#define _UST_BUFFERS_H
#include <assert.h>
-#include <ust/kernelcompat.h>
-#include "usterr.h"
+
+#include <ust/core.h>
+#include <ust/clock.h>
+
+#include "usterr_signal_safe.h"
#include "channels.h"
#include "tracerconst.h"
#include "tracercore.h"
-#include "header-inline.h"
-
-/***** FIXME: SHOULD BE REMOVED ***** */
/*
* BUFFER_TRUNC zeroes the subbuffer offset and the subbuffer number parts of
/**************************************/
+/*
+ * TODO: using "long" type for struct ust_buffer (control structure
+ * shared between traced apps and the consumer) is a very bad idea when
+ * we get to systems with mixed 32/64-bit processes.
+ *
+ * But on 64-bit system, we want the full power of 64-bit counters,
+ * which wraps less often. Therefore, it's not as easy as "use 32-bit
+ * types everywhere".
+ *
+ * One way to deal with this is to:
+ * 1) Design 64-bit consumer so it can detect 32-bit and 64-bit apps.
+ * 2) The 32-bit consumer only supports 32-bit apps.
+ */
+
struct commit_counters {
long cc; /* ATOMIC */
long cc_sb; /* ATOMIC - Incremented _once_ at sb switch */
int data_ready_fd_write;
/* the reading end of the pipe */
int data_ready_fd_read;
+ /*
+ * List of buffers with an open pipe, used for fork and forced subbuffer
+ * switch.
+ */
+ struct cds_list_head open_buffers_list;
unsigned int finalized;
//ust// struct timer_list switch_timer; /* timer for periodical switch */
struct ust_channel *chan;
- struct kref kref;
+ struct urcu_ref urcu_ref;
void *buf_data;
size_t buf_size;
int shmid;
unsigned int cpu;
/* commit count per subbuffer; must be at end of struct */
- long commit_seq[0] ____cacheline_aligned; /* ATOMIC */
+ long commit_seq[0]; /* ATOMIC */
} ____cacheline_aligned;
/*
- * A switch is done during tracing or as a final flush after tracing (so it
- * won't write in the new sub-buffer).
- * FIXME: make this message clearer
+ * A switch is either done during tracing (FORCE_ACTIVE) or as a final
+ * flush after tracing (with FORCE_FLUSH). FORCE_FLUSH ensures we won't
+ * write in the new sub-buffer).
*/
enum force_switch_mode { FORCE_ACTIVE, FORCE_FLUSH };
-extern int ltt_reserve_slot_lockless_slow(struct ust_trace *trace,
- struct ust_channel *ltt_channel, void **transport_data,
- size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc,
- unsigned int *rflags, int largest_align, int cpu);
+extern int ltt_reserve_slot_lockless_slow(struct ust_channel *chan,
+ struct ust_trace *trace, size_t data_size,
+ int largest_align, int cpu,
+ struct ust_buffer **ret_buf,
+ size_t *slot_size, long *buf_offset,
+ u64 *tsc, unsigned int *rflags);
extern void ltt_force_switch_lockless_slow(struct ust_buffer *buf,
enum force_switch_mode mode);
+#ifndef HAVE_EFFICIENT_UNALIGNED_ACCESS
+
+/*
+ * Calculate the offset needed to align the type.
+ * size_of_type must be non-zero.
+ */
+static inline unsigned int ltt_align(size_t align_drift, size_t size_of_type)
+{
+ size_t alignment = min(sizeof(void *), size_of_type);
+ return (alignment - align_drift) & (alignment - 1);
+}
+/* Default arch alignment */
+#define LTT_ALIGN
+
+static inline int ltt_get_alignment(void)
+{
+ return sizeof(void *);
+}
+
+#else /* HAVE_EFFICIENT_UNALIGNED_ACCESS */
+
+static inline unsigned int ltt_align(size_t align_drift,
+ size_t size_of_type)
+{
+ return 0;
+}
+
+#define LTT_ALIGN __attribute__((packed))
+
+static inline int ltt_get_alignment(void)
+{
+ return 0;
+}
+#endif /* HAVE_EFFICIENT_UNALIGNED_ACCESS */
static __inline__ void ust_buffers_do_copy(void *dest, const void *src, size_t len)
{
}
#endif
+/*
+ * ust_get_header_size
+ *
+ * Calculate alignment offset to 32-bits. This is the alignment offset of the
+ * event header.
+ *
+ * Important note :
+ * The event header must be 32-bits. The total offset calculated here :
+ *
+ * Alignment of header struct on 32 bits (min arch size, header size)
+ * + sizeof(header struct) (32-bits)
+ * + (opt) u16 (ext. event id)
+ * + (opt) u16 (event_size) (if event_size == 0xFFFFUL, has ext. event size)
+ * + (opt) u32 (ext. event size)
+ * + (opt) u64 full TSC (aligned on min(64-bits, arch size))
+ *
+ * The payload must itself determine its own alignment from the biggest type it
+ * contains.
+ * */
+static __inline__ unsigned char ust_get_header_size(
+ struct ust_channel *channel,
+ size_t offset,
+ size_t data_size,
+ size_t *before_hdr_pad,
+ unsigned int rflags)
+{
+ size_t orig_offset = offset;
+ size_t padding;
+
+ padding = ltt_align(offset, sizeof(struct ltt_event_header));
+ offset += padding;
+ offset += sizeof(struct ltt_event_header);
+
+ if(unlikely(rflags)) {
+ switch (rflags) {
+ case LTT_RFLAG_ID_SIZE_TSC:
+ offset += sizeof(u16) + sizeof(u16);
+ if (data_size >= 0xFFFFU)
+ offset += sizeof(u32);
+ offset += ltt_align(offset, sizeof(u64));
+ offset += sizeof(u64);
+ break;
+ case LTT_RFLAG_ID_SIZE:
+ offset += sizeof(u16) + sizeof(u16);
+ if (data_size >= 0xFFFFU)
+ offset += sizeof(u32);
+ break;
+ case LTT_RFLAG_ID:
+ offset += sizeof(u16);
+ break;
+ }
+ }
+
+ *before_hdr_pad = padding;
+ return offset - orig_offset;
+}
+
static __inline__ void ltt_reserve_push_reader(
struct ust_channel *rchan,
struct ust_buffer *buf,
consumed_idx = SUBBUF_INDEX(consumed_old, buf->chan);
commit_count = uatomic_read(&buf->commit_count[consumed_idx].cc_sb);
/*
- * No memory barrier here, since we are only interested
+ * No memory cmm_barrier here, since we are only interested
* in a statistically correct polling result. The next poll will
* get the data is we are racing. The mb() that ensures correct
* memory order is in get_subbuf.
*tsc = trace_clock_read64();
-//ust// #ifdef CONFIG_LTT_VMCORE
-//ust// prefetch(&buf->commit_count[SUBBUF_INDEX(*o_begin, rchan)]);
-//ust// prefetch(&buf->commit_seq[SUBBUF_INDEX(*o_begin, rchan)]);
-//ust// #else
-//ust// prefetchw(&buf->commit_count[SUBBUF_INDEX(*o_begin, rchan)]);
-//ust// #endif
if (last_tsc_overflow(buf, *tsc))
*rflags = LTT_RFLAG_ID_SIZE_TSC;
return 0;
}
-static __inline__ int ltt_reserve_slot(struct ust_trace *trace,
- struct ust_channel *chan, void **transport_data,
- size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc,
- unsigned int *rflags, int largest_align, int cpu)
+static __inline__ int ltt_reserve_slot(struct ust_channel *chan,
+ struct ust_trace *trace, size_t data_size,
+ int largest_align, int cpu,
+ struct ust_buffer **ret_buf,
+ size_t *slot_size, long *buf_offset, u64 *tsc,
+ unsigned int *rflags)
{
- struct ust_buffer *buf = chan->buf[cpu];
+ struct ust_buffer *buf = *ret_buf = chan->buf[cpu];
long o_begin, o_end, o_old;
size_t before_hdr_pad;
/*
* Perform retryable operations.
*/
- /* FIXME: make this rellay per cpu? */
- if (unlikely(LOAD_SHARED(ltt_nesting) > 4)) {
+ if (unlikely(CMM_LOAD_SHARED(ltt_nesting) > 4)) {
DBG("Dropping event because nesting is too deep.");
uatomic_inc(&buf->events_lost);
return -EPERM;
*/
ltt_reserve_push_reader(chan, buf, o_end - 1);
- /*
- * Clear noref flag for this subbuffer.
- */
-//ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(o_end - 1, chan));
-
*buf_offset = o_begin + before_hdr_pad;
return 0;
slow_path:
- return ltt_reserve_slot_lockless_slow(trace, chan,
- transport_data, data_size, slot_size, buf_offset, tsc,
- rflags, largest_align, cpu);
+ return ltt_reserve_slot_lockless_slow(chan, trace, data_size,
+ largest_align, cpu, ret_buf,
+ slot_size, buf_offset, tsc,
+ rflags);
}
/*
* commit count reaches back the reserve offset (module subbuffer size). It is
* useful for crash dump.
*/
-//ust// #ifdef CONFIG_LTT_VMCORE
static __inline__ void ltt_write_commit_counter(struct ust_channel *chan,
struct ust_buffer *buf, long idx, long buf_offset,
long commit_count, size_t data_size)
DBG("commit_seq for channel %s_%d, subbuf %ld is now %ld", buf->chan->channel_name, buf->cpu, idx, commit_count);
}
-//ust// #else
-//ust// static __inline__ void ltt_write_commit_counter(struct ust_buffer *buf,
-//ust// long idx, long buf_offset, long commit_count, size_t data_size)
-//ust// {
-//ust// }
-//ust// #endif
/*
* Atomic unordered slot commit. Increments the commit count in the
long endidx = SUBBUF_INDEX(offset_end - 1, chan);
long commit_count;
-#ifdef LTT_NO_IPI_BARRIER
- smp_wmb();
-#else
- /*
- * Must write slot data before incrementing commit count.
- * This compiler barrier is upgraded into a smp_mb() by the IPI
- * sent by get_subbuf().
- */
- barrier();
-#endif
+ cmm_smp_wmb();
+
uatomic_add(&buf->commit_count[endidx].cc, slot_size);
/*
* commit count read can race with concurrent OOO commit count updates.
ltt_write_commit_counter(chan, buf, endidx, buf_offset, commit_count, data_size);
}
-void _ust_buffers_write(struct ust_buffer *buf, size_t offset,
- const void *src, size_t len, ssize_t cpy);
+void _ust_buffers_strncpy_fixup(struct ust_buffer *buf, size_t offset,
+ size_t len, size_t copied, int terminated);
static __inline__ int ust_buffers_write(struct ust_buffer *buf, size_t offset,
const void *src, size_t len)
{
- size_t cpy;
size_t buf_offset = BUFFER_OFFSET(offset, buf->chan);
assert(buf_offset < buf->chan->subbuf_size*buf->chan->subbuf_cnt);
+ assert(buf_offset + len
+ <= buf->chan->subbuf_size*buf->chan->subbuf_cnt);
+
+ ust_buffers_do_copy(buf->buf_data + buf_offset, src, len);
+
+ return len;
+}
+
+/*
+ * ust_buffers_do_memset - write character into dest.
+ * @dest: destination
+ * @src: source character
+ * @len: length to write
+ */
+static __inline__
+void ust_buffers_do_memset(void *dest, char src, size_t len)
+{
+ /*
+ * What we really want here is an __inline__ memset, but we
+ * don't have constants, so gcc generally uses a function call.
+ */
+ for (; len > 0; len--)
+ *(u8 *)dest++ = src;
+}
+
+/*
+ * ust_buffers_do_strncpy - copy a string up to a certain number of bytes
+ * @dest: destination
+ * @src: source
+ * @len: max. length to copy
+ * @terminated: output string ends with \0 (output)
+ *
+ * returns the number of bytes copied. Does not finalize with \0 if len is
+ * reached.
+ */
+static __inline__
+size_t ust_buffers_do_strncpy(void *dest, const void *src, size_t len,
+ int *terminated)
+{
+ size_t orig_len = len;
- cpy = min_t(size_t, len, buf->buf_size - buf_offset);
- ust_buffers_do_copy(buf->buf_data + buf_offset, src, cpy);
+ *terminated = 0;
+ /*
+ * What we really want here is an __inline__ strncpy, but we
+ * don't have constants, so gcc generally uses a function call.
+ */
+ for (; len > 0; len--) {
+ *(u8 *)dest = CMM_LOAD_SHARED(*(const u8 *)src);
+ /* Check with dest, because src may be modified concurrently */
+ if (*(const u8 *)dest == '\0') {
+ len--;
+ *terminated = 1;
+ break;
+ }
+ dest++;
+ src++;
+ }
+ return orig_len - len;
+}
- if (unlikely(len != cpy))
- _ust_buffers_write(buf, buf_offset, src, len, cpy);
+static __inline__
+int ust_buffers_strncpy(struct ust_buffer *buf, size_t offset, const void *src,
+ size_t len)
+{
+ size_t buf_offset = BUFFER_OFFSET(offset, buf->chan);
+ ssize_t copied;
+ int terminated;
+
+ assert(buf_offset < buf->chan->subbuf_size*buf->chan->subbuf_cnt);
+ assert(buf_offset + len
+ <= buf->chan->subbuf_size*buf->chan->subbuf_cnt);
+
+ copied = ust_buffers_do_strncpy(buf->buf_data + buf_offset,
+ src, len, &terminated);
+ if (unlikely(copied < len || !terminated))
+ _ust_buffers_strncpy_fixup(buf, offset, len, copied,
+ terminated);
return len;
}
projects / ust.git / blobdiff