Merge branch 'add-internal-only-bpf-per-cpu-instruction'

Andrii Nakryiko says:

====================
Add internal-only BPF per-CPU instruction

Add a new BPF instruction for resolving per-CPU memory addresses.

New instruction is a special form of BPF_ALU64 | BPF_MOV | BPF_X, with
insns->off set to BPF_ADDR_PERCPU (== -1). It resolves provided per-CPU offset
to an absolute address where per-CPU data resides for "this" CPU.

This patch set implements support for it in x86-64 BPF JIT only.

Using the new instruction, we also implement inlining for three cases:
  - bpf_get_smp_processor_id(), which allows to avoid unnecessary trivial
    function call, saving a bit of performance and also not polluting LBR
    records with unnecessary function call/return records;
  - PERCPU_ARRAY's bpf_map_lookup_elem() is completely inlined, bringing its
    performance to implementing per-CPU data structures using global variables
    in BPF (which is an awesome improvement, see benchmarks below);
  - PERCPU_HASH's bpf_map_lookup_elem() is partially inlined, just like the
    same for non-PERCPU HASH map; this still saves a bit of overhead.

To validate performance benefits, I hacked together a tiny benchmark doing
only bpf_map_lookup_elem() and incrementing the value by 1 for PERCPU_ARRAY
(arr-inc benchmark below) and PERCPU_HASH (hash-inc benchmark below) maps. To
establish a baseline, I also implemented logic similar to PERCPU_ARRAY based
on global variable array using bpf_get_smp_processor_id() to index array for
current CPU (glob-arr-inc benchmark below).

BEFORE
======
glob-arr-inc   :  163.685 ± 0.092M/s
arr-inc        :  138.096 ± 0.160M/s
hash-inc       :   66.855 ± 0.123M/s

AFTER
=====
glob-arr-inc   :  173.921 ± 0.039M/s (+6%)
arr-inc        :  170.729 ± 0.210M/s (+23.7%)
hash-inc       :   68.673 ± 0.070M/s (+2.7%)

As can be seen, PERCPU_HASH gets a modest +2.7% improvement, while global
array-based gets a nice +6% due to inlining of bpf_get_smp_processor_id().

But what's really important is that arr-inc benchmark basically catches up
with glob-arr-inc, resulting in +23.7% improvement. This means that in
practice it won't be necessary to avoid PERCPU_ARRAY anymore if performance is
critical (e.g., high-frequent stats collection, which is often a practical use
for PERCPU_ARRAY today).

v1->v2:
  - use BPF_ALU64 | BPF_MOV instruction instead of LDX (Alexei);
  - dropped the direct per-CPU memory read instruction, it can always be added
    back, if necessary;
  - guarded bpf_get_smp_processor_id() behind x86-64 check (Alexei);
  - switched all per-cpu addr casts to (unsigned long) to avoid sparse
    warnings.
====================

Link: https://lore.kernel.org/r/20240402021307.1012571-1-andrii@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>