1 #ifndef _TOOLS_LINUX_RING_BUFFER_H_ 2 #define _TOOLS_LINUX_RING_BUFFER_H_ 3 4 #include <asm/barrier.h> 5 6 /* 7 * Contract with kernel for walking the perf ring buffer from 8 * user space requires the following barrier pairing (quote 9 * from kernel/events/ring_buffer.c): 10 * 11 * Since the mmap() consumer (userspace) can run on a 12 * different CPU: 13 * 14 * kernel user 15 * 16 * if (LOAD ->data_tail) { LOAD ->data_head 17 * (A) smp_rmb() (C) 18 * STORE $data LOAD $data 19 * smp_wmb() (B) smp_mb() (D) 20 * STORE ->data_head STORE ->data_tail 21 * } 22 * 23 * Where A pairs with D, and B pairs with C. 24 * 25 * In our case A is a control dependency that separates the 26 * load of the ->data_tail and the stores of $data. In case 27 * ->data_tail indicates there is no room in the buffer to 28 * store $data we do not. 29 * 30 * D needs to be a full barrier since it separates the data 31 * READ from the tail WRITE. 32 * 33 * For B a WMB is sufficient since it separates two WRITEs, 34 * and for C an RMB is sufficient since it separates two READs. 35 * 36 * Note, instead of B, C, D we could also use smp_store_release() 37 * in B and D as well as smp_load_acquire() in C. 38 * 39 * However, this optimization does not make sense for all kernel 40 * supported architectures since for a fair number it would 41 * resolve into READ_ONCE() + smp_mb() pair for smp_load_acquire(), 42 * and smp_mb() + WRITE_ONCE() pair for smp_store_release(). 43 * 44 * Thus for those smp_wmb() in B and smp_rmb() in C would still 45 * be less expensive. For the case of D this has either the same 46 * cost or is less expensive, for example, due to TSO x86 can 47 * avoid the CPU barrier entirely. 48 */ 49 50 static inline u64 ring_buffer_read_head(struct perf_event_mmap_page *base) 51 { 52 /* 53 * Architectures where smp_load_acquire() does not fallback to 54 * READ_ONCE() + smp_mb() pair. 55 */ 56 #if defined(__x86_64__) || defined(__aarch64__) || defined(__powerpc64__) || \ 57 defined(__ia64__) || defined(__sparc__) && defined(__arch64__) 58 return smp_load_acquire(&base->data_head); 59 #else 60 u64 head = READ_ONCE(base->data_head); 61 62 smp_rmb(); 63 return head; 64 #endif 65 } 66 67 static inline void ring_buffer_write_tail(struct perf_event_mmap_page *base, 68 u64 tail) 69 { 70 smp_store_release(&base->data_tail, tail); 71 } 72 73 #endif /* _TOOLS_LINUX_RING_BUFFER_H_ */ 74