kernel: 6.6: prevent garbage-cleanup of BPF kfuncs
authorTony Ambardar <itugrok@yahoo.com>
Mon, 10 Jun 2024 09:17:40 +0000 (02:17 -0700)
committerChristian Marangi <ansuelsmth@gmail.com>
Thu, 13 Jun 2024 12:16:34 +0000 (14:16 +0200)
Backport an upstream Linux fix to prevent BPF kfunc functions from being
removed during linker optimization when building BTF-enabled vmlinux.

The telltale sign of this occurring is build log warnings (e.g. mips64el):

      BTFIDS  vmlinux
    WARN: resolve_btfids: unresolved symbol bpf_verify_pkcs7_signature
    WARN: resolve_btfids: unresolved symbol bpf_lookup_user_key
    WARN: resolve_btfids: unresolved symbol bpf_lookup_system_key
    WARN: resolve_btfids: unresolved symbol bpf_key_put
    WARN: resolve_btfids: unresolved symbol bpf_iter_task_next
    WARN: resolve_btfids: unresolved symbol bpf_iter_css_task_new
    WARN: resolve_btfids: unresolved symbol bpf_get_file_xattr
    WARN: resolve_btfids: unresolved symbol bpf_ct_insert_entry
      NM      System.map
      SORTTAB vmlinux

Link: https://lore.kernel.org/bpf/ZlkoM6%2FPSxVcGM6X@kodidev-ubuntu/
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
Link: https://github.com/openwrt/openwrt/pull/15697
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
target/linux/generic/backport-6.6/065-v6.10-compiler_types.h-Define-__retain.patch [new file with mode: 0644]
target/linux/generic/backport-6.6/066-v6.10-bpf-Harden-__bpf_kfunc-against-linker-removal.patch [new file with mode: 0644]

diff --git a/target/linux/generic/backport-6.6/065-v6.10-compiler_types.h-Define-__retain.patch b/target/linux/generic/backport-6.6/065-v6.10-compiler_types.h-Define-__retain.patch
new file mode 100644 (file)
index 0000000..862207f
--- /dev/null
@@ -0,0 +1,58 @@
+From 443df175be581618d6ff781dc3af3aa1a9ba789d Mon Sep 17 00:00:00 2001
+From: Tony Ambardar <Tony.Ambardar@gmail.com>
+Date: Fri, 31 May 2024 23:55:55 -0700
+Subject: [PATCH 1/2] compiler_types.h: Define __retain for
+ __attribute__((__retain__))
+
+Some code includes the __used macro to prevent functions and data from
+being optimized out. This macro implements __attribute__((__used__)), which
+operates at the compiler and IR-level, and so still allows a linker to
+remove objects intended to be kept.
+
+Compilers supporting __attribute__((__retain__)) can address this gap by
+setting the flag SHF_GNU_RETAIN on the section of a function/variable,
+indicating to the linker the object should be retained. This attribute is
+available since gcc 11, clang 13, and binutils 2.36.
+
+Provide a __retain macro implementing __attribute__((__retain__)), whose
+first user will be the '__bpf_kfunc' tag.
+
+Link: https://lore.kernel.org/bpf/ZlmGoT9KiYLZd91S@krava/T/
+Cc: stable@vger.kernel.org # v6.6+
+Signed-off-by: Tony Ambardar <Tony.Ambardar@gmail.com>
+---
+ include/linux/compiler_types.h | 23 +++++++++++++++++++++++
+ 1 file changed, 23 insertions(+)
+
+--- a/include/linux/compiler_types.h
++++ b/include/linux/compiler_types.h
+@@ -145,6 +145,29 @@ static inline void __chk_io_ptr(const vo
+ #define __has_builtin(x) (0)
+ #endif
++/*
++ * Annotating a function/variable with __retain tells the compiler to place
++ * the object in its own section and set the flag SHF_GNU_RETAIN. This flag
++ * instructs the linker to retain the object during garbage-cleanup or LTO
++ * phases.
++ *
++ * Note that the __used macro is also used to prevent functions or data
++ * being optimized out, but operates at the compiler/IR-level and may still
++ * allow unintended removal of objects during linking.
++ *
++ * Optional: only supported since gcc >= 11, clang >= 13
++ *
++ *   gcc: https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html#index-retain-function-attribute
++ * clang: https://clang.llvm.org/docs/AttributeReference.html#retain
++ */
++#if __has_attribute(__retain__) && \
++      (defined(CONFIG_LD_DEAD_CODE_DATA_ELIMINATION) || \
++       defined(CONFIG_LTO_CLANG))
++# define __retain                     __attribute__((__retain__))
++#else
++# define __retain
++#endif
++
+ /* Compiler specific macros. */
+ #ifdef __clang__
+ #include <linux/compiler-clang.h>
diff --git a/target/linux/generic/backport-6.6/066-v6.10-bpf-Harden-__bpf_kfunc-against-linker-removal.patch b/target/linux/generic/backport-6.6/066-v6.10-bpf-Harden-__bpf_kfunc-against-linker-removal.patch
new file mode 100644 (file)
index 0000000..f7e4f43
--- /dev/null
@@ -0,0 +1,65 @@
+From ac507ed9882fd91a94657d68fe9ceac04b957103 Mon Sep 17 00:00:00 2001
+From: Tony Ambardar <Tony.Ambardar@gmail.com>
+Date: Sat, 1 Jun 2024 00:00:21 -0700
+Subject: [PATCH 2/2] bpf: Harden __bpf_kfunc tag against linker kfunc removal
+
+BPF kfuncs are often not directly referenced and may be inadvertently
+removed by optimization steps during kernel builds, thus the __bpf_kfunc
+tag mitigates against this removal by including the __used macro. However,
+this macro alone does not prevent removal during linking, and may still
+yield build warnings (e.g. on mips64el):
+
+    LD      vmlinux
+    BTFIDS  vmlinux
+  WARN: resolve_btfids: unresolved symbol bpf_verify_pkcs7_signature
+  WARN: resolve_btfids: unresolved symbol bpf_lookup_user_key
+  WARN: resolve_btfids: unresolved symbol bpf_lookup_system_key
+  WARN: resolve_btfids: unresolved symbol bpf_key_put
+  WARN: resolve_btfids: unresolved symbol bpf_iter_task_next
+  WARN: resolve_btfids: unresolved symbol bpf_iter_css_task_new
+  WARN: resolve_btfids: unresolved symbol bpf_get_file_xattr
+  WARN: resolve_btfids: unresolved symbol bpf_ct_insert_entry
+  WARN: resolve_btfids: unresolved symbol bpf_cgroup_release
+  WARN: resolve_btfids: unresolved symbol bpf_cgroup_from_id
+  WARN: resolve_btfids: unresolved symbol bpf_cgroup_acquire
+  WARN: resolve_btfids: unresolved symbol bpf_arena_free_pages
+    NM      System.map
+    SORTTAB vmlinux
+    OBJCOPY vmlinux.32
+
+Update the __bpf_kfunc tag to better guard against linker optimization by
+including the new __retain compiler macro, which fixes the warnings above.
+
+Verify the __retain macro with readelf by checking object flags for 'R':
+
+  $ readelf -Wa kernel/trace/bpf_trace.o
+  Section Headers:
+    [Nr]  Name              Type     Address  Off  Size ES Flg Lk Inf Al
+  ...
+    [178] .text.bpf_key_put PROGBITS 00000000 6420 0050 00 AXR  0   0  8
+  ...
+  Key to Flags:
+  ...
+    R (retain), D (mbind), p (processor specific)
+
+Link: https://lore.kernel.org/bpf/ZlmGoT9KiYLZd91S@krava/T/
+Reported-by: kernel test robot <lkp@intel.com>
+Closes: https://lore.kernel.org/r/202401211357.OCX9yllM-lkp@intel.com/
+Fixes: 57e7c169cd6a ("bpf: Add __bpf_kfunc tag for marking kernel functions as kfuncs")
+Cc: stable@vger.kernel.org # v6.6+
+Signed-off-by: Tony Ambardar <Tony.Ambardar@gmail.com>
+---
+ include/linux/btf.h | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+--- a/include/linux/btf.h
++++ b/include/linux/btf.h
+@@ -81,7 +81,7 @@
+  * as to avoid issues such as the compiler inlining or eliding either a static
+  * kfunc, or a global kfunc in an LTO build.
+  */
+-#define __bpf_kfunc __used noinline
++#define __bpf_kfunc __used __retain noinline
+ /*
+  * Return the name of the passed struct, if exists, or halt the build if for