/* * Generic ring buffer * * Copyright (C) 2008 Steven Rostedt */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "trace.h" /* * The ring buffer header is special. We must manually up keep it. */ int ring_buffer_print_entry_header(struct trace_seq *s) { int ret; ret = trace_seq_printf(s, "# compressed entry header\n"); ret = trace_seq_printf(s, "\ttype_len : 5 bits\n"); ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n"); ret = trace_seq_printf(s, "\tarray : 32 bits\n"); ret = trace_seq_printf(s, "\n"); ret = trace_seq_printf(s, "\tpadding : type == %d\n", RINGBUF_TYPE_PADDING); ret = trace_seq_printf(s, "\ttime_extend : type == %d\n", RINGBUF_TYPE_TIME_EXTEND); ret = trace_seq_printf(s, "\tdata max type_len == %d\n", RINGBUF_TYPE_DATA_TYPE_LEN_MAX); return ret; } /* * The ring buffer is made up of a list of pages. A separate list of pages is * allocated for each CPU. A writer may only write to a buffer that is * associated with the CPU it is currently executing on. A reader may read * from any per cpu buffer. * * The reader is special. For each per cpu buffer, the reader has its own * reader page. When a reader has read the entire reader page, this reader * page is swapped with another page in the ring buffer. * * Now, as long as the writer is off the reader page, the reader can do what * ever it wants with that page. The writer will never write to that page * again (as long as it is out of the ring buffer). * * Here's some silly ASCII art. * * +------+ * |reader| RING BUFFER * |page | * +------+ +---+ +---+ +---+ * | |-->| |-->| | * +---+ +---+ +---+ * ^ | * | | * +---------------+ * * * +------+ * |reader| RING BUFFER * |page |------------------v * +------+ +---+ +---+ +---+ * | |-->| |-->| | * +---+ +---+ +---+ * ^ | * | | * +---------------+ * * * +------+ * |reader| RING BUFFER * |page |------------------v * +------+ +---+ +---+ +---+ * ^ | |-->| |-->| | * | +---+ +---+ +---+ * | | * | | * +------------------------------+ * * * +------+ * |buffer| RING BUFFER * |page |------------------v * +------+ +---+ +---+ +---+ * ^ | | | |-->| | * | New +---+ +---+ +---+ * | Reader------^ | * | page | * +------------------------------+ * * * After we make this swap, the reader can hand this page off to the splice * code and be done with it. It can even allocate a new page if it needs to * and swap that into the ring buffer. * * We will be using cmpxchg soon to make all this lockless. * */ /* * A fast way to enable or disable all ring buffers is to * call tracing_on or tracing_off. Turning off the ring buffers * prevents all ring buffers from being recorded to. * Turning this switch on, makes it OK to write to the * ring buffer, if the ring buffer is enabled itself. * * There's three layers that must be on in order to write * to the ring buffer. * * 1) This global flag must be set. * 2) The ring buffer must be enabled for recording. * 3) The per cpu buffer must be enabled for recording. * * In case of an anomaly, this global flag has a bit set that * will permantly disable all ring buffers. */ /* * Global flag to disable all recording to ring buffers * This has two bits: ON, DISABLED * * ON DISABLED * ---- ---------- * 0 0 : ring buffers are off * 1 0 : ring buffers are on * X 1 : ring buffers are permanently disabled */ enum { RB_BUFFERS_ON_BIT = 0, RB_BUFFERS_DISABLED_BIT = 1, }; enum { RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT, RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT, }; static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON; #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data) /** * tracing_on - enable all tracing buffers * * This function enables all tracing buffers that may have been * disabled with tracing_off. */ void tracing_on(void) { set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags); } EXPORT_SYMBOL_GPL(tracing_on); /** * tracing_off - turn off all tracing buffers * * This function stops all tracing buffers from recording data. * It does not disable any overhead the tracers themselves may * be causing. This function simply causes all recording to * the ring buffers to fail. */ void tracing_off(void) { clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags); } EXPORT_SYMBOL_GPL(tracing_off); /** * tracing_off_permanent - permanently disable ring buffers * * This function, once called, will disable all ring buffers * permanently. */ void tracing_off_permanent(void) { set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags); } /** * tracing_is_on - show state of ring buffers enabled */ int tracing_is_on(void) { return ring_buffer_flags == RB_BUFFERS_ON; } EXPORT_SYMBOL_GPL(tracing_is_on); #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array)) #define RB_ALIGNMENT 4U #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX) #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */ #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) # define RB_FORCE_8BYTE_ALIGNMENT 0 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT #else # define RB_FORCE_8BYTE_ALIGNMENT 1 # define RB_ARCH_ALIGNMENT 8U #endif /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */ #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX enum { RB_LEN_TIME_EXTEND = 8, RB_LEN_TIME_STAMP = 16, }; #define skip_time_extend(event) \ ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND)) static inline int rb_null_event(struct ring_buffer_event *event) { return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta; } static void rb_event_set_padding(struct ring_buffer_event *event) { /* padding has a NULL time_delta */ event->type_len = RINGBUF_TYPE_PADDING; event->time_delta = 0; } static unsigned rb_event_data_length(struct ring_buffer_event *event) { unsigned length; if (event->type_len) length = event->type_len * RB_ALIGNMENT; else length = event->array[0]; return length + RB_EVNT_HDR_SIZE; } /* * Return the length of the given event. Will return * the length of the time extend if the event is a * time extend. */ static inline unsigned rb_event_length(struct ring_buffer_event *event) { switch (event->type_len) { case RINGBUF_TYPE_PADDING: if (rb_null_event(event)) /* undefined */ return -1; return event->array[0] + RB_EVNT_HDR_SIZE; case RINGBUF_TYPE_TIME_EXTEND: return RB_LEN_TIME_EXTEND; case RINGBUF_TYPE_TIME_STAMP: return RB_LEN_TIME_STAMP; case RINGBUF_TYPE_DATA: return rb_event_data_length(event); default: BUG(); } /* not hit */ return 0; } /* * Return total length of time extend and data, * or just the event length for all other events. */ static inline unsigned rb_event_ts_length(struct ring_buffer_event *event) { unsigned len = 0; if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) { /* time extends include the data event after it */ len = RB_LEN_TIME_EXTEND; event = skip_time_extend(event); } return len + rb_event_length(event); } /** * ring_buffer_event_length - return the length of the event * @event: the event to get the length of * * Returns the size of the data load of a data event. * If the event is something other than a data event, it * returns the size of the event itself. With the exception * of a TIME EXTEND, where it still returns the size of the * data load of the data event after it. */ unsigned ring_buffer_event_length(struct ring_buffer_event *event) { unsigned length; if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) event = skip_time_extend(event); length = rb_event_length(event); if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX) return length; length -= RB_EVNT_HDR_SIZE; if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0])) length -= sizeof(event->array[0]); return length; } EXPORT_SYMBOL_GPL(ring_buffer_event_length); /* inline for ring buffer fast paths */ static void * rb_event_data(struct ring_buffer_event *event) { if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) event = skip_time_extend(event); BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX); /* If length is in len field, then array[0] has the data */ if (event->type_len) return (void *)&event->array[0]; /* Otherwise length is in array[0] and array[1] has the data */ return (void *)&event->array[1]; } /** * ring_buffer_event_data - return the data of the event * @event: the event to get the data from */ void *ring_buffer_event_data(struct ring_buffer_event *event) { return rb_event_data(event); } EXPORT_SYMBOL_GPL(ring_buffer_event_data); #define for_each_buffer_cpu(buffer, cpu) \ for_each_cpu(cpu, buffer->cpumask) #define TS_SHIFT 27 #define TS_MASK ((1ULL << TS_SHIFT) - 1) #define TS_DELTA_TEST (~TS_MASK) /* Flag when events were overwritten */ #define RB_MISSED_EVENTS (1 << 31) /* Missed count stored at end */ #define RB_MISSED_STORED (1 << 30) struct buffer_data_page { u64 time_stamp; /* page time stamp */ local_t commit; /* write committed index */ unsigned char data[]; /* data of buffer page */ }; /* * Note, the buffer_page list must be first. The buffer pages * are allocated in cache lines, which means that each buffer * page will be at the beginning of a cache line, and thus * the least significant bits will be zero. We use this to * add flags in the list struct pointers, to make the ring buffer * lockless. */ struct buffer_page { struct list_head list; /* list of buffer pages */ local_t write; /* index for next write */ unsigned read; /* index for next read */ local_t entries; /* entries on this page */ unsigned long real_end; /* real end of data */ struct buffer_data_page *page; /* Actual data page */ }; /* * The buffer page counters, write and entries, must be reset * atomically when crossing page boundaries. To synchronize this * update, two counters are inserted into the number. One is * the actual counter for the write position or count on the page. * * The other is a counter of updaters. Before an update happens * the update partition of the counter is incremented. This will * allow the updater to update the counter atomically. * * The counter is 20 bits, and the state data is 12. */ #define RB_WRITE_MASK 0xfffff #define RB_WRITE_INTCNT (1 << 20) static void rb_init_page(struct buffer_data_page *bpage) { local_set(&bpage->commit, 0); } /** * ring_buffer_page_len - the size of data on the page. * @page: The page to read * * Returns the amount of data on the page, including buffer page header. */ size_t ring_buffer_page_len(void *page) { return local_read(&((struct buffer_data_page *)page)->commit) + BUF_PAGE_HDR_SIZE; } /* * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing * this issue out. */ static void free_buffer_page(struct buffer_page *bpage) { free_page((unsigned long)bpage->page); kfree(bpage); } /* * We need to fit the time_stamp delta into 27 bits. */ static inline int test_time_stamp(u64 delta) { if (delta & TS_DELTA_TEST) return 1; return 0; } #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE) /* Max payload is BUF_PAGE_SIZE - header (8bytes) */ #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2)) int ring_buffer_print_page_header(struct trace_seq *s) { struct buffer_data_page field; int ret; ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t" "offset:0;\tsize:%u;\tsigned:%u;\n", (unsigned int)sizeof(field.time_stamp), (unsigned int)is_signed_type(u64)); ret = trace_seq_printf(s, "\tfield: local_t commit;\t" "offset:%u;\tsize:%u;\tsigned:%u;\n", (unsigned int)offsetof(typeof(field), commit), (unsigned int)sizeof(field.commit), (unsigned int)is_signed_type(long)); ret = trace_seq_printf(s, "\tfield: int overwrite;\t" "offset:%u;\tsize:%u;\tsigned:%u;\n", (unsigned int)offsetof(typeof(field), commit), 1, (unsigned int)is_signed_type(long)); ret = trace_seq_printf(s, "\tfield: char data;\t" "offset:%u;\tsize:%u;\tsigned:%u;\n", (unsigned int)offsetof(typeof(field), data), (unsigned int)BUF_PAGE_SIZE, (unsigned int)is_signed_type(char)); return ret; } /* * head_page == tail_page && head == tail then buffer is empty. */ struct ring_buffer_per_cpu { int cpu; atomic_t record_disabled; struct ring_buffer *buffer; raw_spinlock_t reader_lock; /* serialize readers */ arch_spinlock_t lock; struct lock_class_key lock_key; struct list_head *pages; struct buffer_page *head_page; /* read from head */ struct buffer_page *tail_page; /* write to tail */ struct buffer_page *commit_page; /* committed pages */ struct buffer_page *reader_page; unsigned long lost_events; unsigned long last_overrun; local_t entries_bytes; local_t commit_overrun; local_t overrun; local_t entries; local_t committing; local_t commits; unsigned long read; unsigned long read_bytes; u64 write_stamp; u64 read_stamp; }; struct ring_buffer { unsigned pages; unsigned flags; int cpus; atomic_t record_disabled; cpumask_var_t cpumask; struct lock_class_key *reader_lock_key; struct mutex mutex; struct ring_buffer_per_cpu **buffers; #ifdef CONFIG_HOTPLUG_CPU struct notifier_block cpu_notify; #endif u64 (*clock)(void); }; struct ring_buffer_iter { struct ring_buffer_per_cpu *cpu_buffer; unsigned long head; struct buffer_page *head_page; struct buffer_page *cache_reader_page; unsigned long cache_read; u64 read_stamp; }; /* buffer may be either ring_buffer or ring_buffer_per_cpu */ #define RB_WARN_ON(b, cond) \ ({ \ int _____ret = unlikely(cond); \ if (_____ret) { \ if (__same_type(*(b), struct ring_buffer_per_cpu)) { \ struct ring_buffer_per_cpu *__b = \ (void *)b; \ atomic_inc(&__b->buffer->record_disabled); \ } else \ atomic_inc(&b->record_disabled); \ WARN_ON(1); \ } \ _____ret; \ }) /* Up this if you want to test the TIME_EXTENTS and normalization */ #define DEBUG_SHIFT 0 static inline u64 rb_time_stamp(struct ring_buffer *buffer) { /* shift to debug/test normalization and TIME_EXTENTS */ return buffer->clock() << DEBUG_SHIFT; } u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu) { u64 time; preempt_disable_notrace(); time = rb_time_stamp(buffer); preempt_enable_no_resched_notrace(); return time; } EXPORT_SYMBOL_GPL(ring_buffer_time_stamp); void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer, int cpu, u64 *ts) { /* Just stupid testing the normalize function and deltas */ *ts >>= DEBUG_SHIFT; } EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp); /* * Making the ring buffer lockless makes things tricky. * Although writes only happen on the CPU that they are on, * and they only need to worry about interrupts. Reads can * happen on any CPU. * * The reader page is always off the ring buffer, but when the * reader finishes with a page, it needs to swap its page with * a new one from the buffer. The reader needs to take from * the head (writes go to the tail). But if a writer is in overwrite * mode and wraps, it must push the head page forward. * * Here lies the problem. * * The reader must be careful to replace only the head page, and * not another one. As described at the top of the file in the * ASCII art, the reader sets its old page to point to the next * page after head. It then sets the page after head to point to * the old reader page. But if the writer moves the head page * during this operation, the reader could end up with the tail. * * We use cmpxchg to help prevent this race. We also do something * special with the page before head. We set the LSB to 1. * * When the writer must push the page forward, it will clear the * bit that points to the head page, move the head, and then set * the bit that points to the new head page. * * We also don't want an interrupt coming in and moving the head * page on another writer. Thus we use the second LSB to catch * that too. Thus: * * head->list->prev->next bit 1 bit 0 * ------- ------- * Normal page 0 0 * Points to head page 0 1 * New head page 1 0 * * Note we can not trust the prev pointer of the head page, because: * * +----+ +-----+ +-----+ * | |------>| T |---X--->| N | * | |<------| | | | * +----+ +-----+ +-----+ * ^ ^ | * | +-----+ | | * +----------| R |----------+ | * | |<-----------+ * +-----+ * * Key: ---X--> HEAD flag set in pointer * T Tail page * R Reader page * N Next page * * (see __rb_reserve_next() to see where this happens) * * What the above shows is that the reader just swapped out * the reader page with a page in the buffer, but before it * could make the new header point back to the new page added * it was preempted by a writer. The writer moved forward onto * the new page added by the reader and is about to move forward * again. * * You can see, it is legitimate for the previous pointer of * the head (or any page) not to point back to itself. But only * temporarially. */ #define RB_PAGE_NORMAL 0UL #define RB_PAGE_HEAD 1UL #define RB_PAGE_UPDATE 2UL #define RB_FLAG_MASK 3UL /* PAGE_MOVED is not part of the mask */ #define RB_PAGE_MOVED 4UL /* * rb_list_head - remove any bit */ static struct list_head *rb_list_head(struct list_head *list) { unsigned long val = (unsigned long)list; return (struct list_head *)(val & ~RB_FLAG_MASK); } /* * rb_is_head_page - test if the given page is the head page * * Because the reader may move the head_page pointer, we can * not trust what the head page is (it may be pointing to * the reader page). But if the next page is a header page, * its flags will be non zero. */ static inline int rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *page, struct list_head *list) { unsigned long val; val = (unsigned long)list->next; if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list) return RB_PAGE_MOVED; return val & RB_FLAG_MASK; } /* * rb_is_reader_page * * The unique thing about the reader page, is that, if the * writer is ever on it, the previous pointer never points * back to the reader page. */ static int rb_is_reader_page(struct buffer_page *page) { struct list_head *list = page->list.prev; return rb_list_head(list->next) != &page->list; } /* * rb_set_list_to_head - set a list_head to be pointing to head. */ static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer, struct list_head *list) { unsigned long *ptr; ptr = (unsigned long *)&list->next; *ptr |= RB_PAGE_HEAD; *ptr &= ~RB_PAGE_UPDATE; } /* * rb_head_page_activate - sets up head page */ static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer) { struct buffer_page *head; head = cpu_buffer->head_page; if (!head) return; /* * Set the previous list pointer to have the HEAD flag. */ rb_set_list_to_head(cpu_buffer, head->list.prev); } static void rb_list_head_clear(struct list_head *list) { unsigned long *ptr = (unsigned long *)&list->next; *ptr &= ~RB_FLAG_MASK; } /* * rb_head_page_dactivate - clears head page ptr (for free list) */ static void rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer) { struct list_head *hd; /* Go through the whole list and clear any pointers found. */ rb_list_head_clear(cpu_buffer->pages); list_for_each(hd, cpu_buffer->pages) rb_list_head_clear(hd); } static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *head, struct buffer_page *prev, int old_flag, int new_flag) { struct list_head *list; unsigned long val = (unsigned long)&head->list; unsigned long ret; list = &prev->list; val &= ~RB_FLAG_MASK; ret = cmpxchg((unsigned long *)&list->next, val | old_flag, val | new_flag); /* check if the reader took the page */ if ((ret & ~RB_FLAG_MASK) != val) return RB_PAGE_MOVED; return ret & RB_FLAG_MASK; } static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *head, struct buffer_page *prev, int old_flag) { return rb_head_page_set(cpu_buffer, head, prev, old_flag, RB_PAGE_UPDATE); } static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *head, struct buffer_page *prev, int old_flag) { return rb_head_page_set(cpu_buffer, head, prev, old_flag, RB_PAGE_HEAD); } static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *head, struct buffer_page *prev, int old_flag) { return rb_head_page_set(cpu_buffer, head, prev, old_flag, RB_PAGE_NORMAL); } static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page **bpage) { struct list_head *p = rb_list_head((*bpage)->list.next); *bpage = list_entry(p, struct buffer_page, list); } static struct buffer_page * rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer) { struct buffer_page *head; struct buffer_page *page; struct list_head *list; int i; if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page)) return NULL; /* sanity check */ list = cpu_buffer->pages; if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list)) return NULL; page = head = cpu_buffer->head_page; /* * It is possible that the writer moves the header behind * where we started, and we miss in one loop. * A second loop should grab the header, but we'll do * three loops just because I'm paranoid. */ for (i = 0; i < 3; i++) { do { if (rb_is_head_page(cpu_buffer, page, page->list.prev)) { cpu_buffer->head_page = page; return page; } rb_inc_page(cpu_buffer, &page); } while (page != head); } RB_WARN_ON(cpu_buffer, 1); return NULL; } static int rb_head_page_replace(struct buffer_page *old, struct buffer_page *new) { unsigned long *ptr = (unsigned long *)&old->list.prev->next; unsigned long val; unsigned long ret; val = *ptr & ~RB_FLAG_MASK; val |= RB_PAGE_HEAD; ret = cmpxchg(ptr, val, (unsigned long)&new->list); return ret == val; } /* * rb_tail_page_update - move the tail page forward * * Returns 1 if moved tail page, 0 if someone else did. */ static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *tail_page, struct buffer_page *next_page) { struct buffer_page *old_tail; unsigned long old_entries; unsigned long old_write; int ret = 0; /* * The tail page now needs to be moved forward. * * We need to reset the tail page, but without messing * with possible erasing of data brought in by interrupts * that have moved the tail page and are currently on it. * * We add a counter to the write field to denote this. */ old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write); old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries); /* * Just make sure we have seen our old_write and synchronize * with any interrupts that come in. */ barrier(); /* * If the tail page is still the same as what we think * it is, then it is up to us to update the tail * pointer. */ if (tail_page == cpu_buffer->tail_page) { /* Zero the write counter */ unsigned long val = old_write & ~RB_WRITE_MASK; unsigned long eval = old_entries & ~RB_WRITE_MASK; /* * This will only succeed if an interrupt did * not come in and change it. In which case, we * do not want to modify it. * * We add (void) to let the compiler know that we do not care * about the return value of these functions. We use the * cmpxchg to only update if an interrupt did not already * do it for us. If the cmpxchg fails, we don't care. */ (void)local_cmpxchg(&next_page->write, old_write, val); (void)local_cmpxchg(&next_page->entries, old_entries, eval); /* * No need to worry about races with clearing out the commit. * it only can increment when a commit takes place. But that * only happens in the outer most nested commit. */ local_set(&next_page->page->commit, 0); old_tail = cmpxchg(&cpu_buffer->tail_page, tail_page, next_page); if (old_tail == tail_page) ret = 1; } return ret; } static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *bpage) { unsigned long val = (unsigned long)bpage; if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK)) return 1; return 0; } /** * rb_check_list - make sure a pointer to a list has the last bits zero */ static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer, struct list_head *list) { if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev)) return 1; if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next)) return 1; return 0; } /** * check_pages - integrity check of buffer pages * @cpu_buffer: CPU buffer with pages to test * * As a safety measure we check to make sure the data pages have not * been corrupted. */ static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer) { struct list_head *head = cpu_buffer->pages; struct buffer_page *bpage, *tmp; rb_head_page_deactivate(cpu_buffer); if (RB_WARN_ON(cpu_buffer, head->next->prev != head)) return -1; if (RB_WARN_ON(cpu_buffer, head->prev->next != head)) return -1; if (rb_check_list(cpu_buffer, head)) return -1; list_for_each_entry_safe(bpage, tmp, head, list) { if (RB_WARN_ON(cpu_buffer, bpage->list.next->prev != &bpage->list)) return -1; if (RB_WARN_ON(cpu_buffer, bpage->list.prev->next != &bpage->list)) return -1; if (rb_check_list(cpu_buffer, &bpage->list)) return -1; } rb_head_page_activate(cpu_buffer); return 0; } static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages) { struct buffer_page *bpage, *tmp; LIST_HEAD(pages); unsigned i; WARN_ON(!nr_pages); for (i = 0; i < nr_pages; i++) { struct page *page; /* * __GFP_NORETRY flag makes sure that the allocation fails * gracefully without invoking oom-killer and the system is * not destabilized. */ bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), GFP_KERNEL | __GFP_NORETRY, cpu_to_node(cpu_buffer->cpu)); if (!bpage) goto free_pages; rb_check_bpage(cpu_buffer, bpage); list_add(&bpage->list, &pages); page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), GFP_KERNEL | __GFP_NORETRY, 0); if (!page) goto free_pages; bpage->page = page_address(page); rb_init_page(bpage->page); } /* * The ring buffer page list is a circular list that does not * start and end with a list head. All page list items point to * other pages. */ cpu_buffer->pages = pages.next; list_del(&pages); rb_check_pages(cpu_buffer); return 0; free_pages: list_for_each_entry_safe(bpage, tmp, &pages, list) { list_del_init(&bpage->list); free_buffer_page(bpage); } return -ENOMEM; } static struct ring_buffer_per_cpu * rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu) { struct ring_buffer_per_cpu *cpu_buffer; struct buffer_page *bpage; struct page *page; int ret; cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()), GFP_KERNEL, cpu_to_node(cpu)); if (!cpu_buffer) return NULL; cpu_buffer->cpu = cpu; cpu_buffer->buffer = buffer; raw_spin_lock_init(&cpu_buffer->reader_lock); lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key); cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED; bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), GFP_KERNEL, cpu_to_node(cpu)); if (!bpage) goto fail_free_buffer; rb_check_bpage(cpu_buffer, bpage); cpu_buffer->reader_page = bpage; page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0); if (!page) goto fail_free_reader; bpage->page = page_address(page); rb_init_page(bpage->page); INIT_LIST_HEAD(&cpu_buffer->reader_page->list); ret = rb_allocate_pages(cpu_buffer, buffer->pages); if (ret < 0) goto fail_free_reader; cpu_buffer->head_page = list_entry(cpu_buffer->pages, struct buffer_page, list); cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page; rb_head_page_activate(cpu_buffer); return cpu_buffer; fail_free_reader: free_buffer_page(cpu_buffer->reader_page); fail_free_buffer: kfree(cpu_buffer); return NULL; } static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer) { struct list_head *head = cpu_buffer->pages; struct buffer_page *bpage, *tmp; free_buffer_page(cpu_buffer->reader_page); rb_head_page_deactivate(cpu_buffer); if (head) { list_for_each_entry_safe(bpage, tmp, head, list) { list_del_init(&bpage->list); free_buffer_page(bpage); } bpage = list_entry(head, struct buffer_page, list); free_buffer_page(bpage); } kfree(cpu_buffer); } #ifdef CONFIG_HOTPLUG_CPU static int rb_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu); #endif /** * ring_buffer_alloc - allocate a new ring_buffer * @size: the size in bytes per cpu that is needed. * @flags: attributes to set for the ring buffer. * * Currently the only flag that is available is the RB_FL_OVERWRITE * flag. This flag means that the buffer will overwrite old data * when the buffer wraps. If this flag is not set, the buffer will * drop data when the tail hits the head. */ struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags, struct lock_class_key *key) { struct ring_buffer *buffer; int bsize; int cpu; /* keep it in its own cache line */ buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()), GFP_KERNEL); if (!buffer) return NULL; if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL)) goto fail_free_buffer; buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE); buffer->flags = flags; buffer->clock = trace_clock_local; buffer->reader_lock_key = key; /* need at least two pages */ if (buffer->pages < 2) buffer->pages = 2; /* * In case of non-hotplug cpu, if the ring-buffer is allocated * in early initcall, it will not be notified of secondary cpus. * In that off case, we need to allocate for all possible cpus. */ #ifdef CONFIG_HOTPLUG_CPU get_online_cpus(); cpumask_copy(buffer->cpumask, cpu_online_mask); #else cpumask_copy(buffer->cpumask, cpu_possible_mask); #endif buffer->cpus = nr_cpu_ids; bsize = sizeof(void *) * nr_cpu_ids; buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()), GFP_KERNEL); if (!buffer->buffers) goto fail_free_cpumask; for_each_buffer_cpu(buffer, cpu) { buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, cpu); if (!buffer->buffers[cpu]) goto fail_free_buffers; } #ifdef CONFIG_HOTPLUG_CPU buffer->cpu_notify.notifier_call = rb_cpu_notify; buffer->cpu_notify.priority = 0; register_cpu_notifier(&buffer->cpu_notify); #endif put_online_cpus(); mutex_init(&buffer->mutex); return buffer; fail_free_buffers: for_each_buffer_cpu(buffer, cpu) { if (buffer->buffers[cpu]) rb_free_cpu_buffer(buffer->buffers[cpu]); } kfree(buffer->buffers); fail_free_cpumask: free_cpumask_var(buffer->cpumask); put_online_cpus(); fail_free_buffer: kfree(buffer); return NULL; } EXPORT_SYMBOL_GPL(__ring_buffer_alloc); /** * ring_buffer_free - free a ring buffer. * @buffer: the buffer to free. */ void ring_buffer_free(struct ring_buffer *buffer) { int cpu; get_online_cpus(); #ifdef CONFIG_HOTPLUG_CPU unregister_cpu_notifier(&buffer->cpu_notify); #endif for_each_buffer_cpu(buffer, cpu) rb_free_cpu_buffer(buffer->buffers[cpu]); put_online_cpus(); kfree(buffer->buffers); free_cpumask_var(buffer->cpumask); kfree(buffer); } EXPORT_SYMBOL_GPL(ring_buffer_free); void ring_buffer_set_clock(struct ring_buffer *buffer, u64 (*clock)(void)) { buffer->clock = clock; } static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer); static void rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages) { struct buffer_page *bpage; struct list_head *p; unsigned i; raw_spin_lock_irq(&cpu_buffer->reader_lock); rb_head_page_deactivate(cpu_buffer); for (i = 0; i < nr_pages; i++) { if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages))) goto out; p = cpu_buffer->pages->next; bpage = list_entry(p, struct buffer_page, list); list_del_init(&bpage->list); free_buffer_page(bpage); } if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages))) goto out; rb_reset_cpu(cpu_buffer); rb_check_pages(cpu_buffer); out: raw_spin_unlock_irq(&cpu_buffer->reader_lock); } static void rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer, struct list_head *pages, unsigned nr_pages) { struct buffer_page *bpage; struct list_head *p; unsigned i; raw_spin_lock_irq(&cpu_buffer->reader_lock); rb_head_page_deactivate(cpu_buffer); for (i = 0; i < nr_pages; i++) { if (RB_WARN_ON(cpu_buffer, list_empty(pages))) goto out; p = pages->next; bpage = list_entry(p, struct buffer_page, list); list_del_init(&bpage->list); list_add_tail(&bpage->list, cpu_buffer->pages); } rb_reset_cpu(cpu_buffer); rb_check_pages(cpu_buffer); out: raw_spin_unlock_irq(&cpu_buffer->reader_lock); } /** * ring_buffer_resize - resize the ring buffer * @buffer: the buffer to resize. * @size: the new size. * * Minimum size is 2 * BUF_PAGE_SIZE. * * Returns -1 on failure. */ int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size) { struct ring_buffer_per_cpu *cpu_buffer; unsigned nr_pages, rm_pages, new_pages; struct buffer_page *bpage, *tmp; unsigned long buffer_size; LIST_HEAD(pages); int i, cpu; /* * Always succeed at resizing a non-existent buffer: */ if (!buffer) return size; size = DIV_ROUND_UP(size, BUF_PAGE_SIZE); size *= BUF_PAGE_SIZE; buffer_size = buffer->pages * BUF_PAGE_SIZE; /* we need a minimum of two pages */ if (size < BUF_PAGE_SIZE * 2) size = BUF_PAGE_SIZE * 2; if (size == buffer_size) return size; atomic_inc(&buffer->record_disabled); /* Make sure all writers are done with this buffer. */ synchronize_sched(); mutex_lock(&buffer->mutex); get_online_cpus(); nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE); if (size < buffer_size) { /* easy case, just free pages */ if (RB_WARN_ON(buffer, nr_pages >= buffer->pages)) goto out_fail; rm_pages = buffer->pages - nr_pages; for_each_buffer_cpu(buffer, cpu) { cpu_buffer = buffer->buffers[cpu]; rb_remove_pages(cpu_buffer, rm_pages); } goto out; } /* * This is a bit more difficult. We only want to add pages * when we can allocate enough for all CPUs. We do this * by allocating all the pages and storing them on a local * link list. If we succeed in our allocation, then we * add these pages to the cpu_buffers. Otherwise we just free * them all and return -ENOMEM; */ if (RB_WARN_ON(buffer, nr_pages <= buffer->pages)) goto out_fail; new_pages = nr_pages - buffer->pages; for_each_buffer_cpu(buffer, cpu) { for (i = 0; i < new_pages; i++) { struct page *page; /* * __GFP_NORETRY flag makes sure that the allocation * fails gracefully without invoking oom-killer and * the system is not destabilized. */ bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), GFP_KERNEL | __GFP_NORETRY, cpu_to_node(cpu)); if (!bpage) goto free_pages; list_add(&bpage->list, &pages); page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL | __GFP_NORETRY, 0); if (!page) goto free_pages; bpage->page = page_address(page); rb_init_page(bpage->page); } } for_each_buffer_cpu(buffer, cpu) { cpu_buffer = buffer->buffers[cpu]; rb_insert_pages(cpu_buffer, &pages, new_pages); } if (RB_WARN_ON(buffer, !list_empty(&pages))) goto out_fail; out: buffer->pages = nr_pages; put_online_cpus(); mutex_unlock(&buffer->mutex); atomic_dec(&buffer->record_disabled); return size; free_pages: list_for_each_entry_safe(bpage, tmp, &pages, list) { list_del_init(&bpage->list); free_buffer_page(bpage); } put_online_cpus(); mutex_unlock(&buffer->mutex); atomic_dec(&buffer->record_disabled); return -ENOMEM; /* * Something went totally wrong, and we are too paranoid * to even clean up the mess. */ out_fail: put_online_cpus(); mutex_unlock(&buffer->mutex); atomic_dec(&buffer->record_disabled); return -1; } EXPORT_SYMBOL_GPL(ring_buffer_resize); void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val) { mutex_lock(&buffer->mutex); if (val) buffer->flags |= RB_FL_OVERWRITE; else buffer->flags &= ~RB_FL_OVERWRITE; mutex_unlock(&buffer->mutex); } EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite); static inline void * __rb_data_page_index(struct buffer_data_page *bpage, unsigned index) { return bpage->data + index; } static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index) { return bpage->page->data + index; } static inline struct ring_buffer_event * rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer) { return __rb_page_index(cpu_buffer->reader_page, cpu_buffer->reader_page->read); } static inline struct ring_buffer_event * rb_iter_head_event(struct ring_buffer_iter *iter) { return __rb_page_index(iter->head_page, iter->head); } static inline unsigned long rb_page_write(struct buffer_page *bpage) { return local_read(&bpage->write) & RB_WRITE_MASK; } static inline unsigned rb_page_commit(struct buffer_page *bpage) { return local_read(&bpage->page->commit); } static inline unsigned long rb_page_entries(struct buffer_page *bpage) { return local_read(&bpage->entries) & RB_WRITE_MASK; } /* Size is determined by what has been committed */ static inline unsigned rb_page_size(struct buffer_page *bpage) { return rb_page_commit(bpage); } static inline unsigned rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer) { return rb_page_commit(cpu_buffer->commit_page); } static inline unsigned rb_event_index(struct ring_buffer_event *event) { unsigned long addr = (unsigned long)event; return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE; } static inline int rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event) { unsigned long addr = (unsigned long)event; unsigned long index; index = rb_event_index(event); addr &= PAGE_MASK; return cpu_buffer->commit_page->page == (void *)addr && rb_commit_index(cpu_buffer) == index; } static void rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer) { unsigned long max_count; /* * We only race with interrupts and NMIs on this CPU. * If we own the commit event, then we can commit * all others that interrupted us, since the interruptions * are in stack format (they finish before they come * back to us). This allows us to do a simple loop to * assign the commit to the tail. */ again: max_count = cpu_buffer->buffer->pages * 100; while (cpu_buffer->commit_page != cpu_buffer->tail_page) { if (RB_WARN_ON(cpu_buffer, !(--max_count))) return; if (RB_WARN_ON(cpu_buffer, rb_is_reader_page(cpu_buffer->tail_page))) return; local_set(&cpu_buffer->commit_page->page->commit, rb_page_write(cpu_buffer->commit_page)); rb_inc_page(cpu_buffer, &cpu_buffer->commit_page); cpu_buffer->write_stamp = cpu_buffer->commit_page->page->time_stamp; /* add barrier to keep gcc from optimizing too much */ barrier(); } while (rb_commit_index(cpu_buffer) != rb_page_write(cpu_buffer->commit_page)) { local_set(&cpu_buffer->commit_page->page->commit, rb_page_write(cpu_buffer->commit_page)); RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->commit_page->page->commit) & ~RB_WRITE_MASK); barrier(); } /* again, keep gcc from optimizing */ barrier(); /* * If an interrupt came in just after the first while loop * and pushed the tail page forward, we will be left with * a dangling commit that will never go forward. */ if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page)) goto again; } static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer) { cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp; cpu_buffer->reader_page->read = 0; } static void rb_inc_iter(struct ring_buffer_iter *iter) { struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; /* * The iterator could be on the reader page (it starts there). * But the head could have moved, since the reader was * found. Check for this case and assign the iterator * to the head page instead of next. */ if (iter->head_page == cpu_buffer->reader_page) iter->head_page = rb_set_head_page(cpu_buffer); else rb_inc_page(cpu_buffer, &iter->head_page); iter->read_stamp = iter->head_page->page->time_stamp; iter->head = 0; } /* Slow path, do not inline */ static noinline struct ring_buffer_event * rb_add_time_stamp(struct ring_buffer_event *event, u64 delta) { event->type_len = RINGBUF_TYPE_TIME_EXTEND; /* Not the first event on the page? */ if (rb_event_index(event)) { event->time_delta = delta & TS_MASK; event->array[0] = delta >> TS_SHIFT; } else { /* nope, just zero it */ event->time_delta = 0; event->array[0] = 0; } return skip_time_extend(event); } /** * ring_buffer_update_event - update event type and data * @event: the even to update * @type: the type of event * @length: the size of the event field in the ring buffer * * Update the type and data fields of the event. The length * is the actual size that is written to the ring buffer, * and with this, we can determine what to place into the * data field. */ static void rb_update_event(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event, unsigned length, int add_timestamp, u64 delta) { /* Only a commit updates the timestamp */ if (unlikely(!rb_event_is_commit(cpu_buffer, event))) delta = 0; /* * If we need to add a timestamp, then we * add it to the start of the resevered space. */ if (unlikely(add_timestamp)) { event = rb_add_time_stamp(event, delta); length -= RB_LEN_TIME_EXTEND; delta = 0; } event->time_delta = delta; length -= RB_EVNT_HDR_SIZE; if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) { event->type_len = 0; event->array[0] = length; } else event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT); } /* * rb_handle_head_page - writer hit the head page * * Returns: +1 to retry page * 0 to continue * -1 on error */ static int rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *tail_page, struct buffer_page *next_page) { struct buffer_page *new_head; int entries; int type; int ret; entries = rb_page_entries(next_page); /* * The hard part is here. We need to move the head * forward, and protect against both readers on * other CPUs and writers coming in via interrupts. */ type = rb_head_page_set_update(cpu_buffer, next_page, tail_page, RB_PAGE_HEAD); /* * type can be one of four: * NORMAL - an interrupt already moved it for us * HEAD - we are the first to get here. * UPDATE - we are the interrupt interrupting * a current move. * MOVED - a reader on another CPU moved the next * pointer to its reader page. Give up * and try again. */ switch (type) { case RB_PAGE_HEAD: /* * We changed the head to UPDATE, thus * it is our responsibility to update * the counters. */ local_add(entries, &cpu_buffer->overrun); local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes); /* * The entries will be zeroed out when we move the * tail page. */ /* still more to do */ break; case RB_PAGE_UPDATE: /* * This is an interrupt that interrupt the * previous update. Still more to do. */ break; case RB_PAGE_NORMAL: /* * An interrupt came in before the update * and processed this for us. * Nothing left to do. */ return 1; case RB_PAGE_MOVED: /* * The reader is on another CPU and just did * a swap with our next_page. * Try again. */ return 1; default: RB_WARN_ON(cpu_buffer, 1); /* WTF??? */ return -1; } /* * Now that we are here, the old head pointer is * set to UPDATE. This will keep the reader from * swapping the head page with the reader page. * The reader (on another CPU) will spin till * we are finished. * * We just need to protect against interrupts * doing the job. We will set the next pointer * to HEAD. After that, we set the old pointer * to NORMAL, but only if it was HEAD before. * otherwise we are an interrupt, and only * want the outer most commit to reset it. */ new_head = next_page; rb_inc_page(cpu_buffer, &new_head); ret = rb_head_page_set_head(cpu_buffer, new_head, next_page, RB_PAGE_NORMAL); /* * Valid returns are: * HEAD - an interrupt came in and already set it. * NORMAL - One of two things: * 1) We really set it. * 2) A bunch of interrupts came in and moved * the page forward again. */ switch (ret) { case RB_PAGE_HEAD: case RB_PAGE_NORMAL: /* OK */ break; default: RB_WARN_ON(cpu_buffer, 1); return -1; } /* * It is possible that an interrupt came in, * set the head up, then more interrupts came in * and moved it again. When we get back here, * the page would have been set to NORMAL but we * just set it back to HEAD. * * How do you detect this? Well, if that happened * the tail page would have moved. */ if (ret == RB_PAGE_NORMAL) { /* * If the tail had moved passed next, then we need * to reset the pointer. */ if (cpu_buffer->tail_page != tail_page && cpu_buffer->tail_page != next_page) rb_head_page_set_normal(cpu_buffer, new_head, next_page, RB_PAGE_HEAD); } /* * If this was the outer most commit (the one that * changed the original pointer from HEAD to UPDATE), * then it is up to us to reset it to NORMAL. */ if (type == RB_PAGE_HEAD) { ret = rb_head_page_set_normal(cpu_buffer, next_page, tail_page, RB_PAGE_UPDATE); if (RB_WARN_ON(cpu_buffer, ret != RB_PAGE_UPDATE)) return -1; } return 0; } static unsigned rb_calculate_event_length(unsigned length) { struct ring_buffer_event event; /* Used only for sizeof array */ /* zero length can cause confusions */ if (!length) length = 1; if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) length += sizeof(event.array[0]); length += RB_EVNT_HDR_SIZE; length = ALIGN(length, RB_ARCH_ALIGNMENT); return length; } static inline void rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer, struct buffer_page *tail_page, unsigned long tail, unsigned long length) { struct ring_buffer_event *event; /* * Only the event that crossed the page boundary * must fill the old tail_page with padding. */ if (tail >= BUF_PAGE_SIZE) { /* * If the page was filled, then we still need * to update the real_end. Reset it to zero * and the reader will ignore it. */ if (tail == BUF_PAGE_SIZE) tail_page->real_end = 0; local_sub(length, &tail_page->write); return; } event = __rb_page_index(tail_page, tail); kmemcheck_annotate_bitfield(event, bitfield); /* account for padding bytes */ local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes); /* * Save the original length to the meta data. * This will be used by the reader to add lost event * counter. */ tail_page->real_end = tail; /* * If this event is bigger than the minimum size, then * we need to be careful that we don't subtract the * write counter enough to allow another writer to slip * in on this page. * We put in a discarded commit instead, to make sure * that this space is not used again. * * If we are less than the minimum size, we don't need to * worry about it. */ if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) { /* No room for any events */ /* Mark the rest of the page with padding */ rb_event_set_padding(event); /* Set the write back to the previous setting */ local_sub(length, &tail_page->write); return; } /* Put in a discarded event */ event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE; event->type_len = RINGBUF_TYPE_PADDING; /* time delta must be non zero */ event->time_delta = 1; /* Set write to end of buffer */ length = (tail + length) - BUF_PAGE_SIZE; local_sub(length, &tail_page->write); } /* * This is the slow path, force gcc not to inline it. */ static noinline struct ring_buffer_event * rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer, unsigned long length, unsigned long tail, struct buffer_page *tail_page, u64 ts) { struct buffer_page *commit_page = cpu_buffer->commit_page; struct ring_buffer *buffer = cpu_buffer->buffer; struct buffer_page *next_page; int ret; next_page = tail_page; rb_inc_page(cpu_buffer, &next_page); /* * If for some reason, we had an interrupt storm that made * it all the way around the buffer, bail, and warn * about it. */ if (unlikely(next_page == commit_page)) { local_inc(&cpu_buffer->commit_overrun); goto out_reset; } /* * This is where the fun begins! * * We are fighting against races between a reader that * could be on another CPU trying to swap its reader * page with the buffer head. * * We are also fighting against interrupts coming in and * moving the head or tail on us as well. * * If the next page is the head page then we have filled * the buffer, unless the commit page is still on the * reader page. */ if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) { /* * If the commit is not on the reader page, then * move the header page. */ if (!rb_is_reader_page(cpu_buffer->commit_page)) { /* * If we are not in overwrite mode, * this is easy, just stop here. */ if (!(buffer->flags & RB_FL_OVERWRITE)) goto out_reset; ret = rb_handle_head_page(cpu_buffer, tail_page, next_page); if (ret < 0) goto out_reset; if (ret) goto out_again; } else { /* * We need to be careful here too. The * commit page could still be on the reader * page. We could have a small buffer, and * have filled up the buffer with events * from interrupts and such, and wrapped. * * Note, if the tail page is also the on the * reader_page, we let it move out. */ if (unlikely((cpu_buffer->commit_page != cpu_buffer->tail_page) && (cpu_buffer->commit_page == cpu_buffer->reader_page))) { local_inc(&cpu_buffer->commit_overrun); goto out_reset; } } } ret = rb_tail_page_update(cpu_buffer, tail_page, next_page); if (ret) { /* * Nested commits always have zero deltas, so * just reread the time stamp */ ts = rb_time_stamp(buffer); next_page->page->time_stamp = ts; } out_again: rb_reset_tail(cpu_buffer, tail_page, tail, length); /* fail and let the caller try again */ return ERR_PTR(-EAGAIN); out_reset: /* reset write */ rb_reset_tail(cpu_buffer, tail_page, tail, length); return NULL; } static struct ring_buffer_event * __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer, unsigned long length, u64 ts, u64 delta, int add_timestamp) { struct buffer_page *tail_page; struct ring_buffer_event *event; unsigned long tail, write; /* * If the time delta since the last event is too big to * hold in the time field of the event, then we append a * TIME EXTEND event ahead of the data event. */ if (unlikely(add_timestamp)) length += RB_LEN_TIME_EXTEND; tail_page = cpu_buffer->tail_page; write = local_add_return(length, &tail_page->write); /* set write to only the index of the write */ write &= RB_WRITE_MASK; tail = write - length; /* See if we shot pass the end of this buffer page */ if (unlikely(write > BUF_PAGE_SIZE)) return rb_move_tail(cpu_buffer, length, tail, tail_page, ts); /* We reserved something on the buffer */ event = __rb_page_index(tail_page, tail); kmemcheck_annotate_bitfield(event, bitfield); rb_update_event(cpu_buffer, event, length, add_timestamp, delta); local_inc(&tail_page->entries); /* * If this is the first commit on the page, then update * its timestamp. */ if (!tail) tail_page->page->time_stamp = ts; /* account for these added bytes */ local_add(length, &cpu_buffer->entries_bytes); return event; } static inline int rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event) { unsigned long new_index, old_index; struct buffer_page *bpage; unsigned long index; unsigned long addr; new_index = rb_event_index(event); old_index = new_index + rb_event_ts_length(event); addr = (unsigned long)event; addr &= PAGE_MASK; bpage = cpu_buffer->tail_page; if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) { unsigned long write_mask = local_read(&bpage->write) & ~RB_WRITE_MASK; unsigned long event_length = rb_event_length(event); /* * This is on the tail page. It is possible that * a write could come in and move the tail page * and write to the next page. That is fine * because we just shorten what is on this page. */ old_index += write_mask; new_index += write_mask; index = local_cmpxchg(&bpage->write, old_index, new_index); if (index == old_index) { /* update counters */ local_sub(event_length, &cpu_buffer->entries_bytes); return 1; } } /* could not discard */ return 0; } static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer) { local_inc(&cpu_buffer->committing); local_inc(&cpu_buffer->commits); } static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer) { unsigned long commits; if (RB_WARN_ON(cpu_buffer, !local_read(&cpu_buffer->committing))) return; again: commits = local_read(&cpu_buffer->commits); /* synchronize with interrupts */ barrier(); if (local_read(&cpu_buffer->committing) == 1) rb_set_commit_to_write(cpu_buffer); local_dec(&cpu_buffer->committing); /* synchronize with interrupts */ barrier(); /* * Need to account for interrupts coming in between the * updating of the commit page and the clearing of the * committing counter. */ if (unlikely(local_read(&cpu_buffer->commits) != commits) && !local_read(&cpu_buffer->committing)) { local_inc(&cpu_buffer->committing); goto again; } } static struct ring_buffer_event * rb_reserve_next_event(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer, unsigned long length) { struct ring_buffer_event *event; u64 ts, delta; int nr_loops = 0; int add_timestamp; u64 diff; rb_start_commit(cpu_buffer); #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP /* * Due to the ability to swap a cpu buffer from a buffer * it is possible it was swapped before we committed. * (committing stops a swap). We check for it here and * if it happened, we have to fail the write. */ barrier(); if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) { local_dec(&cpu_buffer->committing); local_dec(&cpu_buffer->commits); return NULL; } #endif length = rb_calculate_event_length(length); again: add_timestamp = 0; delta = 0; /* * We allow for interrupts to reenter here and do a trace. * If one does, it will cause this original code to loop * back here. Even with heavy interrupts happening, this * should only happen a few times in a row. If this happens * 1000 times in a row, there must be either an interrupt * storm or we have something buggy. * Bail! */ if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000)) goto out_fail; ts = rb_time_stamp(cpu_buffer->buffer); diff = ts - cpu_buffer->write_stamp; /* make sure this diff is calculated here */ barrier(); /* Did the write stamp get updated already? */ if (likely(ts >= cpu_buffer->write_stamp)) { delta = diff; if (unlikely(test_time_stamp(delta))) { int local_clock_stable = 1; #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK local_clock_stable = sched_clock_stable; #endif WARN_ONCE(delta > (1ULL << 59), KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s", (unsigned long long)delta, (unsigned long long)ts, (unsigned long long)cpu_buffer->write_stamp, local_clock_stable ? "" : "If you just came from a suspend/resume,\n" "please switch to the trace global clock:\n" " echo global > /sys/kernel/debug/tracing/trace_clock\n"); add_timestamp = 1; } } event = __rb_reserve_next(cpu_buffer, length, ts, delta, add_timestamp); if (unlikely(PTR_ERR(event) == -EAGAIN)) goto again; if (!event) goto out_fail; return event; out_fail: rb_end_commit(cpu_buffer); return NULL; } #ifdef CONFIG_TRACING #define TRACE_RECURSIVE_DEPTH 16 /* Keep this code out of the fast path cache */ static noinline void trace_recursive_fail(void) { /* Disable all tracing before we do anything else */ tracing_off_permanent(); printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:" "HC[%lu]:SC[%lu]:NMI[%lu]\n", trace_recursion_buffer(), hardirq_count() >> HARDIRQ_SHIFT, softirq_count() >> SOFTIRQ_SHIFT, in_nmi()); WARN_ON_ONCE(1); } static inline int trace_recursive_lock(void) { trace_recursion_inc(); if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH)) return 0; trace_recursive_fail(); return -1; } static inline void trace_recursive_unlock(void) { WARN_ON_ONCE(!trace_recursion_buffer()); trace_recursion_dec(); } #else #define trace_recursive_lock() (0) #define trace_recursive_unlock() do { } while (0) #endif /** * ring_buffer_lock_reserve - reserve a part of the buffer * @buffer: the ring buffer to reserve from * @length: the length of the data to reserve (excluding event header) * * Returns a reseverd event on the ring buffer to copy directly to. * The user of this interface will need to get the body to write into * and can use the ring_buffer_event_data() interface. * * The length is the length of the data needed, not the event length * which also includes the event header. * * Must be paired with ring_buffer_unlock_commit, unless NULL is returned. * If NULL is returned, then nothing has been allocated or locked. */ struct ring_buffer_event * ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length) { struct ring_buffer_per_cpu *cpu_buffer; struct ring_buffer_event *event; int cpu; if (ring_buffer_flags != RB_BUFFERS_ON) return NULL; /* If we are tracing schedule, we don't want to recurse */ preempt_disable_notrace(); if (atomic_read(&buffer->record_disabled)) goto out_nocheck; if (trace_recursive_lock()) goto out_nocheck; cpu = raw_smp_processor_id(); if (!cpumask_test_cpu(cpu, buffer->cpumask)) goto out; cpu_buffer = buffer->buffers[cpu]; if (atomic_read(&cpu_buffer->record_disabled)) goto out; if (length > BUF_MAX_DATA_SIZE) goto out; event = rb_reserve_next_event(buffer, cpu_buffer, length); if (!event) goto out; return event; out: trace_recursive_unlock(); out_nocheck: preempt_enable_notrace(); return NULL; } EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve); static void rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event) { u64 delta; /* * The event first in the commit queue updates the * time stamp. */ if (rb_event_is_commit(cpu_buffer, event)) { /* * A commit event that is first on a page * updates the write timestamp with the page stamp */ if (!rb_event_index(event)) cpu_buffer->write_stamp = cpu_buffer->commit_page->page->time_stamp; else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) { delta = event->array[0]; delta <<= TS_SHIFT; delta += event->time_delta; cpu_buffer->write_stamp += delta; } else cpu_buffer->write_stamp += event->time_delta; } } static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event) { local_inc(&cpu_buffer->entries); rb_update_write_stamp(cpu_buffer, event); rb_end_commit(cpu_buffer); } /** * ring_buffer_unlock_commit - commit a reserved * @buffer: The buffer to commit to * @event: The event pointer to commit. * * This commits the data to the ring buffer, and releases any locks held. * * Must be paired with ring_buffer_lock_reserve. */ int ring_buffer_unlock_commit(struct ring_buffer *buffer, struct ring_buffer_event *event) { struct ring_buffer_per_cpu *cpu_buffer; int cpu = raw_smp_processor_id(); cpu_buffer = buffer->buffers[cpu]; rb_commit(cpu_buffer, event); trace_recursive_unlock(); preempt_enable_notrace(); return 0; } EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit); static inline void rb_event_discard(struct ring_buffer_event *event) { if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) event = skip_time_extend(event); /* array[0] holds the actual length for the discarded event */ event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE; event->type_len = RINGBUF_TYPE_PADDING; /* time delta must be non zero */ if (!event->time_delta) event->time_delta = 1; } /* * Decrement the entries to the page that an event is on. * The event does not even need to exist, only the pointer * to the page it is on. This may only be called before the commit * takes place. */ static inline void rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event) { unsigned long addr = (unsigned long)event; struct buffer_page *bpage = cpu_buffer->commit_page; struct buffer_page *start; addr &= PAGE_MASK; /* Do the likely case first */ if (likely(bpage->page == (void *)addr)) { local_dec(&bpage->entries); return; } /* * Because the commit page may be on the reader page we * start with the next page and check the end loop there. */ rb_inc_page(cpu_buffer, &bpage); start = bpage; do { if (bpage->page == (void *)addr) { local_dec(&bpage->entries); return; } rb_inc_page(cpu_buffer, &bpage); } while (bpage != start); /* commit not part of this buffer?? */ RB_WARN_ON(cpu_buffer, 1); } /** * ring_buffer_commit_discard - discard an event that has not been committed * @buffer: the ring buffer * @event: non committed event to discard * * Sometimes an event that is in the ring buffer needs to be ignored. * This function lets the user discard an event in the ring buffer * and then that event will not be read later. * * This function only works if it is called before the the item has been * committed. It will try to free the event from the ring buffer * if another event has not been added behind it. * * If another event has been added behind it, it will set the event * up as discarded, and perform the commit. * * If this function is called, do not call ring_buffer_unlock_commit on * the event. */ void ring_buffer_discard_commit(struct ring_buffer *buffer, struct ring_buffer_event *event) { struct ring_buffer_per_cpu *cpu_buffer; int cpu; /* The event is discarded regardless */ rb_event_discard(event); cpu = smp_processor_id(); cpu_buffer = buffer->buffers[cpu]; /* * This must only be called if the event has not been * committed yet. Thus we ca