On Tue, Feb 22, 2022 at 7:13 PM Kumar Kartikeya Dwivedi
<memxor@xxxxxxxxx> wrote:
>
> On Tue, Feb 22, 2022 at 12:34:05PM IST, Alexei Starovoitov wrote:
> > On Sun, Feb 20, 2022 at 07:18:06PM +0530, Kumar Kartikeya Dwivedi wrote:
> > > The changes in this patch deserve closer look, so it has been split into
> > > its own independent patch. While earlier we just had to skip two objects
> > > at most while copying in and out of map, now we have potentially many
> > > objects (at most 8 + 2 = 10, due to the BPF_MAP_VALUE_OFF_MAX limit).
> > >
> > > Hence, divide the copy_map_value function into an inlined fast path and
> > > function call to slowpath. The slowpath handles the case of > 3 offsets,
> > > while we handle the most common cases (0, 1, 2, or 3 offsets) in the
> > > inline function itself.
> > >
> > > In copy_map_value_slow, we use 11 offsets, just to make the for loop
> > > that copies the value free of edge cases for the last offset, by using
> > > map->value_size as final offset to subtract remaining area to copy from.
> > >
> > > Signed-off-by: Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx>
> > > ---
> > > include/linux/bpf.h | 43 +++++++++++++++++++++++++++++++---
> > > kernel/bpf/syscall.c | 55 ++++++++++++++++++++++++++++++++++++++++++++
> > > 2 files changed, 95 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > index ae599aaf8d4c..5d845ca02eba 100644
> > > --- a/include/linux/bpf.h
> > > +++ b/include/linux/bpf.h
> > > @@ -253,12 +253,22 @@ static inline void check_and_init_map_value(struct bpf_map *map, void *dst)
> > > memset(dst + map->spin_lock_off, 0, sizeof(struct bpf_spin_lock));
> > > if (unlikely(map_value_has_timer(map)))
> > > memset(dst + map->timer_off, 0, sizeof(struct bpf_timer));
> > > + if (unlikely(map_value_has_ptr_to_btf_id(map))) {
> > > + struct bpf_map_value_off *tab = map->ptr_off_tab;
> > > + int i;
> > > +
> > > + for (i = 0; i < tab->nr_off; i++)
> > > + *(u64 *)(dst + tab->off[i].offset) = 0;
> > > + }
> > > }
> > >
> > > +void copy_map_value_slow(struct bpf_map *map, void *dst, void *src, u32 s_off,
> > > + u32 s_sz, u32 t_off, u32 t_sz);
> > > +
> > > /* copy everything but bpf_spin_lock and bpf_timer. There could be one of each. */
> > > static inline void copy_map_value(struct bpf_map *map, void *dst, void *src)
> > > {
> > > - u32 s_off = 0, s_sz = 0, t_off = 0, t_sz = 0;
> > > + u32 s_off = 0, s_sz = 0, t_off = 0, t_sz = 0, p_off = 0, p_sz = 0;
> > >
> > > if (unlikely(map_value_has_spin_lock(map))) {
> > > s_off = map->spin_lock_off;
> > > @@ -268,13 +278,40 @@ static inline void copy_map_value(struct bpf_map *map, void *dst, void *src)
> > > t_off = map->timer_off;
> > > t_sz = sizeof(struct bpf_timer);
> > > }
> > > + /* Multiple offset case is slow, offload to function */
> > > + if (unlikely(map_value_has_ptr_to_btf_id(map))) {
> > > + struct bpf_map_value_off *tab = map->ptr_off_tab;
> > > +
> > > + /* Inline the likely common case */
> > > + if (likely(tab->nr_off == 1)) {
> > > + p_off = tab->off[0].offset;
> > > + p_sz = sizeof(u64);
> > > + } else {
> > > + copy_map_value_slow(map, dst, src, s_off, s_sz, t_off, t_sz);
> > > + return;
> > > + }
> > > + }
> > > +
> > > + if (unlikely(s_sz || t_sz || p_sz)) {
> > > + /* The order is p_off, t_off, s_off, use insertion sort */
> > >
> > > - if (unlikely(s_sz || t_sz)) {
> > > + if (t_off < p_off || !t_sz) {
> > > + swap(t_off, p_off);
> > > + swap(t_sz, p_sz);
> > > + }
> > > if (s_off < t_off || !s_sz) {
> > > swap(s_off, t_off);
> > > swap(s_sz, t_sz);
> > > + if (t_off < p_off || !t_sz) {
> > > + swap(t_off, p_off);
> > > + swap(t_sz, p_sz);
> > > + }
> > > }
> > > - memcpy(dst, src, t_off);
> > > +
> > > + memcpy(dst, src, p_off);
> > > + memcpy(dst + p_off + p_sz,
> > > + src + p_off + p_sz,
> > > + t_off - p_off - p_sz);
> > > memcpy(dst + t_off + t_sz,
> > > src + t_off + t_sz,
> > > s_off - t_off - t_sz);
> > > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > > index beb96866f34d..83d71d6912f5 100644
> > > --- a/kernel/bpf/syscall.c
> > > +++ b/kernel/bpf/syscall.c
> > > @@ -30,6 +30,7 @@
> > > #include <linux/pgtable.h>
> > > #include <linux/bpf_lsm.h>
> > > #include <linux/poll.h>
> > > +#include <linux/sort.h>
> > > #include <linux/bpf-netns.h>
> > > #include <linux/rcupdate_trace.h>
> > > #include <linux/memcontrol.h>
> > > @@ -230,6 +231,60 @@ static int bpf_map_update_value(struct bpf_map *map, struct fd f, void *key,
> > > return err;
> > > }
> > >
> > > +static int copy_map_value_cmp(const void *_a, const void *_b)
> > > +{
> > > + const u32 a = *(const u32 *)_a;
> > > + const u32 b = *(const u32 *)_b;
> > > +
> > > + /* We only need to sort based on offset */
> > > + if (a < b)
> > > + return -1;
> > > + else if (a > b)
> > > + return 1;
> > > + return 0;
> > > +}
> > > +
> > > +void copy_map_value_slow(struct bpf_map *map, void *dst, void *src, u32 s_off,
> > > + u32 s_sz, u32 t_off, u32 t_sz)
> > > +{
> > > + struct bpf_map_value_off *tab = map->ptr_off_tab; /* already set to non-NULL */
> > > + /* 3 = 2 for bpf_timer, bpf_spin_lock, 1 for map->value_size sentinel */
> > > + struct {
> > > + u32 off;
> > > + u32 sz;
> > > + } off_arr[BPF_MAP_VALUE_OFF_MAX + 3];
> > > + int i, cnt = 0;
> > > +
> > > + /* Reconsider stack usage when bumping BPF_MAP_VALUE_OFF_MAX */
> > > + BUILD_BUG_ON(sizeof(off_arr) != 88);
> > > +
> > > + for (i = 0; i < tab->nr_off; i++) {
> > > + off_arr[cnt].off = tab->off[i].offset;
> > > + off_arr[cnt++].sz = sizeof(u64);
> > > + }
> > > + if (s_sz) {
> > > + off_arr[cnt].off = s_off;
> > > + off_arr[cnt++].sz = s_sz;
> > > + }
> > > + if (t_sz) {
> > > + off_arr[cnt].off = t_off;
> > > + off_arr[cnt++].sz = t_sz;
> > > + }
> > > + off_arr[cnt].off = map->value_size;
> > > +
> > > + sort(off_arr, cnt, sizeof(off_arr[0]), copy_map_value_cmp, NULL);
> >
> > Ouch. sort every time we need to copy map value?
> > sort it once please. 88 bytes in a map are worth it.
> > Especially since "slow" version will trigger with just 2 kptrs.
> > (if I understand this correctly).
>
> Ok, also think we can reduce the size of the 88 bytes down to 55 bytes (32-bit
> off + 8-bit size), and embed it in struct map. Then the shuffling needed for
> timer and spin lock should also be gone.
We can probably make this copy_map_value_slow to be the only function.
I suspect
+ /* There is always at least one element */
+ memcpy(dst, src, off_arr[0].off);
+ /* Copy the rest, while skipping other regions */
+ for (i = 1; i < cnt; i++) {
+ u32 curr_off = off_arr[i - 1].off + off_arr[i - 1].sz;
+ u32 next_off = off_arr[i].off;
+
+ memcpy(dst + curr_off, src + curr_off, next_off - curr_off);
+ }
is faster than the dance we currently do in copy_map_value with a bunch of if-s.
