Merge branch 'no_caller_saved_registers-attribute-for-helper-calls'

Eduard Zingerman says:

====================
no_caller_saved_registers attribute for helper calls

This patch-set seeks to allow using no_caller_saved_registers gcc/clang
attribute with some BPF helper functions (and kfuncs in the future).

As documented in [1], this attribute means that function scratches
only some of the caller saved registers defined by ABI.
For BPF the set of such registers could be defined as follows:
- R0 is scratched only if function is non-void;
- R1-R5 are scratched only if corresponding parameter type is defined
  in the function prototype.

The goal of the patch-set is to implement no_caller_saved_registers
(nocsr for short) in a backwards compatible manner:
- for kernels that support the feature, gain some performance boost
  from better register allocation;
- for kernels that don't support the feature, allow programs execution
  with minor performance losses.

To achieve this, use a scheme suggested by Alexei Starovoitov:
- for nocsr calls clang allocates registers as-if relevant r0-r5
  registers are not scratched by the call;
- as a post-processing step, clang visits each nocsr call and adds
  spill/fill for every live r0-r5;
- stack offsets used for spills/fills are allocated as lowest
  stack offsets in whole function and are not used for any other
  purpose;
- when kernel loads a program, it looks for such patterns
  (nocsr function surrounded by spills/fills) and checks if
  spill/fill stack offsets are used exclusively in nocsr patterns;
- if so, and if current JIT inlines the call to the nocsr function
  (e.g. a helper call), kernel removes unnecessary spill/fill pairs;
- when old kernel loads a program, presence of spill/fill pairs
  keeps BPF program valid, albeit slightly less efficient.

Corresponding clang/llvm changes are available in [2].

The patch-set uses bpf_get_smp_processor_id() function as a canary,
making it the first helper with nocsr attribute.

For example, consider the following program:

  #define __no_csr __attribute__((no_caller_saved_registers))
  #define SEC(name) __attribute__((section(name), used))
  #define bpf_printk(fmt, ...) bpf_trace_printk((fmt), sizeof(fmt), __VA_ARGS__)

  typedef unsigned int __u32;

  static long (* const bpf_trace_printk)(const char *fmt, __u32 fmt_size, ...) = (void *) 6;
  static __u32 (*const bpf_get_smp_processor_id)(void) __no_csr = (void *)8;

  SEC("raw_tp")
  int test(void *ctx)
  {
          __u32 task = bpf_get_smp_processor_id();
  	bpf_printk("ctx=%p, smp=%d", ctx, task);
  	return 0;
  }

  char _license[] SEC("license") = "GPL";

Compiled (using [2]) as follows:

  $ clang --target=bpf -O2 -g -c -o nocsr.bpf.o nocsr.bpf.c
  $ llvm-objdump --no-show-raw-insn -Sd nocsr.bpf.o
    ...
  3rd parameter for printk call     removable spill/fill pair
  .--- 0:       r3 = r1                             |
; |       __u32 task = bpf_get_smp_processor_id();  |
  |    1:       *(u64 *)(r10 - 0x8) = r3 <----------|
  |    2:       call 0x8                            |
  |    3:       r3 = *(u64 *)(r10 - 0x8) <----------'
; |     bpf_printk("ctx=%p, smp=%d", ctx, task);
  |    4:       r1 = 0x0 ll
  |    6:       r2 = 0xf
  |    7:       r4 = r0
  '--> 8:       call 0x6
;       return 0;
       9:       r0 = 0x0
      10:       exit

Here is how the program looks after verifier processing:

  # bpftool prog load ./nocsr.bpf.o /sys/fs/bpf/nocsr-test
  # bpftool prog dump xlated pinned /sys/fs/bpf/nocsr-test

  int test(void * ctx):
     0: (bf) r3 = r1                         <--- 3rd printk parameter
  ; __u32 task = bpf_get_smp_processor_id();
     1: (b4) w0 = 197324                     <--. inlined helper call,
     2: (bf) r0 = &(void __percpu *)(r0)     <--- spill/fill
     3: (61) r0 = *(u32 *)(r0 +0)            <--' pair removed
  ; bpf_printk("ctx=%p, smp=%d", ctx, task);
     4: (18) r1 = map[id:5][0]+0
     6: (b7) r2 = 15
     7: (bf) r4 = r0
     8: (85) call bpf_trace_printk#-125920
  ; return 0;
     9: (b7) r0 = 0
    10: (95) exit

[1] https://clang.llvm.org/docs/AttributeReference.html#no-caller-saved-registers
[2] https://github.com/eddyz87/llvm-project/tree/bpf-no-caller-saved-registers

Change list:
- v3 -> v4:
  - When nocsr spills/fills are removed in the subprogram, allow these
    spills/fills to reside in [-MAX_BPF_STACK-48..MAX_BPF_STACK) range
    (suggested by Alexei);
  - Dropped patches with special handling for bpf_probe_read_kernel()
    (requested by Alexei);
  - Reset aux .nocsr_pattern and .nocsr_spills_num fields in
    check_nocsr_stack_contract() (requested by Andrii).
    Andrii, I have not added an additional flag to
    struct bpf_subprog_info, it currently does not have holes
    and I really don't like adding a bool field there just as an
    alternative indicator that nocsr is disabled.
    Indicator at the moment:
    - nocsr_stack_off >= S16_MIN means that nocsr rewrite is enabled;
    - nocsr_stack_off == S16_MIN means that nocsr rewrite is disabled.
- v2 -> v3:
  - As suggested by Andrii, 'nocsr_stack_off' is no longer checked at
    rewrite time, instead mark_nocsr_patterns() now does two passes
    over BPF program:
    - on a first pass it computes the lowest stack spill offset for
      the subprogram;
    - on a second pass this offset is used to recognize nocsr pattern.
  - As suggested by Alexei, a new mechanic is added to work around a
    situation mentioned by Andrii, when more helper functions are
    marked as nocsr at compile time than current kernel supports:
    - all {spill*,helper call,fill*} patterns are now marked as
      insn_aux_data[*].nocsr_pattern, thus relaxing failure condition
      for check_nocsr_stack_contract();
    - spill/fill pairs are not removed for patterns where helper can't
      be inlined;
    - see mark_nocsr_pattern_for_call() for details an example.
  - As suggested by Alexei, subprogram stack depth is now adjusted
    if all spill/fill pairs could be removed. This adjustment has
    to take place before optimize_bpf_loop(), hence the rewrite
    is moved from do_misc_fixups() to remove_nocsr_spills_fills()
    (again).
  - As suggested by Andrii, special measures are taken to work around
    bpf_probe_read_kernel() access to BPF stack, see patches 11, 12.
    Patch #11 is very simplistic, a more comprehensive solution would
    be to change the type of the third parameter of the
    bpf_probe_read_kernel() from ARG_ANYTHING to something else and
    not only check nocsr contract, but also propagate stack slot
    liveness information. However, such change would require update in
    struct bpf_call_arg_meta processing, which currently implies that
    every memory parameter is followed by a size parameter.
    I can work on these changes, please comment.
  - Stylistic changes suggested by Andrii.
  - Added acks from Andrii.
  - Dropped RFC tag.
- v1 -> v2:
  - assume that functions inlined by either jit or verifier
    conform to no_caller_saved_registers contract (Andrii, Puranjay);
  - allow nocsr rewrite for bpf_get_smp_processor_id()
    on arm64 and riscv64 architectures (Puranjay);
  - __arch_{x86_64,arm64,riscv64} macro for test_loader;
  - moved remove_nocsr_spills_fills() inside do_misc_fixups() (Andrii);
  - moved nocsr pattern detection from check_cfg() to a separate pass
    (Andrii);
  - various stylistic/correctness changes according to Andrii's
    comments.

Revisions:
- v1 https://lore.kernel.org/bpf/[email protected]/
- v2 https://lore.kernel.org/bpf/[email protected]/
- v3 https://lore.kernel.org/bpf/[email protected]/
====================

Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Alexei Starovoitov <[email protected]>
Signed-off-by: Andrii Nakryiko <[email protected]>

Author: Alexei Starovoitov

include/linux/bpf.h

@@ -808,6 +808,12 @@ struct bpf_func_proto {
 	bool gpl_only;
 	bool pkt_access;
 	bool might_sleep;
+	/* set to true if helper follows contract for gcc/llvm
+	 * attribute no_caller_saved_registers:
+	 * - void functions do not scratch r0
+	 * - functions taking N arguments scratch only registers r1-rN
+	 */
+	bool allow_nocsr;
 	enum bpf_return_type ret_type;
 	union {
 		struct {

include/linux/bpf_verifier.h

@@ -576,6 +576,14 @@ struct bpf_insn_aux_data {
 	bool is_iter_next; /* bpf_iter_<type>_next() kfunc call */
 	bool call_with_percpu_alloc_ptr; /* {this,per}_cpu_ptr() with prog percpu alloc */
 	u8 alu_state; /* used in combination with alu_limit */
+	/* true if STX or LDX instruction is a part of a spill/fill
+	 * pattern for a no_caller_saved_registers call.
+	 */
+	u8 nocsr_pattern:1;
+	/* for CALL instructions, a number of spill/fill pairs in the
+	 * no_caller_saved_registers pattern.
+	 */
+	u8 nocsr_spills_num:3;
 
 	/* below fields are initialized once */
 	unsigned int orig_idx; /* original instruction index */
@@ -645,13 +653,19 @@ struct bpf_subprog_info {
 	u32 linfo_idx; /* The idx to the main_prog->aux->linfo */
 	u16 stack_depth; /* max. stack depth used by this function */
 	u16 stack_extra;
+	/* offsets in range [stack_depth .. nocsr_stack_off)
+	 * are used for no_caller_saved_registers spills and fills.
+	 */
+	s16 nocsr_stack_off;
 	bool has_tail_call: 1;
 	bool tail_call_reachable: 1;
 	bool has_ld_abs: 1;
 	bool is_cb: 1;
 	bool is_async_cb: 1;
 	bool is_exception_cb: 1;
 	bool args_cached: 1;
+	/* true if nocsr stack region is used by functions that can't be inlined */
+	bool keep_nocsr_stack: 1;
 
 	u8 arg_cnt;
 	struct bpf_subprog_arg_info args[MAX_BPF_FUNC_REG_ARGS];

kernel/bpf/helpers.c

@@ -158,6 +158,7 @@ const struct bpf_func_proto bpf_get_smp_processor_id_proto = {
 	.func		= bpf_get_smp_processor_id,
 	.gpl_only	= false,
 	.ret_type	= RET_INTEGER,
+	.allow_nocsr	= true,
 };
 
 BPF_CALL_0(bpf_get_numa_node_id)