c | 17 ++++++---
> > > 3 files changed, 53 insertions(+), 13 deletions(-)
> > >
> > > diff --git a/kernel/bpf/local_storage.c b/kernel/bpf/local_storage.c
> > > index 7ed2a14dc0de..035e9e3a7132 100644
> > > --- a/kernel/bpf/local_storage.c
> > > +++ b/kernel/bpf/local_storage.c
> > > @@ -1,6 +1,7 @@
> > > //SPDX-License-Identifier: GPL-2.0
> > > #include <linux/bpf-cgroup.h>
> > > #include <linux/bpf.h>
> > > +#include <linux/bpf_local_storage.h>
> > > #include <linux/btf.h>
> > > #include <linux/bug.h>
> > > #include <linux/filter.h>
> > > @@ -283,9 +284,17 @@ static int cgroup_storage_get_next_key(struct
> > bpf_map *_map, void *key,
> > >
> > > static struct bpf_map *cgroup_storage_map_alloc(union bpf_attr *attr)
> > > {
> > > + __u32 max_value_size = BPF_LOCAL_STORAGE_MAX_VALUE_SIZE;
> > > int numa_node = bpf_map_attr_numa_node(attr);
> > > struct bpf_cgroup_storage_map *map;
> > >
> > > + /* percpu is bound by PCPU_MIN_UNIT_SIZE, non-percu
> > > + * is the same as other local storages.
> > > + */
> > > + if (attr->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
> > > + max_value_size = min_t(__u32, max_value_size,
> > > + PCPU_MIN_UNIT_SIZE);
> > > +
> > > if (attr->key_size != sizeof(struct bpf_cgroup_storage_key) &&
> > > attr->key_size != sizeof(__u64))
> > > return ERR_PTR(-EINVAL);
> > > @@ -293,7 +302,7 @@ static struct bpf_map
> > *cgroup_storage_map_alloc(union bpf_attr *attr)
> > > if (attr->value_size == 0)
> > > return ERR_PTR(-EINVAL);
> > >
> > > - if (attr->value_size > PAGE_SIZE)
> > > + if (attr->value_size > max_value_size)
> > > return ERR_PTR(-E2BIG);
> > >
> > > if (attr->map_flags & ~LOCAL_STORAGE_CREATE_FLAG_MASK ||
> > > diff --git a/tools/testing/selftests/bpf/netcnt_common.h
> > b/tools/testing/selftests/bpf/netcnt_common.h
> > > index 81084c1c2c23..87f5b97e1932 100644
> > > --- a/tools/testing/selftests/bpf/netcnt_common.h
> > > +++ b/tools/testing/selftests/bpf/netcnt_common.h
> > > @@ -6,19 +6,43 @@
> > >
> > > #define MAX_PERCPU_PACKETS 32
> > >
> > > +/* sizeof(struct bpf_local_storage_elem):
> > > + *
> > > + * It really is about 128 bytes on x86_64, but allocate more to
> > account for
> > > + * possible layout changes, different architectures, etc.
> > > + * The kernel will wrap up to PAGE_SIZE internally anyway.
> > > + */
> > > +#define SIZEOF_BPF_LOCAL_STORAGE_ELEM 256
> > > +
> > > +/* Try to estimate kernel's BPF_LOCAL_STORAGE_MAX_VALUE_SIZE: */
> > > +#define BPF_LOCAL_STORAGE_MAX_VALUE_SIZE (0xFFFF - \
> > > +
> > SIZEOF_BPF_LOCAL_STORAGE_ELEM)
> > > +
> > > +#define PCPU_MIN_UNIT_SIZE 32768
> > > +
> > > struct percpu_net_cnt {
> > > - __u64 packets;
> > > - __u64 bytes;
> > > + union {
>
> > so you have a struct with a single anonymous union inside, isn't that
> > right? Any problems with just making struct percpu_net_cnt into union
> > percpu_net_cnt?
> We'd have to s/struct/union/ everywhere in this case, not sure
> we want to add more churn? Seemed easier to do anonymous union+struct.
4 occurrences for net_cnt and another 4 for percpu_net_cnt, not much
churn (and all pretty localized). But I honestly don't care, just
wanted to note that you don't need this extra nesting.
>
> > > + struct {
> > > + __u64 packets;
> > > + __u64 bytes;
> > >
> > > - __u64 prev_ts;
> > > + __u64 prev_ts;
> > >
> > > - __u64 prev_packets;
> > > - __u64 prev_bytes;
> > > + __u64 prev_packets;
> > > + __u64 prev_bytes;
> > > + };
> > > + __u8 data[PCPU_MIN_UNIT_SIZE];
> > > + };
> > > };
> > >
> > > struct net_cnt {
> > > - __u64 packets;
> > > - __u64 bytes;
> > > + union {
>
> > similarly here
>
> > > + struct {
> > > + __u64 packets;
> > > + __u64 bytes;
> > > + };
> > > + __u8 data[BPF_LOCAL_STORAGE_MAX_VALUE_SIZE];
> > > + };
> > > };
> > >
> > > #endif
> > > diff --git a/tools/testing/selftests/bpf/test_netcnt.c
> > b/tools/testing/selftests/bpf/test_netcnt.c
> > > index a7b9a69f4fd5..372afccf2d17 100644
> > > --- a/tools/testing/selftests/bpf/test_netcnt.c
> > > +++ b/tools/testing/selftests/bpf/test_netcnt.c
> > > @@ -33,11 +33,11 @@ static int bpf_find_map(const char *test, struct
> > bpf_object *obj,
> > >
> > > int main(int argc, char **argv)
> > > {
> > > - struct percpu_net_cnt *percpu_netcnt;
> > > + struct percpu_net_cnt *percpu_netcnt = NULL;
> > > struct bpf_cgroup_storage_key key;
> > > + struct net_cnt *netcnt = NULL;
> > > int map_fd, percpu_map_fd;
> > > int error = EXIT_FAILURE;
> > > - struct net_cnt netcnt;
> > > struct bpf_object *obj;
> > > int prog_fd, cgroup_fd;
> > > unsigned long packets;
> > > @@ -52,6 +52,12 @@ int main(int argc, char **argv)
> > > goto err;
> > > }
> > >
> > > + netcnt = malloc(sizeof(*netcnt));
>
> > curious, was it too big to be just allocated on the stack? Isn't the
> > thread stack size much bigger than 64KB (at least by default)?
> I haven't tried really, I just moved it to malloc because it crossed
> some unconscious boundary for the 'stuff I allocate on the stack'.
> I can try it out if you prefer to keep it on the stack, let me know.
Yeah, if it can stay on the stack. Less thinking about freeing memory.
