Problem description: gcc happily align on 32-byte structures defined statically. Ftrace trace events and Tracepoints both statically define structures into custom sections (using the "section" attribute), to then assign these to symbols with the linker scripts to iterate the these sections as an array. However, gcc uses different alignments for these structures when they are defined statically than when they are globally visible and/or in an array. Therefore iteration on these arrays sees "holes" of padding. gcc is within its rights to increase the alignment of the statically defined structures because, normally, there should be no other accesses to them than in the local object. We are actually iterating on the generated structures as if they were an array without letting gcc knowing anything about it. This patch introduces __u64_packed_aligned and __u64_aligned to force gcc to use the u64 type alignment, up-aligning or down-aligning the target type if necessary. The memory accesses to the target structure are efficient (does not require bytewise memory accesses) and the atomic pointer update guarantees required by RCU are kept. u64 is considered as the largest type that can generate a trap for unaligned accesses (u64 on sparc32 needs to be aligned on 64-bit). Specifying both "packed" and "aligned" generates decent code (without the bytewise memory accesses generated by simply using "packed"), and forces gcc to down-align the structure alignment to the alignment of a u64 type. This down-alignment provided by "packed" guarantees that the structure alignment for the type declaration will match the alignment of the custom structure, and won't be larger than the alignment of the start of the custom section in the linker script. This alignment should be used for both structure definitions and declarations (as *both* the type and variable attribute) when using the "section" attribute to generate arrays of structures. Declarations should use the __u64_packed_aligned macro. Definitions should use the __u64_aligned macro. We cannot have one single macro for both, because the use of __u64_packed_aligned with variable definitions causes "warning: ‘__packed__’ attribute ignored" messages. This is understandable because the "aligned" variable attribute enforces a strict alignment (compared to the "aligned" type attribute which only specifies a minimum alignment, hence requiring the "packed" attribute too). This is explained by the gcc Variable Attributes documentation: "When used as part of a typedef, the aligned attribute can both increase and decrease alignment, and specifying the packed attribute will generate a warning." The type attribute takes care of appropriately iterating on the extern array, therefore, no variable attribute needs to be applied to the extern array definition. Also introduce the linker script U64_ALIGN() macro for specification of custom section alignment that matches that of __u64_packed_aligned and __u64_aligned. Signed-off-by: Mathieu Desnoyers CC: David Miller CC: Steven Rostedt CC: Frederic Weisbecker CC: Ingo Molnar --- include/asm-generic/vmlinux.lds.h | 6 +++ include/linux/align-section.h | 62 ++++++++++++++++++++++++++++++++++++++ include/linux/compiler.h | 14 ++++++++ 3 files changed, 82 insertions(+) Index: linux-2.6-lttng/include/linux/compiler.h =================================================================== --- linux-2.6-lttng.orig/include/linux/compiler.h +++ linux-2.6-lttng/include/linux/compiler.h @@ -57,6 +57,20 @@ extern void __chk_io_ptr(const volatile # include #endif +#include + +/* + * Use as variable attribute for custom section structure definition. + * It should also be applied to any static or extern definition of the + * structure that would override the definition to which the "section" + * attribute is applied, e.g. + * + * extern struct __u64_aligned custom; + * struct custom __u64_aligned __attribute__((section("__custom")) identifier; + */ +#define __u64_aligned \ + __attribute__((__aligned__(__alignof__(long long)))) + /* * Generic compiler-dependent macros required for kernel * build go below this comment. Actual compiler/compiler version Index: linux-2.6-lttng/include/asm-generic/vmlinux.lds.h =================================================================== --- linux-2.6-lttng.orig/include/asm-generic/vmlinux.lds.h +++ linux-2.6-lttng/include/asm-generic/vmlinux.lds.h @@ -69,6 +69,12 @@ */ #define STRUCT_ALIGN() . = ALIGN(32) +/* + * Align to a 8 byte boundary. For use with custom section made from structures + * declared with __u64_packed_aligned and defined with __u64_aligned. + */ +#define U64_ALIGN() . = ALIGN(8) + /* The actual configuration determine if the init/exit sections * are handled as text/data or they can be discarded (which * often happens at runtime) Index: linux-2.6-lttng/include/linux/align-section.h =================================================================== --- /dev/null +++ linux-2.6-lttng/include/linux/align-section.h @@ -0,0 +1,62 @@ +#ifndef _LINUX_ALIGN_SECTION_H +#define _LINUX_ALIGN_SECTION_H + +/* + * __u64_packed_aligned and __u64_aligned: + * + * __u64_aligned should be used as type attribute for structure definitions, and + * __u64_packed_aligned as type attribute for structure declaration when using + * the "section" attribute to generate arrays of structures. U64_ALIGN() must be + * used prior to these section definitions in the linker script. + * + * It forces the compiler to use the u64 type alignment, up-aligning or + * down-aligning the target type if necessary. The memory accesses to the target + * structure are efficient (does not require bytewise memory accesses) and the + * atomic pointer update guarantees required by RCU are kept. u64 is considered + * as the largest type that can generate a trap for unaligned accesses (u64 on + * sparc32 needs to be aligned on 64-bit). + * + * Specifying both "packed" and "aligned" generates decent code (without the + * bytewise memory accesses generated by simply using "packed"), and forces + * gcc to down-align the structure alignment to the alignment of a u64 type. + */ + +/* + * Use __u64_packed_aligned as type attribute for custom section structure + * declaration, e.g. + * + * struct custom { + * unsigned long field; + * ... + * } __u64_packed_aligned; + * + * The array can then be defined with: + * + * extern struct custom __start___custom[]; + * extern struct custom __stop___custom[]; + * + * With linking performed by the linker script: + * + * U64_ALIGN(); + * VMLINUX_SYMBOL(__start___custom) = .; + * *(__custom) + * VMLINUX_SYMBOL(__stop___custom) = .; + */ + +#define __u64_packed_aligned \ + __attribute__((__packed__, __aligned__(__alignof__(long long)))) + +/* + * Use __u64_aligned as variable attribute for custom section structure + * definition. It should also be applied to any static or extern definition of + * the structure that would override the definition to which the "section" + * attribute is applied, e.g. + * + * extern struct __u64_aligned custom; + * struct custom __u64_aligned __attribute__((section("__custom")) identifier; + */ + +#define __u64_aligned \ + __attribute__((__aligned__(__alignof__(long long)))) + +#endif /* _LINUX_ALIGN_SECTION_H */ -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/