From 87b15cebdaca21f6725637abeceec0f19f4eec51 Mon Sep 17 00:00:00 2001 From: Florin Coras Date: Sun, 28 Apr 2019 21:16:30 -0700 Subject: svm: more fifo refactor/cleanup Change-Id: Ie76c69641c8598164d0d00fd498018037258fd86 Signed-off-by: Florin Coras --- src/svm/svm_fifo.c | 910 +++++++++++++++++++++++++++-------------------------- src/svm/svm_fifo.h | 678 ++++++++++++++++++++++++++------------- 2 files changed, 917 insertions(+), 671 deletions(-) (limited to 'src/svm') diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c index 4c84fce4c01..3512c72d9f1 100644 --- a/src/svm/svm_fifo.c +++ b/src/svm/svm_fifo.c @@ -101,29 +101,25 @@ svm_fifo_copy_from_chunk (svm_fifo_t * f, svm_fifo_chunk_t * c, u32 head_idx, static inline u8 position_lt (svm_fifo_t * f, u32 a, u32 b, u32 tail) { - return (ooo_segment_distance_from_tail (f, a, tail) - < ooo_segment_distance_from_tail (f, b, tail)); + return (f_distance_to (f, a, tail) < f_distance_to (f, b, tail)); } static inline u8 position_leq (svm_fifo_t * f, u32 a, u32 b, u32 tail) { - return (ooo_segment_distance_from_tail (f, a, tail) - <= ooo_segment_distance_from_tail (f, b, tail)); + return (f_distance_to (f, a, tail) <= f_distance_to (f, b, tail)); } static inline u8 position_gt (svm_fifo_t * f, u32 a, u32 b, u32 tail) { - return (ooo_segment_distance_from_tail (f, a, tail) - > ooo_segment_distance_from_tail (f, b, tail)); + return (f_distance_to (f, a, tail) > f_distance_to (f, b, tail)); } static inline u32 -position_diff (svm_fifo_t * f, u32 posa, u32 posb, u32 tail) +position_diff (svm_fifo_t * f, u32 a, u32 b, u32 tail) { - return ooo_segment_distance_from_tail (f, posa, tail) - - ooo_segment_distance_from_tail (f, posb, tail); + return f_distance_to (f, a, tail) - f_distance_to (f, b, tail); } static inline u32 @@ -132,385 +128,30 @@ ooo_segment_end_pos (svm_fifo_t * f, ooo_segment_t * s) return s->start + s->length; } -u8 * -format_ooo_segment (u8 * s, va_list * args) -{ - svm_fifo_t *f = va_arg (*args, svm_fifo_t *); - ooo_segment_t *seg = va_arg (*args, ooo_segment_t *); - u32 normalized_start = (seg->start + f->nitems - f->tail) % f->size; - s = format (s, "[%u, %u], len %u, next %d, prev %d", normalized_start, - (normalized_start + seg->length) % f->size, seg->length, - seg->next, seg->prev); - return s; -} - -u8 * -svm_fifo_dump_trace (u8 * s, svm_fifo_t * f) -{ -#if SVM_FIFO_TRACE - svm_fifo_trace_elem_t *seg = 0; - int i = 0; - - if (f->trace) - { - vec_foreach (seg, f->trace) - { - s = format (s, "{%u, %u, %u}, ", seg->offset, seg->len, seg->action); - i++; - if (i % 5 == 0) - s = format (s, "\n"); - } - s = format (s, "\n"); - } - return s; -#else - return 0; -#endif -} - -u8 * -svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose) -{ - int i, trace_len; - u8 *data = 0; - svm_fifo_trace_elem_t *trace; - u32 offset; - svm_fifo_t *dummy_fifo; - - if (!f) - return s; - -#if SVM_FIFO_TRACE - trace = f->trace; - trace_len = vec_len (trace); -#else - trace = 0; - trace_len = 0; -#endif - - dummy_fifo = svm_fifo_create (f->size); - clib_memset (f->head_chunk->data, 0xFF, f->nitems); - vec_validate (data, f->nitems); - for (i = 0; i < vec_len (data); i++) - data[i] = i; - - for (i = 0; i < trace_len; i++) - { - offset = trace[i].offset; - if (trace[i].action == 1) - { - if (verbose) - s = format (s, "adding [%u, %u]:", trace[i].offset, - (trace[i].offset + trace[i].len) % dummy_fifo->size); - svm_fifo_enqueue_with_offset (dummy_fifo, trace[i].offset, - trace[i].len, &data[offset]); - } - else if (trace[i].action == 2) - { - if (verbose) - s = format (s, "adding [%u, %u]:", 0, trace[i].len); - svm_fifo_enqueue_nowait (dummy_fifo, trace[i].len, &data[offset]); - } - else if (!no_read) - { - if (verbose) - s = format (s, "read: %u", trace[i].len); - svm_fifo_dequeue_drop (dummy_fifo, trace[i].len); - } - if (verbose) - s = format (s, "%U", format_svm_fifo, dummy_fifo, 1); - } - - s = format (s, "result: %U", format_svm_fifo, dummy_fifo, 1); - - return s; -} - -u8 * -format_ooo_list (u8 * s, va_list * args) -{ - svm_fifo_t *f = va_arg (*args, svm_fifo_t *); - u32 indent = va_arg (*args, u32); - u32 ooo_segment_index = f->ooos_list_head; - ooo_segment_t *seg; - - while (ooo_segment_index != OOO_SEGMENT_INVALID_INDEX) - { - seg = pool_elt_at_index (f->ooo_segments, ooo_segment_index); - s = format (s, "%U%U\n", format_white_space, indent, format_ooo_segment, - f, seg); - ooo_segment_index = seg->next; - } - - return s; -} - -u8 * -format_svm_fifo (u8 * s, va_list * args) -{ - svm_fifo_t *f = va_arg (*args, svm_fifo_t *); - int verbose = va_arg (*args, int); - u32 indent; - - if (!s) - return s; - - indent = format_get_indent (s); - s = format (s, "cursize %u nitems %u has_event %d\n", - svm_fifo_max_dequeue (f), f->nitems, f->has_event); - s = format (s, "%Uhead %u tail %u segment manager %u\n", format_white_space, - indent, (f->head % f->size), (f->tail % f->size), - f->segment_manager); - - if (verbose > 1) - s = format (s, "%Uvpp session %d thread %d app session %d thread %d\n", - format_white_space, indent, f->master_session_index, - f->master_thread_index, f->client_session_index, - f->client_thread_index); - - if (verbose) - { - s = format (s, "%Uooo pool %d active elts newest %u\n", - format_white_space, indent, pool_elts (f->ooo_segments), - f->ooos_newest); - if (svm_fifo_has_ooo_data (f)) - s = format (s, " %U", format_ooo_list, f, indent, verbose); - } - return s; -} - -void -svm_fifo_init (svm_fifo_t * f, u32 size) -{ - f->size = size; - /* - * usable size of the fifo set to rounded_data_size - 1 - * to differentiate between free fifo and empty fifo. - */ - f->nitems = f->size - 1; - f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX; - f->ct_session_index = SVM_FIFO_INVALID_SESSION_INDEX; - f->segment_index = SVM_FIFO_INVALID_INDEX; - f->refcnt = 1; - f->default_chunk.start_byte = 0; - f->default_chunk.length = f->size; - f->default_chunk.next = f->start_chunk = f->end_chunk = &f->default_chunk; - f->head_chunk = f->tail_chunk = f->ooo_enq = f->ooo_deq = f->start_chunk; -} - -/** - * Creates a fifo in the current heap. Fails vs blow up the process - */ -svm_fifo_t * -svm_fifo_create (u32 data_size_in_bytes) -{ - svm_fifo_t *f; - u32 rounded_data_size; - - /* always round fifo data size to the next highest power-of-two */ - rounded_data_size = (1 << (max_log2 (data_size_in_bytes))); - f = clib_mem_alloc_aligned_or_null (sizeof (*f) + rounded_data_size, - CLIB_CACHE_LINE_BYTES); - if (f == 0) - return 0; - - clib_memset (f, 0, sizeof (*f)); - svm_fifo_init (f, data_size_in_bytes); - return f; -} - -/** - * Creates a fifo chunk in the current heap - */ -svm_fifo_chunk_t * -svm_fifo_chunk_alloc (u32 size) -{ - svm_fifo_chunk_t *c; - u32 rounded_size; - - /* round chunk size to the next highest power-of-two */ - rounded_size = (1 << (max_log2 (size))); - c = clib_mem_alloc_aligned_or_null (sizeof (*c) + rounded_size, - CLIB_CACHE_LINE_BYTES); - if (c == 0) - return 0; - - clib_memset (c, 0, sizeof (*c)); - c->length = rounded_size; - return c; -} - -static inline void -svm_fifo_size_update (svm_fifo_t * f, svm_fifo_chunk_t * c) -{ - svm_fifo_chunk_t *prev; - u32 add_bytes = 0; - - if (!c) - return; - - f->end_chunk->next = c; - while (c) - { - add_bytes += c->length; - prev = c; - c = c->next; - } - f->end_chunk = prev; - prev->next = f->start_chunk; - f->size += add_bytes; - f->nitems = f->size - 1; - f->new_chunks = 0; -} - -static void -svm_fifo_try_size_update (svm_fifo_t * f, u32 new_head) -{ - if (new_head % f->size > f->tail % f->size) - return; - - svm_fifo_size_update (f, f->new_chunks); - f->flags &= ~SVM_FIFO_F_SIZE_UPDATE; -} - -void -svm_fifo_add_chunk (svm_fifo_t * f, svm_fifo_chunk_t * c) -{ - svm_fifo_chunk_t *cur, *prev; - - /* Initialize rbtree if needed and add default chunk to it */ - if (!(f->flags & SVM_FIFO_F_MULTI_CHUNK)) - { - rb_tree_init (&f->chunk_lookup); - rb_tree_add2 (&f->chunk_lookup, 0, pointer_to_uword (f->start_chunk)); - f->flags |= SVM_FIFO_F_MULTI_CHUNK; - } - - /* Initialize chunks and add to lookup rbtree. Expectation is that this is - * called with the heap where the rbtree's pool is pushed. */ - cur = c; - if (f->new_chunks) - { - prev = f->new_chunks; - while (prev->next) - prev = prev->next; - prev->next = c; - } - else - prev = f->end_chunk; - - while (cur) - { - cur->start_byte = prev->start_byte + prev->length; - rb_tree_add2 (&f->chunk_lookup, cur->start_byte, - pointer_to_uword (cur)); - prev = cur; - cur = cur->next; - } - - /* If fifo is not wrapped, update the size now */ - if (!svm_fifo_is_wrapped (f)) - { - ASSERT (!f->new_chunks); - svm_fifo_size_update (f, c); - return; - } - - /* Postpone size update */ - if (!f->new_chunks) - { - f->new_chunks = c; - f->flags |= SVM_FIFO_F_SIZE_UPDATE; - } -} - -static inline u8 -svm_fifo_chunk_includes_pos (svm_fifo_chunk_t * c, u32 pos) -{ - return (pos >= c->start_byte && pos < c->start_byte + c->length); -} - -/** - * Find chunk for given byte position - * - * @param f fifo - * @param pos normalized position in fifo - * - * @return chunk that includes given position or 0 - */ -static svm_fifo_chunk_t * -svm_fifo_find_chunk (svm_fifo_t * f, u32 pos) -{ - rb_tree_t *rt = &f->chunk_lookup; - rb_node_t *cur, *prev; - svm_fifo_chunk_t *c; - - cur = rb_node (rt, rt->root); - while (pos != cur->key) - { - prev = cur; - if (pos < cur->key) - cur = rb_node_left (rt, cur); - else - cur = rb_node_right (rt, cur); - - if (rb_node_is_tnil (rt, cur)) - { - /* Hit tnil as a left child. Find predecessor */ - if (pos < prev->key) - { - cur = rb_tree_predecessor (rt, prev); - c = uword_to_pointer (cur->opaque, svm_fifo_chunk_t *); - if (svm_fifo_chunk_includes_pos (c, pos)) - return c; - return 0; - } - /* Hit tnil as a right child. Check if this is the one, otherwise - * search for successor */ - c = uword_to_pointer (prev->opaque, svm_fifo_chunk_t *); - if (svm_fifo_chunk_includes_pos (c, pos)) - return c; - - cur = rb_tree_successor (rt, prev); - c = uword_to_pointer (cur->opaque, svm_fifo_chunk_t *); - if (svm_fifo_chunk_includes_pos (c, pos)) - return c; - return 0; - } - } - - if (!rb_node_is_tnil (rt, cur)) - return uword_to_pointer (cur->opaque, svm_fifo_chunk_t *); - return 0; -} - -void -svm_fifo_free_chunk_lookup (svm_fifo_t * f) -{ - rb_tree_free_nodes (&f->chunk_lookup); -} - void svm_fifo_free_ooo_data (svm_fifo_t * f) { pool_free (f->ooo_segments); } -void -svm_fifo_free (svm_fifo_t * f) +static inline ooo_segment_t * +ooo_segment_get_prev (svm_fifo_t * f, ooo_segment_t * s) { - ASSERT (f->refcnt > 0); + if (s->prev == OOO_SEGMENT_INVALID_INDEX) + return 0; + return pool_elt_at_index (f->ooo_segments, s->prev); +} - if (--f->refcnt == 0) - { - /* ooo data is not allocated on segment heap */ - svm_fifo_free_chunk_lookup (f); - clib_mem_free (f); - } +static inline ooo_segment_t * +ooo_segment_next (svm_fifo_t * f, ooo_segment_t * s) +{ + if (s->next == OOO_SEGMENT_INVALID_INDEX) + return 0; + return pool_elt_at_index (f->ooo_segments, s->next); } -always_inline ooo_segment_t * -ooo_segment_new (svm_fifo_t * f, u32 start, u32 length) +static inline ooo_segment_t * +ooo_segment_alloc (svm_fifo_t * f, u32 start, u32 length) { ooo_segment_t *s; @@ -518,14 +159,13 @@ ooo_segment_new (svm_fifo_t * f, u32 start, u32 length) s->start = start; s->length = length; - s->prev = s->next = OOO_SEGMENT_INVALID_INDEX; return s; } -always_inline void -ooo_segment_del (svm_fifo_t * f, u32 index) +static inline void +ooo_segment_free (svm_fifo_t * f, u32 index) { ooo_segment_t *cur, *prev = 0, *next = 0; cur = pool_elt_at_index (f->ooo_segments, index); @@ -560,8 +200,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 head, u32 tail, u32 length) u32 new_index, s_end_pos, s_index; u32 offset_pos, offset_end_pos; - ASSERT (offset + length <= ooo_segment_distance_from_tail (f, head, tail) - || head == tail); + ASSERT (offset + length <= f_distance_to (f, head, tail) || head == tail); offset_pos = tail + offset; offset_end_pos = tail + offset + length; @@ -570,7 +209,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 head, u32 tail, u32 length) if (f->ooos_list_head == OOO_SEGMENT_INVALID_INDEX) { - s = ooo_segment_new (f, offset_pos, length); + s = ooo_segment_alloc (f, offset_pos, length); f->ooos_list_head = s - f->ooo_segments; f->ooos_newest = f->ooos_list_head; return; @@ -602,7 +241,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 head, u32 tail, u32 length) /* No overlap, add before current segment */ if (position_lt (f, offset_end_pos, s->start, tail)) { - new_s = ooo_segment_new (f, offset_pos, length); + new_s = ooo_segment_alloc (f, offset_pos, length); new_index = new_s - f->ooo_segments; /* Pool might've moved, get segment again */ @@ -627,7 +266,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 head, u32 tail, u32 length) /* No overlap, add after current segment */ else if (position_gt (f, offset_pos, s_end_pos, tail)) { - new_s = ooo_segment_new (f, offset_pos, length); + new_s = ooo_segment_alloc (f, offset_pos, length); new_index = new_s - f->ooo_segments; /* Pool might've moved, get segment again */ @@ -668,7 +307,7 @@ check_tail: offset_end_pos, tail)) { next = ooo_segment_next (f, it); - ooo_segment_del (f, it - f->ooo_segments); + ooo_segment_free (f, it - f->ooo_segments); it = next; } @@ -677,66 +316,284 @@ check_tail: { s->length = position_diff (f, ooo_segment_end_pos (f, it), s->start, tail); - ooo_segment_del (f, it - f->ooo_segments); + ooo_segment_free (f, it - f->ooo_segments); } f->ooos_newest = s - f->ooo_segments; } } /** - * Removes segments that can now be enqueued because the fifo's tail has - * advanced. Returns the number of bytes added to tail. + * Removes segments that can now be enqueued because the fifo's tail has + * advanced. Returns the number of bytes added to tail. + */ +static int +ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued, u32 * tail) +{ + u32 s_index, bytes = 0; + ooo_segment_t *s; + i32 diff; + + s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head); + diff = f_distance_from (f, s->start, *tail); + + ASSERT (diff != n_bytes_enqueued); + + if (diff > n_bytes_enqueued) + return 0; + + /* If last tail update overlaps one/multiple ooo segments, remove them */ + while (0 <= diff && diff < n_bytes_enqueued) + { + s_index = s - f->ooo_segments; + + /* Segment end is beyond the tail. Advance tail and remove segment */ + if (s->length > diff) + { + bytes = s->length - diff; + *tail = *tail + bytes; + ooo_segment_free (f, s_index); + break; + } + + /* If we have next go on */ + if (s->next != OOO_SEGMENT_INVALID_INDEX) + { + s = pool_elt_at_index (f->ooo_segments, s->next); + diff = f_distance_from (f, s->start, *tail); + ooo_segment_free (f, s_index); + } + /* End of search */ + else + { + ooo_segment_free (f, s_index); + break; + } + } + + ASSERT (bytes <= f->nitems); + return bytes; +} + +void +svm_fifo_init (svm_fifo_t * f, u32 size) +{ + f->size = size; + /* + * usable size of the fifo set to rounded_data_size - 1 + * to differentiate between free fifo and empty fifo. + */ + f->nitems = f->size - 1; + f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX; + f->segment_index = SVM_FIFO_INVALID_INDEX; + f->refcnt = 1; + f->default_chunk.start_byte = 0; + f->default_chunk.length = f->size; + f->default_chunk.next = f->start_chunk = f->end_chunk = &f->default_chunk; + f->head_chunk = f->tail_chunk = f->ooo_enq = f->ooo_deq = f->start_chunk; +} + +/** + * Creates a fifo in the current heap. Fails vs blow up the process + */ +svm_fifo_t * +svm_fifo_create (u32 data_size_in_bytes) +{ + svm_fifo_t *f; + u32 rounded_data_size; + + /* always round fifo data size to the next highest power-of-two */ + rounded_data_size = (1 << (max_log2 (data_size_in_bytes))); + f = clib_mem_alloc_aligned_or_null (sizeof (*f) + rounded_data_size, + CLIB_CACHE_LINE_BYTES); + if (f == 0) + return 0; + + clib_memset (f, 0, sizeof (*f)); + svm_fifo_init (f, data_size_in_bytes); + return f; +} + +/** + * Creates a fifo chunk in the current heap + */ +svm_fifo_chunk_t * +svm_fifo_chunk_alloc (u32 size) +{ + svm_fifo_chunk_t *c; + u32 rounded_size; + + /* round chunk size to the next highest power-of-two */ + rounded_size = (1 << (max_log2 (size))); + c = clib_mem_alloc_aligned_or_null (sizeof (*c) + rounded_size, + CLIB_CACHE_LINE_BYTES); + if (c == 0) + return 0; + + clib_memset (c, 0, sizeof (*c)); + c->length = rounded_size; + return c; +} + +static inline void +svm_fifo_size_update (svm_fifo_t * f, svm_fifo_chunk_t * c) +{ + svm_fifo_chunk_t *prev; + u32 add_bytes = 0; + + if (!c) + return; + + f->end_chunk->next = c; + while (c) + { + add_bytes += c->length; + prev = c; + c = c->next; + } + f->end_chunk = prev; + prev->next = f->start_chunk; + f->size += add_bytes; + f->nitems = f->size - 1; + f->new_chunks = 0; +} + +static void +svm_fifo_try_size_update (svm_fifo_t * f, u32 new_head) +{ + if (new_head % f->size > f->tail % f->size) + return; + + svm_fifo_size_update (f, f->new_chunks); + f->flags &= ~SVM_FIFO_F_SIZE_UPDATE; +} + +void +svm_fifo_add_chunk (svm_fifo_t * f, svm_fifo_chunk_t * c) +{ + svm_fifo_chunk_t *cur, *prev; + + /* Initialize rbtree if needed and add default chunk to it */ + if (!(f->flags & SVM_FIFO_F_MULTI_CHUNK)) + { + rb_tree_init (&f->chunk_lookup); + rb_tree_add2 (&f->chunk_lookup, 0, pointer_to_uword (f->start_chunk)); + f->flags |= SVM_FIFO_F_MULTI_CHUNK; + } + + /* Initialize chunks and add to lookup rbtree. Expectation is that this is + * called with the heap where the rbtree's pool is pushed. */ + cur = c; + if (f->new_chunks) + { + prev = f->new_chunks; + while (prev->next) + prev = prev->next; + prev->next = c; + } + else + prev = f->end_chunk; + + while (cur) + { + cur->start_byte = prev->start_byte + prev->length; + rb_tree_add2 (&f->chunk_lookup, cur->start_byte, + pointer_to_uword (cur)); + prev = cur; + cur = cur->next; + } + + /* If fifo is not wrapped, update the size now */ + if (!svm_fifo_is_wrapped (f)) + { + ASSERT (!f->new_chunks); + svm_fifo_size_update (f, c); + return; + } + + /* Postpone size update */ + if (!f->new_chunks) + { + f->new_chunks = c; + f->flags |= SVM_FIFO_F_SIZE_UPDATE; + } +} + +static inline u8 +svm_fifo_chunk_includes_pos (svm_fifo_chunk_t * c, u32 pos) +{ + return (pos >= c->start_byte && pos < c->start_byte + c->length); +} + +/** + * Find chunk for given byte position + * + * @param f fifo + * @param pos normalized position in fifo + * + * @return chunk that includes given position or 0 */ -static int -ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued, u32 * tail) +static svm_fifo_chunk_t * +svm_fifo_find_chunk (svm_fifo_t * f, u32 pos) { - ooo_segment_t *s; - u32 index, bytes = 0; - i32 diff; - - s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head); - diff = ooo_segment_distance_to_tail (f, s->start, *tail); - - ASSERT (diff != n_bytes_enqueued); - - if (diff > n_bytes_enqueued) - return 0; + rb_tree_t *rt = &f->chunk_lookup; + rb_node_t *cur, *prev; + svm_fifo_chunk_t *c; - /* If last tail update overlaps one/multiple ooo segments, remove them */ - while (0 <= diff && diff < n_bytes_enqueued) + cur = rb_node (rt, rt->root); + while (pos != cur->key) { - index = s - f->ooo_segments; + prev = cur; + if (pos < cur->key) + cur = rb_node_left (rt, cur); + else + cur = rb_node_right (rt, cur); - /* Segment end is beyond the tail. Advance tail and remove segment */ - if (s->length > diff) + if (rb_node_is_tnil (rt, cur)) { - bytes = s->length - diff; - *tail = *tail + bytes; - ooo_segment_del (f, index); - break; - } + /* Hit tnil as a left child. Find predecessor */ + if (pos < prev->key) + { + cur = rb_tree_predecessor (rt, prev); + c = uword_to_pointer (cur->opaque, svm_fifo_chunk_t *); + if (svm_fifo_chunk_includes_pos (c, pos)) + return c; + return 0; + } + /* Hit tnil as a right child. Check if this is the one */ + c = uword_to_pointer (prev->opaque, svm_fifo_chunk_t *); + if (svm_fifo_chunk_includes_pos (c, pos)) + return c; - /* If we have next go on */ - if (s->next != OOO_SEGMENT_INVALID_INDEX) - { - s = pool_elt_at_index (f->ooo_segments, s->next); - diff = ooo_segment_distance_to_tail (f, s->start, *tail); - ooo_segment_del (f, index); - } - /* End of search */ - else - { - ooo_segment_del (f, index); - break; + return 0; } } - ASSERT (bytes <= f->nitems); - return bytes; + if (!rb_node_is_tnil (rt, cur)) + return uword_to_pointer (cur->opaque, svm_fifo_chunk_t *); + return 0; +} + +void +svm_fifo_free_chunk_lookup (svm_fifo_t * f) +{ + rb_tree_free_nodes (&f->chunk_lookup); +} + +void +svm_fifo_free (svm_fifo_t * f) +{ + ASSERT (f->refcnt > 0); + + if (--f->refcnt == 0) + { + /* ooo data is not allocated on segment heap */ + svm_fifo_free_chunk_lookup (f); + clib_mem_free (f); + } } void -svm_fifo_overwrite_head (svm_fifo_t * f, u8 * data, u32 len) +svm_fifo_overwrite_head (svm_fifo_t * f, u8 * src, u32 len) { u32 n_chunk; u32 head, tail, head_idx; @@ -750,16 +607,16 @@ svm_fifo_overwrite_head (svm_fifo_t * f, u8 * data, u32 len) head_idx -= c->start_byte; n_chunk = c->length - head_idx; if (len <= n_chunk) - clib_memcpy_fast (&c->data[head_idx], data, len); + clib_memcpy_fast (&c->data[head_idx], src, len); else { - clib_memcpy_fast (&c->data[head_idx], data, n_chunk); - clib_memcpy_fast (&c->next->data[0], data + n_chunk, len - n_chunk); + clib_memcpy_fast (&c->data[head_idx], src, n_chunk); + clib_memcpy_fast (&c->next->data[0], src + n_chunk, len - n_chunk); } } int -svm_fifo_enqueue_nowait (svm_fifo_t * f, u32 len, const u8 * src) +svm_fifo_enqueue (svm_fifo_t * f, u32 len, const u8 * src) { u32 tail, head, free_count; @@ -771,16 +628,15 @@ svm_fifo_enqueue_nowait (svm_fifo_t * f, u32 len, const u8 * src) f->ooos_newest = OOO_SEGMENT_INVALID_INDEX; if (PREDICT_FALSE (free_count == 0)) - return SVM_FIFO_FULL; + return SVM_FIFO_EFULL; /* number of bytes we're going to copy */ len = clib_min (free_count, len); - svm_fifo_copy_to_chunk (f, f->tail_chunk, tail % f->size, src, len, &f->tail_chunk); tail += len; - svm_fifo_trace_add (f, head, n_total, 2); + svm_fifo_trace_add (f, head, len, 2); /* collect out-of-order segments */ if (PREDICT_FALSE (f->ooos_list_head != OOO_SEGMENT_INVALID_INDEX)) @@ -811,14 +667,11 @@ svm_fifo_enqueue_with_offset (svm_fifo_t * f, u32 offset, u32 len, u8 * src) /* will this request fit? */ if ((len + offset) > free_count) - return -1; + return SVM_FIFO_EFULL; f->ooos_newest = OOO_SEGMENT_INVALID_INDEX; - svm_fifo_trace_add (f, offset, len, 1); - ooo_segment_add (f, offset, head, tail, len); - tail_idx = (tail % f->size + offset) % f->size; if (!svm_fifo_chunk_includes_pos (f->ooo_enq, tail_idx)) @@ -830,7 +683,7 @@ svm_fifo_enqueue_with_offset (svm_fifo_t * f, u32 offset, u32 len, u8 * src) } int -svm_fifo_dequeue_nowait (svm_fifo_t * f, u32 len, u8 * dst) +svm_fifo_dequeue (svm_fifo_t * f, u32 len, u8 * dst) { u32 tail, head, cursize; @@ -840,10 +693,9 @@ svm_fifo_dequeue_nowait (svm_fifo_t * f, u32 len, u8 * dst) cursize = f_cursize (f, head, tail); if (PREDICT_FALSE (cursize == 0)) - return -2; /* nothing in the fifo */ + return SVM_FIFO_EEMPTY; len = clib_min (cursize, len); - svm_fifo_copy_from_chunk (f, f->head_chunk, head % f->size, dst, len, &f->head_chunk); head += len; @@ -868,7 +720,7 @@ svm_fifo_peek (svm_fifo_t * f, u32 offset, u32 len, u8 * dst) cursize = f_cursize (f, head, tail); if (PREDICT_FALSE (cursize < offset)) - return -2; /* nothing in the fifo */ + return SVM_FIFO_EEMPTY; len = clib_min (cursize - offset, len); head_idx = (head % f->size + offset) % f->size; @@ -880,28 +732,25 @@ svm_fifo_peek (svm_fifo_t * f, u32 offset, u32 len, u8 * dst) } int -svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes) +svm_fifo_dequeue_drop (svm_fifo_t * f, u32 len) { - u32 total_drop_bytes; - u32 tail, head, cursize; + u32 total_drop_bytes, tail, head, cursize; f_load_head_tail_cons (f, &head, &tail); - /* number of bytes we're going to drop */ + /* number of bytes available */ cursize = f_cursize (f, head, tail); - if (PREDICT_FALSE (cursize == 0)) - return -2; /* nothing in the fifo */ + return SVM_FIFO_EEMPTY; svm_fifo_trace_add (f, tail, total_drop_bytes, 3); /* number of bytes we're going to drop */ - total_drop_bytes = (cursize < max_bytes) ? cursize : max_bytes; + total_drop_bytes = clib_min (cursize, len); /* move head */ head += total_drop_bytes; - ASSERT (cursize >= total_drop_bytes); /* store-rel: consumer owned index (paired with load-acq in producer) */ clib_atomic_store_rel_n (&f->head, head); @@ -928,7 +777,7 @@ svm_fifo_segments (svm_fifo_t * f, svm_fifo_seg_t * fs) cursize = f_cursize (f, head, tail); if (PREDICT_FALSE (cursize == 0)) - return -2; /* nothing in the fifo */ + return SVM_FIFO_EEMPTY; head_idx = head % f->size; @@ -964,7 +813,13 @@ svm_fifo_segments_free (svm_fifo_t * f, svm_fifo_seg_t * fs) clib_atomic_store_rel_n (&f->head, head); } -/* Assumption: no prod and cons are accessing either dest or src fifo */ +/** + * Clones fifo + * + * Assumptions: + * - no prod and cons are accessing either dest or src fifo + * - fifo is not multi chunk + */ void svm_fifo_clone (svm_fifo_t * df, svm_fifo_t * sf) { @@ -977,7 +832,7 @@ svm_fifo_clone (svm_fifo_t * df, svm_fifo_t * sf) } u32 -svm_fifo_number_ooo_segments (svm_fifo_t * f) +svm_fifo_n_ooo_segments (svm_fifo_t * f) { return pool_elts (f->ooo_segments); } @@ -1031,6 +886,153 @@ svm_fifo_del_subscriber (svm_fifo_t * f, u8 subscriber) } } +u8 * +format_ooo_segment (u8 * s, va_list * args) +{ + svm_fifo_t *f = va_arg (*args, svm_fifo_t *); + ooo_segment_t *seg = va_arg (*args, ooo_segment_t *); + u32 normalized_start = (seg->start + f->nitems - f->tail) % f->size; + s = format (s, "[%u, %u], len %u, next %d, prev %d", normalized_start, + (normalized_start + seg->length) % f->size, seg->length, + seg->next, seg->prev); + return s; +} + +u8 * +svm_fifo_dump_trace (u8 * s, svm_fifo_t * f) +{ +#if SVM_FIFO_TRACE + svm_fifo_trace_elem_t *seg = 0; + int i = 0; + + if (f->trace) + { + vec_foreach (seg, f->trace) + { + s = format (s, "{%u, %u, %u}, ", seg->offset, seg->len, seg->action); + i++; + if (i % 5 == 0) + s = format (s, "\n"); + } + s = format (s, "\n"); + } + return s; +#else + return 0; +#endif +} + +u8 * +svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose) +{ + int i, trace_len; + u8 *data = 0; + svm_fifo_trace_elem_t *trace; + u32 offset; + svm_fifo_t *dummy_fifo; + + if (!f) + return s; + +#if SVM_FIFO_TRACE + trace = f->trace; + trace_len = vec_len (trace); +#else + trace = 0; + trace_len = 0; +#endif + + dummy_fifo = svm_fifo_create (f->size); + clib_memset (f->head_chunk->data, 0xFF, f->nitems); + vec_validate (data, f->nitems); + for (i = 0; i < vec_len (data); i++) + data[i] = i; + + for (i = 0; i < trace_len; i++) + { + offset = trace[i].offset; + if (trace[i].action == 1) + { + if (verbose) + s = format (s, "adding [%u, %u]:", trace[i].offset, + (trace[i].offset + trace[i].len) % dummy_fifo->size); + svm_fifo_enqueue_with_offset (dummy_fifo, trace[i].offset, + trace[i].len, &data[offset]); + } + else if (trace[i].action == 2) + { + if (verbose) + s = format (s, "adding [%u, %u]:", 0, trace[i].len); + svm_fifo_enqueue (dummy_fifo, trace[i].len, &data[offset]); + } + else if (!no_read) + { + if (verbose) + s = format (s, "read: %u", trace[i].len); + svm_fifo_dequeue_drop (dummy_fifo, trace[i].len); + } + if (verbose) + s = format (s, "%U", format_svm_fifo, dummy_fifo, 1); + } + + s = format (s, "result: %U", format_svm_fifo, dummy_fifo, 1); + + return s; +} + +u8 * +format_ooo_list (u8 * s, va_list * args) +{ + svm_fifo_t *f = va_arg (*args, svm_fifo_t *); + u32 indent = va_arg (*args, u32); + u32 ooo_segment_index = f->ooos_list_head; + ooo_segment_t *seg; + + while (ooo_segment_index != OOO_SEGMENT_INVALID_INDEX) + { + seg = pool_elt_at_index (f->ooo_segments, ooo_segment_index); + s = format (s, "%U%U\n", format_white_space, indent, format_ooo_segment, + f, seg); + ooo_segment_index = seg->next; + } + + return s; +} + +u8 * +format_svm_fifo (u8 * s, va_list * args) +{ + svm_fifo_t *f = va_arg (*args, svm_fifo_t *); + int verbose = va_arg (*args, int); + u32 indent; + + if (!s) + return s; + + indent = format_get_indent (s); + s = format (s, "cursize %u nitems %u has_event %d\n", + svm_fifo_max_dequeue (f), f->nitems, f->has_event); + s = format (s, "%Uhead %u tail %u segment manager %u\n", format_white_space, + indent, (f->head % f->size), (f->tail % f->size), + f->segment_manager); + + if (verbose > 1) + s = format (s, "%Uvpp session %d thread %d app session %d thread %d\n", + format_white_space, indent, f->master_session_index, + f->master_thread_index, f->client_session_index, + f->client_thread_index); + + if (verbose) + { + s = format (s, "%Uooo pool %d active elts newest %u\n", + format_white_space, indent, pool_elts (f->ooo_segments), + f->ooos_newest); + if (svm_fifo_has_ooo_data (f)) + s = format (s, " %U", format_ooo_list, f, indent, verbose); + } + return s; +} + #endif /* * fd.io coding-style-patch-verification: ON diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h index c9370955dfb..96ca3ee01c6 100644 --- a/src/svm/svm_fifo.h +++ b/src/svm/svm_fifo.h @@ -30,7 +30,6 @@ typedef struct { u32 next; /**< Next linked-list element pool index */ u32 prev; /**< Previous linked-list element pool index */ - u32 start; /**< Start of segment, normalized*/ u32 length; /**< Length of segment */ } ooo_segment_t; @@ -41,12 +40,12 @@ typedef struct #define SVM_FIFO_INVALID_INDEX ((u32)~0) #define SVM_FIFO_MAX_EVT_SUBSCRIBERS 7 -enum svm_fifo_tx_ntf_ +typedef enum svm_fifo_tx_ntf_ { SVM_FIFO_NO_TX_NOTIF = 0, SVM_FIFO_WANT_TX_NOTIF = 1, SVM_FIFO_WANT_TX_NOTIF_IF_FULL = 2, -}; +} svm_fifo_tx_ntf_t; typedef struct { @@ -57,10 +56,10 @@ typedef struct typedef struct svm_fifo_chunk_ { - u32 start_byte; - u32 length; - struct svm_fifo_chunk_ *next; - u8 data[0]; + u32 start_byte; /**< chunk start byte */ + u32 length; /**< length of chunk in bytes */ + struct svm_fifo_chunk_ *next; /**< pointer to next chunk in linked-lists */ + u8 data[0]; /**< start of chunk data */ } svm_fifo_chunk_t; typedef enum svm_fifo_flag_ @@ -73,8 +72,8 @@ typedef enum svm_fifo_flag_ typedef struct _svm_fifo { CLIB_CACHE_LINE_ALIGN_MARK (shared_first); - u32 size; /**< size of the fifo */ - u32 nitems; /**< usable size(size-1) */ + u32 size; /**< size of the fifo in bytes */ + u32 nitems; /**< usable size (size-1) */ u8 flags; /**< fifo flags */ svm_fifo_chunk_t *start_chunk;/**< first chunk in fifo chunk list */ svm_fifo_chunk_t *end_chunk; /**< end chunk in fifo chunk list */ @@ -83,14 +82,12 @@ typedef struct _svm_fifo CLIB_CACHE_LINE_ALIGN_MARK (shared_second); volatile u32 has_event; /**< non-zero if deq event exists */ - - u32 master_session_index; - u32 client_session_index; - u8 master_thread_index; - u8 client_thread_index; - u32 segment_manager; - u32 segment_index; - u32 ct_session_index; /**< Local session index for vpp */ + u32 master_session_index; /**< session layer session index */ + u32 client_session_index; /**< app session index */ + u8 master_thread_index; /**< session layer thread index */ + u8 client_thread_index; /**< app worker index */ + u32 segment_manager; /**< session layer segment manager index */ + u32 segment_index; /**< segment index in segment manager */ u32 freelist_index; /**< aka log2(allocated_size) - const. */ i8 refcnt; /**< reference count */ struct _svm_fifo *next; /**< next in freelist/active chain */ @@ -110,7 +107,7 @@ typedef struct _svm_fifo svm_fifo_chunk_t *ooo_enq; /**< last chunk used for ooo enqueue */ ooo_segment_t *ooo_segments; /**< Pool of ooo segments */ u32 ooos_newest; /**< Last segment to have been updated */ - volatile u8 n_subscribers; + volatile u8 n_subscribers; /**< Number of subscribers for io events */ u8 subscribers[SVM_FIFO_MAX_EVT_SUBSCRIBERS]; #if SVM_FIFO_TRACE @@ -122,7 +119,8 @@ typedef struct _svm_fifo typedef enum { - SVM_FIFO_FULL = -2, + SVM_FIFO_EFULL = -2, + SVM_FIFO_EEMPTY = -3, } svm_fifo_err_t; typedef struct svm_fifo_seg_ @@ -147,7 +145,11 @@ typedef struct svm_fifo_seg_ u8 *svm_fifo_dump_trace (u8 * s, svm_fifo_t * f); u8 *svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose); -/* internal function */ +/** + * Load head and tail optimized for consumer + * + * Internal function. + */ static inline void f_load_head_tail_cons (svm_fifo_t * f, u32 * head, u32 * tail) { @@ -157,7 +159,10 @@ f_load_head_tail_cons (svm_fifo_t * f, u32 * head, u32 * tail) *tail = clib_atomic_load_acq_n (&f->tail); } -/* internal function */ +/** Load head and tail optimized for producer + * + * Internal function + */ static inline void f_load_head_tail_prod (svm_fifo_t * f, u32 * head, u32 * tail) { @@ -167,8 +172,10 @@ f_load_head_tail_prod (svm_fifo_t * f, u32 * head, u32 * tail) *head = clib_atomic_load_acq_n (&f->head); } -/* producer consumer role independent */ -/* internal function */ +/* Load head and tail independent of producer/consumer role + * + * Internal function. + */ static inline void f_load_head_tail_all_acq (svm_fifo_t * f, u32 * head, u32 * tail) { @@ -178,21 +185,247 @@ f_load_head_tail_all_acq (svm_fifo_t * f, u32 * head, u32 * tail) *head = clib_atomic_load_acq_n (&f->head); } -/* internal function */ +/** + * Fifo free bytes, i.e., number of free bytes + * + * Internal function + */ static inline u32 f_free_count (svm_fifo_t * f, u32 head, u32 tail) { return (f->nitems + head - tail); } -/* internal function */ +/** + * Fifo current size, i.e., number of bytes enqueued + * + * Internal function. + */ static inline u32 f_cursize (svm_fifo_t * f, u32 head, u32 tail) { return (f->nitems - f_free_count (f, head, tail)); } -/* used by consumer */ +/** + * Distance to a from b, i.e., a - b in the fifo + * + * Internal function. + */ +static inline u32 +f_distance_to (svm_fifo_t * f, u32 a, u32 b) +{ + return ((f->size + a - b) % f->size); +} + +/** + * Distance from a to b, i.e., b - a in the fifo + * + * Internal function. + */ +static inline u32 +f_distance_from (svm_fifo_t * f, u32 a, u32 b) +{ + return ((f->size + b - a) % f->size); +} + +/** + * Create fifo of requested size + * + * Allocates fifo on current heap. + * + * @param size data size in bytes for fifo to be allocated. Will be + * rounded to the next highest power-of-two value. + * @return pointer to new fifo + */ +svm_fifo_t *svm_fifo_create (u32 size); +/** + * Initialize fifo + * + * @param size size for fifo + */ +void svm_fifo_init (svm_fifo_t * f, u32 size); +/** + * Allocate a fifo chunk on heap + * + * If the chunk is allocated on a fifo segment, this should be called + * with the segment's heap pushed. + * + * @param size chunk size in bytes. Will be rounded to the next highest + * power-of-two + * @return new chunk or 0 if alloc failed + */ +svm_fifo_chunk_t *svm_fifo_chunk_alloc (u32 size); +/** + * Grow fifo size by adding chunk to chunk list + * + * If fifos are allocated on a segment, this should be called with + * the segment's heap pushed. + * + * @param f fifo to be extended + * @param c chunk or linked list of chunks to be added + */ +void svm_fifo_add_chunk (svm_fifo_t * f, svm_fifo_chunk_t * c); +/** + * Free fifo and associated state + * + * @param f fifo + */ +void svm_fifo_free (svm_fifo_t * f); +/** + * Cleanup fifo chunk lookup rb tree + * + * The rb tree is allocated in segment heap so this should be called + * with it pushed. + * + * @param f fifo to cleanup + */ +void svm_fifo_free_chunk_lookup (svm_fifo_t * f); +/** + * Cleanup fifo ooo data + * + * The ooo data is allocated in producer process memory. The fifo + * segment heap should not be pushed. + * + * @param f fifo to cleanup + */ +void svm_fifo_free_ooo_data (svm_fifo_t * f); +/** + * Init fifo head and tail + * + * @param f fifo + * @param head head value that will be matched to a chunk + * @param tail tail value that will be matched to a chunk + */ +void svm_fifo_init_pointers (svm_fifo_t * f, u32 head, u32 tail); +/** + * Clone fifo + * + * Clones single/default chunk fifo. It does not work for fifos with + * multiple chunks. + */ +void svm_fifo_clone (svm_fifo_t * df, svm_fifo_t * sf); +/** + * Enqueue data to fifo + * + * Data is enqueued and tail pointer is updated atomically. If the new data + * enqueued partly overlaps or "touches" an out-of-order segment, said segment + * is "consumed" and the number of bytes returned is appropriately updated. + * + * @param f fifo + * @param len length of data to copy + * @param src buffer from where to copy the data + * @return number of contiguous bytes that can be consumed or error + */ +int svm_fifo_enqueue (svm_fifo_t * f, u32 len, const u8 * src); +/** + * Enqueue data to fifo with offset + * + * Data is enqueued without updating tail pointer. Instead, an out-of-order + * list of segments is generated and maintained. Fifo takes care of coalescing + * contiguous or overlapping segments. + * + * @param f fifo + * @param offset offset at which to copy the data + * @param len len of data to copy + * @param src buffer from where to copy the data + * @return 0 if enqueue was successful, error otherwise + */ +int svm_fifo_enqueue_with_offset (svm_fifo_t * f, u32 offset, u32 len, + u8 * src); +/** + * Overwrite fifo head with new data + * + * This should be typically used by dgram transport protocols that need + * to update the dgram header after dequeueing a chunk of data. It assumes + * that the dgram header is at most spread over two chunks. + * + * @param f fifo + * @param src src of new data + * @param len length of new data + */ +void svm_fifo_overwrite_head (svm_fifo_t * f, u8 * src, u32 len); +/** + * Dequeue data from fifo + * + * Data is dequeued to consumer provided buffer and head is atomically + * updated. + * + * @param f fifo + * @param len length of data to dequeue + * @param dst buffer to where to dequeue the data + * @return number of bytes dequeued or error + */ +int svm_fifo_dequeue (svm_fifo_t * f, u32 len, u8 * dst); +/** + * Peek data from fifo + * + * Data is copied from requested offset into provided dst buffer. Head is + * not updated. + * + * @param f fifo + * @param offset offset from which to copy the data + * @param len length of data to copy + * @param dst buffer to where to dequeue the data + * @return number of bytes peeked + */ +int svm_fifo_peek (svm_fifo_t * f, u32 offset, u32 len, u8 * dst); +/** + * Dequeue and drop bytes from fifo + * + * Advances fifo head by requested amount of bytes. + * + * @param f fifo + * @param len number of bytes to drop + * @return number of bytes dropped + */ +int svm_fifo_dequeue_drop (svm_fifo_t * f, u32 len); +/** + * Dequeue and drop all bytes from fifo + * + * Advances head to tail position. + * + * @param f fifo + */ +void svm_fifo_dequeue_drop_all (svm_fifo_t * f); +int svm_fifo_segments (svm_fifo_t * f, svm_fifo_seg_t * fs); +void svm_fifo_segments_free (svm_fifo_t * f, svm_fifo_seg_t * fs); +/** + * Add io events subscriber to list + * + * @param f fifo + * @param sub subscriber opaque index (typically app worker index) + */ +void svm_fifo_add_subscriber (svm_fifo_t * f, u8 sub); +/** + * Remove io events subscriber form list + * + * @param f fifo + * @param sub subscriber index to be removed + */ +void svm_fifo_del_subscriber (svm_fifo_t * f, u8 subscriber); +/** + * Number of out-of-order segments for fifo + * + * @param f fifo + * @return number of out of order segments + */ +u32 svm_fifo_n_ooo_segments (svm_fifo_t * f); +/* + * First out-of-order segment for fifo + * + * @param f fifo + * @return first out-of-order segment for fifo + */ +ooo_segment_t *svm_fifo_first_ooo_segment (svm_fifo_t * f); +format_function_t format_svm_fifo; + +/** + * Fifo max bytes to dequeue optimized for consumer + * + * @param f fifo + * @return max number of bytes that can be dequeued + */ static inline u32 svm_fifo_max_dequeue_cons (svm_fifo_t * f) { @@ -201,7 +434,12 @@ svm_fifo_max_dequeue_cons (svm_fifo_t * f) return f_cursize (f, head, tail); } -/* used by producer*/ +/** + * Fifo max bytes to dequeue optimized for producer + * + * @param f fifo + * @return max number of bytes that can be dequeued + */ static inline u32 svm_fifo_max_dequeue_prod (svm_fifo_t * f) { @@ -210,9 +448,12 @@ svm_fifo_max_dequeue_prod (svm_fifo_t * f) return f_cursize (f, head, tail); } -/* use producer or consumer specific functions for perfomance. - * svm_fifo_max_dequeue_cons (svm_fifo_t *f) - * svm_fifo_max_dequeue_prod (svm_fifo_t *f) +/** + * Fifo max bytes to dequeue + * + * Note: use producer or consumer specific functions for performance: + * @ref svm_fifo_max_dequeue_cons (svm_fifo_t *f) + * @ref svm_fifo_max_dequeue_prod (svm_fifo_t *f) */ static inline u32 svm_fifo_max_dequeue (svm_fifo_t * f) @@ -222,15 +463,22 @@ svm_fifo_max_dequeue (svm_fifo_t * f) return f_cursize (f, head, tail); } -/* used by producer */ +/** + * Check if fifo is full optimized for producer + * + * @param f fifo + * @return 1 if fifo is full 0 otherwise + */ static inline int svm_fifo_is_full_prod (svm_fifo_t * f) { return (svm_fifo_max_dequeue_prod (f) == f->nitems); } -/* use producer or consumer specific functions for perfomance. - * svm_fifo_is_full_prod (svm_fifo_t * f) +/* Check if fifo is full. + * + * Note: use producer or consumer specific functions for performance. + * @ref svm_fifo_is_full_prod (svm_fifo_t * f) * add cons version if needed */ static inline int @@ -239,23 +487,36 @@ svm_fifo_is_full (svm_fifo_t * f) return (svm_fifo_max_dequeue (f) == f->nitems); } -/* used by consumer */ +/** + * Check if fifo is empty optimized for consumer + * + * @param f fifo + * @return 1 if fifo is empty 0 otherwise + */ static inline int svm_fifo_is_empty_cons (svm_fifo_t * f) { return (svm_fifo_max_dequeue_cons (f) == 0); } -/* used by producer */ +/** + * Check if fifo is empty optimized for producer + * + * @param f fifo + * @return 1 if fifo is empty 0 otherwise + */ static inline int svm_fifo_is_empty_prod (svm_fifo_t * f) { return (svm_fifo_max_dequeue_prod (f) == 0); } -/* use producer or consumer specific functions for perfomance. - * svm_fifo_is_empty_cons (svm_fifo_t * f) - * svm_fifo_is_empty_prod (svm_fifo_t * f) +/** + * Check if fifo is empty + * + * Note: use producer or consumer specific functions for perfomance. + * @ref svm_fifo_is_empty_cons (svm_fifo_t * f) + * @ref svm_fifo_is_empty_prod (svm_fifo_t * f) */ static inline int svm_fifo_is_empty (svm_fifo_t * f) @@ -263,6 +524,12 @@ svm_fifo_is_empty (svm_fifo_t * f) return (svm_fifo_max_dequeue (f) == 0); } +/** + * Check if fifo is wrapped + * + * @param f fifo + * @return 1 if 'normalized' head is ahead of tail + */ static inline u8 svm_fifo_is_wrapped (svm_fifo_t * f) { @@ -271,7 +538,14 @@ svm_fifo_is_wrapped (svm_fifo_t * f) return head % f->size > tail % f->size; } -/* used by producer*/ +/** + * Maximum number of bytes that can be enqueued into fifo + * + * Optimized for producer + * + * @param f fifo + * @return max number of bytes that can be enqueued into fifo + */ static inline u32 svm_fifo_max_enqueue_prod (svm_fifo_t * f) { @@ -280,8 +554,10 @@ svm_fifo_max_enqueue_prod (svm_fifo_t * f) return f_free_count (f, head, tail); } -/* use producer or consumer specfic functions for perfomance. - * svm_fifo_max_enqueue_prod (svm_fifo_t *f) +/* Maximum number of bytes that can be enqueued into fifo + * + * Note: use producer or consumer specific functions for performance. + * @ref svm_fifo_max_enqueue_prod (svm_fifo_t *f) * add consumer specific version if needed. */ static inline u32 @@ -292,111 +568,10 @@ svm_fifo_max_enqueue (svm_fifo_t * f) return f_free_count (f, head, tail); } -static inline int -svm_fifo_has_event (svm_fifo_t * f) -{ - return f->has_event; -} - -static inline u8 -svm_fifo_has_ooo_data (svm_fifo_t * f) -{ - return f->ooos_list_head != OOO_SEGMENT_INVALID_INDEX; -} - -/** - * Sets fifo event flag. - * - * Also acts as a release ordering. - * - * @return 1 if flag was not set. - */ -always_inline u8 -svm_fifo_set_event (svm_fifo_t * f) -{ - /* return __sync_lock_test_and_set (&f->has_event, 1) == 0; - return __sync_bool_compare_and_swap (&f->has_event, 0, 1); */ - return !clib_atomic_swap_rel_n (&f->has_event, 1); -} - -/** - * Unsets fifo event flag. - * - * Also acts as an acquire barrier. - */ -always_inline void -svm_fifo_unset_event (svm_fifo_t * f) -{ - clib_atomic_swap_acq_n (&f->has_event, 0); -} - -svm_fifo_t *svm_fifo_create (u32 data_size_in_bytes); -void svm_fifo_init (svm_fifo_t * f, u32 size); - -/** - * Allocate a fifo chunk on heap - * - * If the chunk is allocated on a fifo segment, this should be called - * with the segment's heap pushed. - * - * @param size chunk size in bytes. Will be rounded to the next highest - * power-of-two - * @return new chunk or 0 if alloc failed - */ -svm_fifo_chunk_t *svm_fifo_chunk_alloc (u32 size); - -/** - * Grow fifo size by adding chunk to chunk list - * - * If fifos are allocated on a segment, this should be called with - * the segment's heap pushed. - * - * @param f fifo to be extended - * @param c chunk or linked list of chunks to be added - */ -void svm_fifo_add_chunk (svm_fifo_t * f, svm_fifo_chunk_t * c); -void svm_fifo_free (svm_fifo_t * f); -/** - * Cleanup fifo chunk lookup rb tree - * - * The rb tree is allocated in segment heap so this should be called - * with it pushed. - * - * @param f fifo to cleanup - */ -void svm_fifo_free_chunk_lookup (svm_fifo_t * f); -/** - * Cleanup fifo ooo data - * - * The ooo data is allocated in producer process memory. The fifo - * segment heap should not be pushed. - * - * @param f fifo to cleanup - */ -void svm_fifo_free_ooo_data (svm_fifo_t * f); - -int svm_fifo_enqueue_nowait (svm_fifo_t * f, u32 max_bytes, - const u8 * copy_from_here); -int svm_fifo_enqueue_with_offset (svm_fifo_t * f, u32 offset, - u32 required_bytes, u8 * copy_from_here); -int svm_fifo_dequeue_nowait (svm_fifo_t * f, u32 max_bytes, u8 * copy_here); - -int svm_fifo_peek (svm_fifo_t * f, u32 offset, u32 max_bytes, u8 * copy_here); -int svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes); -void svm_fifo_dequeue_drop_all (svm_fifo_t * f); -int svm_fifo_segments (svm_fifo_t * f, svm_fifo_seg_t * fs); -void svm_fifo_segments_free (svm_fifo_t * f, svm_fifo_seg_t * fs); -void svm_fifo_init_pointers (svm_fifo_t * f, u32 head, u32 tail); -void svm_fifo_clone (svm_fifo_t * df, svm_fifo_t * sf); -void svm_fifo_overwrite_head (svm_fifo_t * f, u8 * data, u32 len); -void svm_fifo_add_subscriber (svm_fifo_t * f, u8 subscriber); -void svm_fifo_del_subscriber (svm_fifo_t * f, u8 subscriber); -format_function_t format_svm_fifo; - /** * Max contiguous chunk of data that can be read */ -always_inline u32 +static inline u32 svm_fifo_max_read_chunk (svm_fifo_t * f) { u32 head, tail; @@ -410,7 +585,7 @@ svm_fifo_max_read_chunk (svm_fifo_t * f) /** * Max contiguous chunk of data that can be written */ -always_inline u32 +static inline u32 svm_fifo_max_write_chunk (svm_fifo_t * f) { u32 head, tail; @@ -425,19 +600,22 @@ svm_fifo_max_write_chunk (svm_fifo_t * f) * Advance tail pointer * * Useful for moving tail pointer after external enqueue. + * + * @param f fifo + * @param len number of bytes to add to tail */ -always_inline void -svm_fifo_enqueue_nocopy (svm_fifo_t * f, u32 bytes) +static inline void +svm_fifo_enqueue_nocopy (svm_fifo_t * f, u32 len) { - ASSERT (bytes <= svm_fifo_max_enqueue_prod (f)); + ASSERT (len <= svm_fifo_max_enqueue_prod (f)); /* load-relaxed: producer owned index */ u32 tail = f->tail; - tail += bytes; + tail += len; /* store-rel: producer owned index (paired with load-acq in consumer) */ clib_atomic_store_rel_n (&f->tail, tail); } -always_inline u8 * +static inline u8 * svm_fifo_head (svm_fifo_t * f) { /* load-relaxed: consumer owned index */ @@ -445,7 +623,7 @@ svm_fifo_head (svm_fifo_t * f) + ((f->head % f->size) - f->head_chunk->start_byte)); } -always_inline u8 * +static inline u8 * svm_fifo_tail (svm_fifo_t * f) { /* load-relaxed: producer owned index */ @@ -453,18 +631,132 @@ svm_fifo_tail (svm_fifo_t * f) + ((f->tail % f->size) - f->tail_chunk->start_byte)); } +static inline u8 +svm_fifo_n_subscribers (svm_fifo_t * f) +{ + return f->n_subscribers; +} + +/** + * Check if fifo has out-of-order data + * + * @param f fifo + * @return 1 if fifo has ooo data, 0 otherwise + */ +static inline u8 +svm_fifo_has_ooo_data (svm_fifo_t * f) +{ + return f->ooos_list_head != OOO_SEGMENT_INVALID_INDEX; +} + +static inline ooo_segment_t * +svm_fifo_newest_ooo_segment (svm_fifo_t * f) +{ + if (f->ooos_newest == OOO_SEGMENT_INVALID_INDEX) + return 0; + return pool_elt_at_index (f->ooo_segments, f->ooos_newest); +} + +static inline void +svm_fifo_newest_ooo_segment_reset (svm_fifo_t * f) +{ + f->ooos_newest = OOO_SEGMENT_INVALID_INDEX; +} + +static inline u32 +ooo_segment_offset_prod (svm_fifo_t * f, ooo_segment_t * s) +{ + u32 tail; + /* load-relaxed: producer owned index */ + tail = f->tail; + + return f_distance_to (f, s->start, tail); +} + +static inline u32 +ooo_segment_length (svm_fifo_t * f, ooo_segment_t * s) +{ + return s->length; +} + +/** + * Check if fifo has io event + * + * @param f fifo + * @return 1 if fifo has event, 0 otherwise + */ +static inline int +svm_fifo_has_event (svm_fifo_t * f) +{ + return f->has_event; +} + +/** + * Set fifo event flag. + * + * Forces release semantics. + * + * @param f fifo + * @return 1 if flag was not set, 0 otherwise + */ +always_inline u8 +svm_fifo_set_event (svm_fifo_t * f) +{ + return !clib_atomic_swap_rel_n (&f->has_event, 1); +} + +/** + * Unset fifo event flag. + * + * Forces acquire semantics + * + * @param f fifo + */ +always_inline void +svm_fifo_unset_event (svm_fifo_t * f) +{ + clib_atomic_swap_acq_n (&f->has_event, 0); +} + +/** + * Set specific want tx notification flag + * + * For list of flags see @ref svm_fifo_tx_ntf_t + * + * @param f fifo + * @param ntf_type type of notification requested + */ static inline void svm_fifo_add_want_tx_ntf (svm_fifo_t * f, u8 ntf_type) { f->want_tx_ntf |= ntf_type; } +/** + * Clear specific want tx notification flag + * + * For list of flags see @ref svm_fifo_tx_ntf_t + * + * @param f fifo + * @param ntf_type type of notification to be cleared + */ static inline void svm_fifo_del_want_tx_ntf (svm_fifo_t * f, u8 ntf_type) { f->want_tx_ntf &= ~ntf_type; } +/** + * Clear the want tx notification flag and set has tx notification + * + * Should be used after enqueuing a tx event. This clears the + * SVM_FIFO_WANT_TX_NOTIF flag but it does not clear + * SVM_FIFO_WANT_TX_NOTIF_IF_FULL. If the latter was set, has_tx_ntf is + * set to avoid enqueueing tx events for for all dequeue operations until + * it is manually cleared. + * + * @param f fifo + */ static inline void svm_fifo_clear_tx_ntf (svm_fifo_t * f) { @@ -473,12 +765,31 @@ svm_fifo_clear_tx_ntf (svm_fifo_t * f) svm_fifo_del_want_tx_ntf (f, SVM_FIFO_WANT_TX_NOTIF); } +/** + * Clear has tx notification flag + * + * The fifo generates only one event per SVM_FIFO_WANT_TX_NOTIF_IF_FULL + * request and sets has_tx_ntf. To received new events the flag must be + * cleared using this function. + * + * @param f fifo + */ static inline void svm_fifo_reset_tx_ntf (svm_fifo_t * f) { f->has_tx_ntf = 0; } +/** + * Check if fifo needs tx notification + * + * Determines based on tx notification request flags and state of the fifo if + * a tx io event should be generated. + * + * @param f fifo + * @param n_last_deq number of bytes last dequeued + * @return 1 if tx io event should be generated, 0 otherwise + */ static inline u8 svm_fifo_needs_tx_ntf (svm_fifo_t * f, u32 n_last_deq) { @@ -501,73 +812,6 @@ svm_fifo_needs_tx_ntf (svm_fifo_t * f, u32 n_last_deq) return 0; } -always_inline u8 -svm_fifo_n_subscribers (svm_fifo_t * f) -{ - return f->n_subscribers; -} - -u32 svm_fifo_number_ooo_segments (svm_fifo_t * f); -ooo_segment_t *svm_fifo_first_ooo_segment (svm_fifo_t * f); - -always_inline ooo_segment_t * -svm_fifo_newest_ooo_segment (svm_fifo_t * f) -{ - if (f->ooos_newest == OOO_SEGMENT_INVALID_INDEX) - return 0; - return pool_elt_at_index (f->ooo_segments, f->ooos_newest); -} - -always_inline void -svm_fifo_newest_ooo_segment_reset (svm_fifo_t * f) -{ - f->ooos_newest = OOO_SEGMENT_INVALID_INDEX; -} - -always_inline u32 -ooo_segment_distance_from_tail (svm_fifo_t * f, u32 pos, u32 tail) -{ - return ((f->size + pos - tail) % f->size); -} - -always_inline u32 -ooo_segment_distance_to_tail (svm_fifo_t * f, u32 pos, u32 tail) -{ - return ((f->size + tail - pos) % f->size); -} - -always_inline u32 -ooo_segment_offset_prod (svm_fifo_t * f, ooo_segment_t * s) -{ - u32 tail; - /* load-relaxed: producer owned index */ - tail = f->tail; - - return ooo_segment_distance_from_tail (f, s->start, tail); -} - -always_inline u32 -ooo_segment_length (svm_fifo_t * f, ooo_segment_t * s) -{ - return s->length; -} - -always_inline ooo_segment_t * -ooo_segment_get_prev (svm_fifo_t * f, ooo_segment_t * s) -{ - if (s->prev == OOO_SEGMENT_INVALID_INDEX) - return 0; - return pool_elt_at_index (f->ooo_segments, s->prev); -} - -always_inline ooo_segment_t * -ooo_segment_next (svm_fifo_t * f, ooo_segment_t * s) -{ - if (s->next == OOO_SEGMENT_INVALID_INDEX) - return 0; - return pool_elt_at_index (f->ooo_segments, s->next); -} - #endif /* __included_ssvm_fifo_h__ */ /* -- cgit 1.2.3-korg