diff options
Diffstat (limited to 'src/vlib')
| -rw-r--r-- | src/vlib/buffer.c | 599 | ||||
| -rw-r--r-- | src/vlib/buffer.h | 122 | ||||
| -rw-r--r-- | src/vlib/buffer_funcs.h | 516 | ||||
| -rw-r--r-- | src/vlib/log.h | 2 | ||||
| -rw-r--r-- | src/vlib/main.c | 9 | ||||
| -rw-r--r-- | src/vlib/main.h | 3 | ||||
| -rw-r--r-- | src/vlib/threads.c | 22 | ||||
| -rw-r--r-- | src/vlib/vlib.h | 1 |
8 files changed, 591 insertions, 683 deletions
diff --git a/src/vlib/buffer.c b/src/vlib/buffer.c index 643f2c50342..3e194113675 100644 --- a/src/vlib/buffer.c +++ b/src/vlib/buffer.c @@ -43,18 +43,24 @@ * Allocate/free network buffers. */ +#include <vppinfra/linux/sysfs.h> #include <vlib/vlib.h> #include <vlib/unix/unix.h> -vlib_buffer_callbacks_t *vlib_buffer_callbacks = 0; +#define VLIB_BUFFER_DEFAULT_BUFFERS_PER_NUMA 16384 +#define VLIB_BUFFER_DEFAULT_BUFFERS_PER_NUMA_UNPRIV 8192 -/* when running unpriviledged we are limited by RLIMIT_MEMLOCK which is - typically set to 16MB so setting default size for buffer memory to 14MB - */ -static u32 vlib_buffer_physmem_sz = 14 << 20; +#ifdef CLIB_HAVE_VEC128 +/* Assumptions by vlib_buffer_free_inline: */ +STATIC_ASSERT_FITS_IN (vlib_buffer_t, flags, 16); +STATIC_ASSERT_FITS_IN (vlib_buffer_t, ref_count, 16); +STATIC_ASSERT_FITS_IN (vlib_buffer_t, buffer_pool_index, 16); +#endif -/* logging */ -static vlib_log_class_t buffer_log_default; +/* Make sure that buffer template size is not accidentally changed */ +STATIC_ASSERT_OFFSET_OF (vlib_buffer_t, template_end, 64); + +u16 __vlib_buffer_external_hdr_size = 0; uword vlib_buffer_length_in_chain_slow_path (vlib_main_t * vm, @@ -86,8 +92,8 @@ format_vlib_buffer (u8 * s, va_list * args) foreach_vlib_buffer_flag #undef _ s = format (s, "current data %d, length %d, buffer-pool %d, " - "clone-count %u", b->current_data, b->current_length, - b->buffer_pool_index, b->n_add_refs); + "ref-count %u", b->current_data, b->current_length, + b->buffer_pool_index, b->ref_count); if (b->flags & VLIB_BUFFER_TOTAL_LENGTH_VALID) s = format (s, ", totlen-nifb %d", @@ -107,9 +113,9 @@ format_vlib_buffer (u8 * s, va_list * args) b = vlib_get_buffer (vm, next_buffer); s = - format (s, "\n%Unext-buffer 0x%x, segment length %d, clone-count %u", + format (s, "\n%Unext-buffer 0x%x, segment length %d, ref-count %u", format_white_space, indent, next_buffer, b->current_length, - b->n_add_refs); + b->ref_count); } return s; @@ -177,14 +183,18 @@ vlib_validate_buffer_helper (vlib_main_t * vm, u32 bi, uword follow_buffer_next, uword ** unique_hash) { + vlib_buffer_main_t *bm = vm->buffer_main; vlib_buffer_t *b = vlib_get_buffer (vm, bi); + if (vec_len (bm->buffer_pools) <= b->buffer_pool_index) + return format (0, "unknown buffer pool 0x%x", b->buffer_pool_index); + if ((signed) b->current_data < (signed) -VLIB_BUFFER_PRE_DATA_SIZE) return format (0, "current data %d before pre-data", b->current_data); if (b->current_data + b->current_length > VLIB_BUFFER_DATA_SIZE) - return format (0, "%d-%d beyond end of buffer %d", - b->current_data, b->current_length, VLIB_BUFFER_DATA_SIZE); + return format (0, "%d-%d beyond end of buffer %d", b->current_data, + b->current_length, VLIB_BUFFER_DATA_SIZE); if (follow_buffer_next && (b->flags & VLIB_BUFFER_NEXT_PRESENT)) { @@ -306,15 +316,13 @@ vlib_buffer_validate_alloc_free (vlib_main_t * vm, uword n_buffers, vlib_buffer_known_state_t expected_state) { + vlib_buffer_main_t *bm = vm->buffer_main; u32 *b; uword i, bi, is_free; if (CLIB_DEBUG == 0) return; - if (vlib_buffer_callbacks) - return; - is_free = expected_state == VLIB_BUFFER_KNOWN_ALLOCATED; b = buffers; for (i = 0; i < n_buffers; i++) @@ -324,287 +332,26 @@ vlib_buffer_validate_alloc_free (vlib_main_t * vm, bi = b[0]; b += 1; known = vlib_buffer_is_known (vm, bi); + + if (known == VLIB_BUFFER_UNKNOWN && + expected_state == VLIB_BUFFER_KNOWN_FREE) + known = VLIB_BUFFER_KNOWN_FREE; + if (known != expected_state) { - ASSERT (0); vlib_panic_with_msg (vm, "%s %U buffer 0x%x", is_free ? "freeing" : "allocating", format_vlib_buffer_known_state, known, bi); } - vlib_buffer_set_known_state (vm, bi, is_free ? VLIB_BUFFER_KNOWN_FREE : - VLIB_BUFFER_KNOWN_ALLOCATED); - } -} - -/* Add buffer free list. */ -static vlib_buffer_free_list_index_t -vlib_buffer_create_free_list_helper (vlib_main_t * vm, - u32 n_data_bytes, - u32 is_public, u32 is_default, u8 * name) -{ - vlib_buffer_main_t *bm = vm->buffer_main; - vlib_buffer_free_list_t *f; - int i; - - ASSERT (vlib_get_thread_index () == 0); - - if (!is_default && pool_elts (vm->buffer_free_list_pool) == 0) - { - vlib_buffer_free_list_index_t default_free_free_list_index; - - /* *INDENT-OFF* */ - default_free_free_list_index = - vlib_buffer_create_free_list_helper - (vm, - /* default buffer size */ VLIB_BUFFER_DATA_SIZE, - /* is_public */ 1, - /* is_default */ 1, - (u8 *) "default"); - /* *INDENT-ON* */ - ASSERT (default_free_free_list_index == - VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); - - if (n_data_bytes == VLIB_BUFFER_DATA_SIZE && is_public) - return default_free_free_list_index; - } - - pool_get_aligned (vm->buffer_free_list_pool, f, CLIB_CACHE_LINE_BYTES); - - clib_memset (f, 0, sizeof (f[0])); - f->index = f - vm->buffer_free_list_pool; - vec_validate (f->buffers, 0); - vec_reset_length (f->buffers); - f->min_n_buffers_each_alloc = VLIB_FRAME_SIZE; - f->buffer_pool_index = 0; - f->name = clib_mem_is_vec (name) ? name : format (0, "%s", name); - - /* Setup free buffer template. */ - f->buffer_init_template.n_add_refs = 0; - - if (is_public) - { - uword *p = hash_get (bm->free_list_by_size, VLIB_BUFFER_DATA_SIZE); - if (!p) - hash_set (bm->free_list_by_size, VLIB_BUFFER_DATA_SIZE, f->index); - } - - for (i = 1; i < vec_len (vlib_mains); i++) - { - vlib_main_t *wvm = vlib_mains[i]; - vlib_buffer_free_list_t *wf; - pool_get_aligned (wvm->buffer_free_list_pool, - wf, CLIB_CACHE_LINE_BYTES); - ASSERT (f - vm->buffer_free_list_pool == - wf - wvm->buffer_free_list_pool); - wf[0] = f[0]; - wf->buffers = 0; - vec_validate (wf->buffers, 0); - vec_reset_length (wf->buffers); - wf->n_alloc = 0; - } - - return f->index; -} - -vlib_buffer_free_list_index_t -vlib_buffer_create_free_list (vlib_main_t * vm, u32 n_data_bytes, - char *fmt, ...) -{ - va_list va; - u8 *name; - - va_start (va, fmt); - name = va_format (0, fmt, &va); - va_end (va); - - return vlib_buffer_create_free_list_helper (vm, n_data_bytes, - /* is_public */ 0, - /* is_default */ 0, - name); -} - -static_always_inline void * -vlib_buffer_pool_get_buffer (vlib_main_t * vm, vlib_buffer_pool_t * bp) -{ - return vlib_physmem_alloc_from_map (vm, bp->physmem_map_index, - bp->buffer_size, CLIB_CACHE_LINE_BYTES); -} - -/* Make sure free list has at least given number of free buffers. */ -static uword -vlib_buffer_fill_free_list_internal (vlib_main_t * vm, - vlib_buffer_free_list_t * fl, - uword min_free_buffers) -{ - vlib_buffer_t *b; - vlib_buffer_pool_t *bp = vlib_buffer_pool_get (vm, fl->buffer_pool_index); - int n; - u32 *bi; - u32 n_alloc = 0; - - /* Already have enough free buffers on free list? */ - n = min_free_buffers - vec_len (fl->buffers); - if (n <= 0) - return min_free_buffers; - - if (vec_len (bp->buffers) > 0) - { - int n_copy, n_left; - clib_spinlock_lock (&bp->lock); - n_copy = clib_min (vec_len (bp->buffers), n); - n_left = vec_len (bp->buffers) - n_copy; - vec_add_aligned (fl->buffers, bp->buffers + n_left, n_copy, - CLIB_CACHE_LINE_BYTES); - _vec_len (bp->buffers) = n_left; - clib_spinlock_unlock (&bp->lock); - n = min_free_buffers - vec_len (fl->buffers); - if (n <= 0) - return min_free_buffers; - } - - /* Always allocate round number of buffers. */ - n = round_pow2 (n, CLIB_CACHE_LINE_BYTES / sizeof (u32)); - - /* Always allocate new buffers in reasonably large sized chunks. */ - n = clib_max (n, fl->min_n_buffers_each_alloc); - - clib_spinlock_lock (&bp->lock); - while (n_alloc < n) - { - if ((b = vlib_buffer_pool_get_buffer (vm, bp)) == 0) - goto done; - - n_alloc += 1; - - vec_add2_aligned (fl->buffers, bi, 1, CLIB_CACHE_LINE_BYTES); - bi[0] = vlib_get_buffer_index (vm, b); - - if (CLIB_DEBUG > 0) - vlib_buffer_set_known_state (vm, bi[0], VLIB_BUFFER_KNOWN_FREE); - - clib_memset (b, 0, sizeof (vlib_buffer_t)); - - if (fl->buffer_init_function) - fl->buffer_init_function (vm, fl, bi, 1); - } - -done: - clib_spinlock_unlock (&bp->lock); - fl->n_alloc += n_alloc; - return n_alloc; -} - -void * -vlib_set_buffer_free_callback (vlib_main_t * vm, void *fp) -{ - vlib_buffer_main_t *bm = vm->buffer_main; - void *rv = bm->buffer_free_callback; - - bm->buffer_free_callback = fp; - return rv; -} - -static_always_inline void -recycle_or_free (vlib_main_t * vm, vlib_buffer_main_t * bm, u32 bi, - vlib_buffer_t * b, u32 follow_buffer_next) -{ - vlib_buffer_free_list_t *fl; - u32 flags, next; - - fl = pool_elt_at_index (vm->buffer_free_list_pool, - VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); - - do - { - vlib_buffer_t *nb = vlib_get_buffer (vm, bi); - flags = nb->flags; - next = nb->next_buffer; - if (nb->n_add_refs) - nb->n_add_refs--; - else - { - vlib_buffer_validate_alloc_free (vm, &bi, 1, - VLIB_BUFFER_KNOWN_ALLOCATED); - vlib_buffer_add_to_free_list (vm, fl, bi, 1); - } - bi = next; - } - while (follow_buffer_next && (flags & VLIB_BUFFER_NEXT_PRESENT)); -} - -static_always_inline void -vlib_buffer_free_inline (vlib_main_t * vm, - u32 * buffers, u32 n_buffers, u32 follow_buffer_next) -{ - vlib_buffer_main_t *bm = vm->buffer_main; - vlib_buffer_t *p, *b0, *b1, *b2, *b3; - int i = 0; - u32 (*cb) (vlib_main_t * vm, u32 * buffers, u32 n_buffers, - u32 follow_buffer_next); - - cb = bm->buffer_free_callback; - - if (PREDICT_FALSE (cb != 0)) - n_buffers = (*cb) (vm, buffers, n_buffers, follow_buffer_next); - - if (!n_buffers) - return; - - while (i + 11 < n_buffers) - { - p = vlib_get_buffer (vm, buffers[i + 8]); - vlib_prefetch_buffer_header (p, LOAD); - p = vlib_get_buffer (vm, buffers[i + 9]); - vlib_prefetch_buffer_header (p, LOAD); - p = vlib_get_buffer (vm, buffers[i + 10]); - vlib_prefetch_buffer_header (p, LOAD); - p = vlib_get_buffer (vm, buffers[i + 11]); - vlib_prefetch_buffer_header (p, LOAD); - - b0 = vlib_get_buffer (vm, buffers[i]); - b1 = vlib_get_buffer (vm, buffers[i + 1]); - b2 = vlib_get_buffer (vm, buffers[i + 2]); - b3 = vlib_get_buffer (vm, buffers[i + 3]); - - VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); - VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b1); - VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b2); - VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b3); - - recycle_or_free (vm, bm, buffers[i], b0, follow_buffer_next); - recycle_or_free (vm, bm, buffers[i + 1], b1, follow_buffer_next); - recycle_or_free (vm, bm, buffers[i + 2], b2, follow_buffer_next); - recycle_or_free (vm, bm, buffers[i + 3], b3, follow_buffer_next); - - i += 4; - } - - while (i < n_buffers) - { - b0 = vlib_get_buffer (vm, buffers[i]); - VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); - recycle_or_free (vm, bm, buffers[i], b0, follow_buffer_next); - i++; + clib_spinlock_lock (&bm->buffer_known_hash_lockp); + hash_set (bm->buffer_known_hash, bi, is_free ? VLIB_BUFFER_KNOWN_FREE : + VLIB_BUFFER_KNOWN_ALLOCATED); + clib_spinlock_unlock (&bm->buffer_known_hash_lockp); } } -static void -vlib_buffer_free_internal (vlib_main_t * vm, u32 * buffers, u32 n_buffers) -{ - vlib_buffer_free_inline (vm, buffers, n_buffers, /* follow_buffer_next */ - 1); -} - -static void -vlib_buffer_free_no_next_internal (vlib_main_t * vm, u32 * buffers, - u32 n_buffers) -{ - vlib_buffer_free_inline (vm, buffers, n_buffers, /* follow_buffer_next */ - 0); -} - void vlib_packet_template_init (vlib_main_t * vm, vlib_packet_template_t * t, @@ -624,7 +371,6 @@ vlib_packet_template_init (vlib_main_t * vm, vec_add (t->packet_data, packet_data, n_packet_data_bytes); t->min_n_buffers_each_alloc = min_n_buffers_each_alloc; - vlib_worker_thread_barrier_release (vm); } @@ -706,8 +452,6 @@ out_of_buffers: u16 vlib_buffer_chain_append_data_with_alloc (vlib_main_t * vm, - vlib_buffer_free_list_index_t - free_list_index, vlib_buffer_t * first, vlib_buffer_t ** last, void *data, u16 data_len) @@ -721,9 +465,8 @@ vlib_buffer_chain_append_data_with_alloc (vlib_main_t * vm, u16 max = n_buffer_bytes - l->current_length - l->current_data; if (max == 0) { - if (1 != - vlib_buffer_alloc_from_free_list (vm, &l->next_buffer, 1, - free_list_index)) + if (1 != vlib_buffer_alloc_from_pool (vm, &l->next_buffer, 1, + first->buffer_pool_index)) return copied; *last = l = vlib_buffer_chain_buffer (vm, l, l->next_buffer); max = n_buffer_bytes - l->current_length - l->current_data; @@ -739,14 +482,26 @@ vlib_buffer_chain_append_data_with_alloc (vlib_main_t * vm, return copied; } -u8 -vlib_buffer_register_physmem_map (vlib_main_t * vm, u32 physmem_map_index) +clib_error_t * +vlib_buffer_pool_create (vlib_main_t * vm, u8 index, char *name, + u32 data_size, u32 physmem_map_index) { vlib_buffer_main_t *bm = vm->buffer_main; - vlib_buffer_pool_t *p; + vlib_buffer_pool_t *bp; vlib_physmem_map_t *m = vlib_physmem_get_map (vm, physmem_map_index); uword start = pointer_to_uword (m->base); uword size = (uword) m->n_pages << m->log2_page_size; + u32 alloc_size, n_alloc_per_page;; + + vec_validate_aligned (bm->buffer_pools, index, CLIB_CACHE_LINE_BYTES); + bp = vec_elt_at_index (bm->buffer_pools, index); + + if (bp->start) + return clib_error_return (0, "buffer with index %u already exists", + index); + + if (index >= 255) + return clib_error_return (0, "buffer index must be < 255", index); if (bm->buffer_mem_size == 0) { @@ -773,36 +528,77 @@ vlib_buffer_register_physmem_map (vlib_main_t * vm, u32 physmem_map_index) clib_panic ("buffer memory size out of range!"); } - vec_add2 (bm->buffer_pools, p, 1); - p->start = start; - p->size = size; - p->physmem_map_index = physmem_map_index; + bp->start = start; + bp->size = size; + bp->index = bp - bm->buffer_pools; + bp->buffer_template.buffer_pool_index = bp->index; + bp->buffer_template.ref_count = 1; + bp->physmem_map_index = physmem_map_index; + bp->name = format (0, "%s%c", name, 0); + bp->data_size = data_size; + bp->numa_node = m->numa_node; - ASSERT (p - bm->buffer_pools < 256); - return p - bm->buffer_pools; -} + vec_validate_aligned (bp->threads, vec_len (vlib_mains) - 1, + CLIB_CACHE_LINE_BYTES); -static u8 * -format_vlib_buffer_free_list (u8 * s, va_list * va) -{ - vlib_buffer_free_list_t *f = va_arg (*va, vlib_buffer_free_list_t *); - u32 threadnum = va_arg (*va, u32); - uword bytes_alloc, bytes_free, n_free, size; + alloc_size = data_size + sizeof (vlib_buffer_t) + bm->ext_hdr_size; + n_alloc_per_page = (1ULL << m->log2_page_size) / alloc_size; + + /* preallocate buffer indices memory */ + vec_validate_aligned (bp->buffers, m->n_pages * n_alloc_per_page, + CLIB_CACHE_LINE_BYTES); + vec_reset_length (bp->buffers); + + clib_spinlock_init (&bp->lock); + + while (1) + { + u8 *p; + u32 bi; + + p = vlib_physmem_alloc_from_map (vm, bp->physmem_map_index, alloc_size, + CLIB_CACHE_LINE_BYTES); + + if (p == 0) + break; + + p += bm->ext_hdr_size; + + vlib_buffer_copy_template ((vlib_buffer_t *) p, &bp->buffer_template); - if (!f) - return format (s, "%=7s%=30s%=12s%=12s%=12s%=12s%=12s%=12s", - "Thread", "Name", "Index", "Size", "Alloc", "Free", - "#Alloc", "#Free"); + bi = vlib_get_buffer_index (vm, (vlib_buffer_t *) p); - size = sizeof (vlib_buffer_t) + VLIB_BUFFER_DATA_SIZE; - n_free = vec_len (f->buffers); - bytes_alloc = size * f->n_alloc; - bytes_free = size * n_free; + vec_add1_aligned (bp->buffers, bi, CLIB_CACHE_LINE_BYTES); + + bp->n_buffers += 1; + } + + return 0; +} - s = format (s, "%7d%30v%12d%12d%=12U%=12U%=12d%=12d", threadnum, - f->name, f->index, VLIB_BUFFER_DATA_SIZE, - format_memory_size, bytes_alloc, - format_memory_size, bytes_free, f->n_alloc, n_free); +static u8 * +format_vlib_buffer_pool (u8 * s, va_list * va) +{ + vlib_main_t *vm = va_arg (*va, vlib_main_t *); + vlib_buffer_pool_t *bp = va_arg (*va, vlib_buffer_pool_t *); + vlib_buffer_pool_thread_t *bpt; + u32 cached = 0; + + if (!bp) + return format (s, "%-20s%=6s%=6s%=6s%=11s%=6s%=8s%=8s%=8s", + "Pool Name", "Index", "NUMA", "Size", "Data Size", + "Total", "Avail", "Cached", "Used"); + + /* *INDENT-OFF* */ + vec_foreach (bpt, bp->threads) + cached += vec_len (bpt->cached_buffers); + /* *INDENT-ON* */ + + s = format (s, "%-20s%=6d%=6d%=6u%=11u%=6u%=8u%=8u%=8u", + bp->name, bp->index, bp->numa_node, bp->data_size + + sizeof (vlib_buffer_t) + vm->buffer_main->ext_hdr_size, + bp->data_size, bp->n_buffers, vec_len (bp->buffers), cached, + bp->n_buffers - vec_len (bp->buffers) - cached); return s; } @@ -811,25 +607,15 @@ static clib_error_t * show_buffers (vlib_main_t * vm, unformat_input_t * input, vlib_cli_command_t * cmd) { - vlib_buffer_free_list_t *f; - vlib_main_t *curr_vm; - u32 vm_index = 0; - - vlib_cli_output (vm, "%U", format_vlib_buffer_free_list, 0, 0); - - do - { - curr_vm = vlib_mains[vm_index]; + vlib_buffer_main_t *bm = vm->buffer_main; + vlib_buffer_pool_t *bp; - /* *INDENT-OFF* */ - pool_foreach (f, curr_vm->buffer_free_list_pool, ({ - vlib_cli_output (vm, "%U", format_vlib_buffer_free_list, f, vm_index); - })); - /* *INDENT-ON* */ + vlib_cli_output (vm, "%U", format_vlib_buffer_pool, vm, 0); - vm_index++; - } - while (vm_index < vec_len (vlib_mains)); + /* *INDENT-OFF* */ + vec_foreach (bp, bm->buffer_pools) + vlib_cli_output (vm, "%U", format_vlib_buffer_pool, vm, bp); + /* *INDENT-ON* */ return 0; } @@ -843,71 +629,134 @@ VLIB_CLI_COMMAND (show_buffers_command, static) = { /* *INDENT-ON* */ clib_error_t * -vlib_buffer_main_init (struct vlib_main_t * vm) +vlib_buffer_worker_init (vlib_main_t * vm) { - vlib_buffer_main_t *bm; - clib_error_t *error; - u32 physmem_map_index; - u8 pool_index; - int log2_page_size = 0; - - buffer_log_default = vlib_log_register_class ("buffer", 0); - - bm = vm->buffer_main = clib_mem_alloc (sizeof (bm[0])); - clib_memset (bm, 0, sizeof (bm[0])); + vlib_buffer_main_t *bm = vm->buffer_main; + vlib_buffer_pool_t *bp; - if (vlib_buffer_callbacks) + /* *INDENT-OFF* */ + vec_foreach (bp, bm->buffer_pools) { - /* external plugin has registered own buffer callbacks - so we just copy them and quit */ - clib_memcpy_fast (&bm->cb, vlib_buffer_callbacks, - sizeof (vlib_buffer_callbacks_t)); - bm->callbacks_registered = 1; - return 0; + clib_spinlock_lock (&bp->lock); + vec_validate_aligned (bp->threads, vec_len (vlib_mains) - 1, + CLIB_CACHE_LINE_BYTES); + clib_spinlock_unlock (&bp->lock); } + /* *INDENT-ON* */ - bm->cb.vlib_buffer_fill_free_list_cb = &vlib_buffer_fill_free_list_internal; - bm->cb.vlib_buffer_free_cb = &vlib_buffer_free_internal; - bm->cb.vlib_buffer_free_no_next_cb = &vlib_buffer_free_no_next_internal; - clib_spinlock_init (&bm->buffer_known_hash_lockp); + return 0; +} + +VLIB_WORKER_INIT_FUNCTION (vlib_buffer_worker_init); + +static clib_error_t * +vlib_buffer_main_init_numa_node (struct vlib_main_t *vm, u32 numa_node) +{ + vlib_buffer_main_t *bm = vm->buffer_main; + clib_error_t *error; + u32 physmem_map_index; + uword n_pages, pagesize; + u32 buffers_per_numa; + u32 buffer_size = CLIB_CACHE_LINE_ROUND (bm->ext_hdr_size + + sizeof (vlib_buffer_t) + + VLIB_BUFFER_DATA_SIZE); + u8 *name; + + pagesize = clib_mem_get_default_hugepage_size (); + name = format (0, "buffers-numa-%d%c", numa_node, 0); + + buffers_per_numa = bm->buffers_per_numa ? bm->buffers_per_numa : + VLIB_BUFFER_DEFAULT_BUFFERS_PER_NUMA; retry: - error = vlib_physmem_shared_map_create (vm, "buffers", - vlib_buffer_physmem_sz, - log2_page_size, - CLIB_PMALLOC_NUMA_LOCAL, + n_pages = (buffers_per_numa - 1) / (pagesize / buffer_size) + 1; + error = vlib_physmem_shared_map_create (vm, (char *) name, + n_pages * pagesize, + min_log2 (pagesize), numa_node, &physmem_map_index); - if (error && log2_page_size == 0) + if (error && pagesize != clib_mem_get_page_size ()) { - vlib_log_warn (buffer_log_default, "%U", format_clib_error, error); + vlib_log_warn (bm->log_default, "%U", format_clib_error, error); clib_error_free (error); - vlib_log_warn (buffer_log_default, "falling back to non-hugepage " + vlib_log_warn (bm->log_default, "falling back to non-hugepage " "backed buffer pool"); - log2_page_size = min_log2 (clib_mem_get_page_size ()); + pagesize = clib_mem_get_page_size (); + buffers_per_numa = bm->buffers_per_numa ? bm->buffers_per_numa : + VLIB_BUFFER_DEFAULT_BUFFERS_PER_NUMA_UNPRIV; goto retry; } if (error) return error; - pool_index = vlib_buffer_register_physmem_map (vm, physmem_map_index); - vlib_buffer_pool_t *bp = vlib_buffer_pool_get (vm, pool_index); - clib_spinlock_init (&bp->lock); - bp->buffer_size = VLIB_BUFFER_DATA_SIZE + sizeof (vlib_buffer_t); + vec_reset_length (name); + name = format (name, "default-numa-%d%c", numa_node, 0); - return 0; + return vlib_buffer_pool_create (vm, numa_node, (char *) name, + VLIB_BUFFER_DATA_SIZE, physmem_map_index); +} + +clib_error_t * +vlib_buffer_main_init (struct vlib_main_t *vm) +{ + vlib_buffer_main_t *bm; + clib_error_t *err; + clib_bitmap_t *bmp = 0; + u32 numa_node; + + if (vm->buffer_main == 0) + { + vm->buffer_main = clib_mem_alloc (sizeof (bm[0])); + clib_memset (vm->buffer_main, 0, sizeof (bm[0])); + } + + bm = vm->buffer_main; + bm->log_default = vlib_log_register_class ("buffer", 0); + bm->ext_hdr_size = __vlib_buffer_external_hdr_size; + + clib_spinlock_init (&bm->buffer_known_hash_lockp); + + err = clib_sysfs_read ("/sys/devices/system/node/possible", "%U", + unformat_bitmap_list, &bmp); + if (err) + { + /* no info from sysfs, assuming that only numa 0 exists */ + clib_error_free (err); + bmp = clib_bitmap_set (bmp, 0, 1); + } + + /* *INDENT-OFF* */ + clib_bitmap_foreach (numa_node, bmp, { + if ((err = vlib_buffer_main_init_numa_node(vm, numa_node))) + goto done; + }); + /* *INDENT-ON* */ + + bm->n_numa_nodes = clib_bitmap_last_set (bmp) + 1; + +done: + vec_free (bmp); + return err; } static clib_error_t * vlib_buffers_configure (vlib_main_t * vm, unformat_input_t * input) { - u32 size_in_mb; + vlib_buffer_main_t *bm; + + if (vm->buffer_main == 0) + { + vm->buffer_main = clib_mem_alloc (sizeof (bm[0])); + clib_memset (vm->buffer_main, 0, sizeof (bm[0])); + } + + bm = vm->buffer_main; while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) { - if (unformat (input, "memory-size-in-mb %d", &size_in_mb)) - vlib_buffer_physmem_sz = size_in_mb << 20; + if (unformat (input, "buffers-per-numa %u", &bm->buffers_per_numa)) + ; else return unformat_parse_error (input); } diff --git a/src/vlib/buffer.h b/src/vlib/buffer.h index f7fdf208a83..1adde73d4ee 100644 --- a/src/vlib/buffer.h +++ b/src/vlib/buffer.h @@ -59,8 +59,6 @@ /* Amount of head buffer data copied to each replica head buffer */ #define VLIB_BUFFER_CLONE_HEAD_SIZE (256) -typedef u8 vlib_buffer_free_list_index_t; - /** \file vlib buffer structure definition and a few select access methods. This structure and the buffer allocation @@ -127,8 +125,8 @@ typedef union /** Generic flow identifier */ u32 flow_id; - /** Number of additional references to this buffer. */ - u8 n_add_refs; + /** Reference count for this buffer. */ + volatile u8 ref_count; /** index of buffer pool this buffer belongs. */ u8 buffer_pool_index; @@ -367,66 +365,32 @@ vlib_buffer_pull (vlib_buffer_t * b, u8 size) /* Forward declaration. */ struct vlib_main_t; -typedef struct vlib_buffer_free_list_t -{ - /* Template buffer used to initialize first 16 bytes of buffers - allocated on this free list. */ - vlib_buffer_t buffer_init_template; - - /* Our index into vlib_main_t's buffer_free_list_pool. */ - vlib_buffer_free_list_index_t index; - - /* Number of buffers to allocate when we need to allocate new buffers */ - u32 min_n_buffers_each_alloc; - - /* Total number of buffers allocated from this free list. */ - u32 n_alloc; - - /* Vector of free buffers. Each element is a byte offset into I/O heap. */ - u32 *buffers; - - /* index of buffer pool used to get / put buffers */ - u8 buffer_pool_index; - - /* Free list name. */ - u8 *name; - - /* Callback functions to initialize newly allocated buffers. - If null buffers are zeroed. */ - void (*buffer_init_function) (struct vlib_main_t * vm, - struct vlib_buffer_free_list_t * fl, - u32 * buffers, u32 n_buffers); - - uword buffer_init_function_opaque; -} __attribute__ ((aligned (16))) vlib_buffer_free_list_t; - -typedef uword (vlib_buffer_fill_free_list_cb_t) (struct vlib_main_t * vm, - vlib_buffer_free_list_t * fl, - uword min_free_buffers); -typedef void (vlib_buffer_free_cb_t) (struct vlib_main_t * vm, u32 * buffers, - u32 n_buffers); -typedef void (vlib_buffer_free_no_next_cb_t) (struct vlib_main_t * vm, - u32 * buffers, u32 n_buffers); - typedef struct { - vlib_buffer_fill_free_list_cb_t *vlib_buffer_fill_free_list_cb; - vlib_buffer_free_cb_t *vlib_buffer_free_cb; - vlib_buffer_free_no_next_cb_t *vlib_buffer_free_no_next_cb; -} vlib_buffer_callbacks_t; - -extern vlib_buffer_callbacks_t *vlib_buffer_callbacks; - + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + u32 *cached_buffers; + u32 n_alloc; +} vlib_buffer_pool_thread_t; typedef struct { CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); uword start; uword size; uword log2_page_size; + u8 index; + u32 numa_node; u32 physmem_map_index; - u32 buffer_size; + u32 data_size; + u32 n_buffers; u32 *buffers; + u8 *name; clib_spinlock_t lock; + + /* per-thread data */ + vlib_buffer_pool_thread_t *threads; + + /* buffer metadata template */ + vlib_buffer_t buffer_template; } vlib_buffer_pool_t; typedef struct @@ -438,36 +402,24 @@ typedef struct uword buffer_mem_size; vlib_buffer_pool_t *buffer_pools; - /* Buffer free callback, for subversive activities */ - u32 (*buffer_free_callback) (struct vlib_main_t * vm, - u32 * buffers, - u32 n_buffers, u32 follow_buffer_next); -#define VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX (0) - - /* Hash table mapping buffer size (rounded to next unit of - sizeof (vlib_buffer_t)) to free list index. */ - uword *free_list_by_size; - /* Hash table mapping buffer index into number 0 => allocated but free, 1 => allocated and not-free. If buffer index is not in hash table then this buffer has never been allocated. */ uword *buffer_known_hash; clib_spinlock_t buffer_known_hash_lockp; + u32 n_numa_nodes; - /* Callbacks */ - vlib_buffer_callbacks_t cb; - int callbacks_registered; -} vlib_buffer_main_t; + /* config */ + u32 buffers_per_numa; + u16 ext_hdr_size; -u8 vlib_buffer_register_physmem_map (struct vlib_main_t *vm, - u32 physmem_map_index); + /* logging */ + vlib_log_class_t log_default; +} vlib_buffer_main_t; clib_error_t *vlib_buffer_main_init (struct vlib_main_t *vm); - -void *vlib_set_buffer_free_callback (struct vlib_main_t *vm, void *fp); - /* */ @@ -488,23 +440,17 @@ extern void vlib_buffer_trace_trajectory_init (vlib_buffer_t * b); #define VLIB_BUFFER_TRACE_TRAJECTORY_INIT(b) #endif /* VLIB_BUFFER_TRACE_TRAJECTORY */ -#endif /* included_vlib_buffer_h */ +extern u16 __vlib_buffer_external_hdr_size; +#define VLIB_BUFFER_SET_EXT_HDR_SIZE(x) \ +static void __clib_constructor \ +vnet_buffer_set_ext_hdr_size() \ +{ \ + if (__vlib_buffer_external_hdr_size) \ + clib_error ("buffer external header space already set"); \ + __vlib_buffer_external_hdr_size = CLIB_CACHE_LINE_ROUND (x); \ +} -#define VLIB_BUFFER_REGISTER_CALLBACKS(x,...) \ - __VA_ARGS__ vlib_buffer_callbacks_t __##x##_buffer_callbacks; \ -static void __vlib_add_buffer_callbacks_t_##x (void) \ - __attribute__((__constructor__)) ; \ -static void __vlib_add_buffer_callbacks_t_##x (void) \ -{ \ - if (vlib_buffer_callbacks) \ - clib_panic ("vlib buffer callbacks already registered"); \ - vlib_buffer_callbacks = &__##x##_buffer_callbacks; \ -} \ -static void __vlib_rm_buffer_callbacks_t_##x (void) \ - __attribute__((__destructor__)) ; \ -static void __vlib_rm_buffer_callbacks_t_##x (void) \ -{ vlib_buffer_callbacks = 0; } \ -__VA_ARGS__ vlib_buffer_callbacks_t __##x##_buffer_callbacks +#endif /* included_vlib_buffer_h */ /* * fd.io coding-style-patch-verification: ON diff --git a/src/vlib/buffer_funcs.h b/src/vlib/buffer_funcs.h index cedd7d02646..ce8d1ef9078 100644 --- a/src/vlib/buffer_funcs.h +++ b/src/vlib/buffer_funcs.h @@ -89,6 +89,13 @@ vlib_buffer_copy_template (vlib_buffer_t * b, vlib_buffer_t * bt) #endif } +always_inline u8 +vlib_buffer_pool_get_default_for_numa (vlib_main_t * vm, u32 numa_node) +{ + ASSERT (numa_node < vm->buffer_main->n_numa_nodes); + return numa_node; +} + /** \brief Translate array of buffer indices into buffer pointers with offset @param vm - (vlib_main_t *) vlib main data structure pointer @@ -394,29 +401,48 @@ vlib_buffer_is_known (vlib_main_t * vm, u32 buffer_index) return p ? p[0] : VLIB_BUFFER_UNKNOWN; } -always_inline void -vlib_buffer_set_known_state (vlib_main_t * vm, u32 buffer_index, - vlib_buffer_known_state_t state) -{ - vlib_buffer_main_t *bm = vm->buffer_main; - - clib_spinlock_lock (&bm->buffer_known_hash_lockp); - hash_set (bm->buffer_known_hash, buffer_index, state); - clib_spinlock_unlock (&bm->buffer_known_hash_lockp); -} - /* Validates sanity of a single buffer. Returns format'ed vector with error message if any. */ u8 *vlib_validate_buffer (vlib_main_t * vm, u32 buffer_index, uword follow_chain); -always_inline u32 -vlib_buffer_round_size (u32 size) +static_always_inline vlib_buffer_pool_t * +vlib_get_buffer_pool (vlib_main_t * vm, u8 buffer_pool_index) { - return round_pow2 (size, sizeof (vlib_buffer_t)); + vlib_buffer_main_t *bm = vm->buffer_main; + return vec_elt_at_index (bm->buffer_pools, buffer_pool_index); } -/** \brief Allocate buffers from specific freelist into supplied array +static_always_inline uword +vlib_buffer_pool_get (vlib_main_t * vm, u8 buffer_pool_index, u32 * buffers, + u32 n_buffers) +{ + vlib_buffer_pool_t *bp = vlib_get_buffer_pool (vm, buffer_pool_index); + u32 len; + + ASSERT (bp->buffers); + + clib_spinlock_lock (&bp->lock); + len = vec_len (bp->buffers); + if (PREDICT_TRUE (n_buffers < len)) + { + len -= n_buffers; + vlib_buffer_copy_indices (buffers, bp->buffers + len, n_buffers); + _vec_len (bp->buffers) = len; + clib_spinlock_unlock (&bp->lock); + return n_buffers; + } + else + { + vlib_buffer_copy_indices (buffers, bp->buffers, len); + _vec_len (bp->buffers) = 0; + clib_spinlock_unlock (&bp->lock); + return len; + } +} + + +/** \brief Allocate buffers from specific pool into supplied array @param vm - (vlib_main_t *) vlib main data structure pointer @param buffers - (u32 * ) buffer index array @@ -424,57 +450,87 @@ vlib_buffer_round_size (u32 size) @return - (u32) number of buffers actually allocated, may be less than the number requested or zero */ + always_inline u32 -vlib_buffer_alloc_from_free_list (vlib_main_t * vm, - u32 * buffers, - u32 n_buffers, - vlib_buffer_free_list_index_t index) +vlib_buffer_alloc_from_pool (vlib_main_t * vm, u32 * buffers, u32 n_buffers, + u8 buffer_pool_index) { - vlib_buffer_free_list_t *fl; vlib_buffer_main_t *bm = vm->buffer_main; - u32 *src; - uword len; - - ASSERT (bm->cb.vlib_buffer_fill_free_list_cb); + vlib_buffer_pool_t *bp; + vlib_buffer_pool_thread_t *bpt; + u32 *src, *dst, len, n_left; - fl = pool_elt_at_index (vm->buffer_free_list_pool, index); + bp = vec_elt_at_index (bm->buffer_pools, buffer_pool_index); + bpt = vec_elt_at_index (bp->threads, vm->thread_index); - len = vec_len (fl->buffers); + dst = buffers; + n_left = n_buffers; + len = vec_len (bpt->cached_buffers); - if (PREDICT_FALSE (len < n_buffers)) + /* per-thread cache contains enough buffers */ + if (len >= n_buffers) { - bm->cb.vlib_buffer_fill_free_list_cb (vm, fl, n_buffers); - if (PREDICT_FALSE ((len = vec_len (fl->buffers)) == 0)) - return 0; + src = bpt->cached_buffers + len - n_buffers; + vlib_buffer_copy_indices (dst, src, n_buffers); + _vec_len (bpt->cached_buffers) -= n_buffers; - /* even if fill free list didn't manage to refill free list - we should give what we have */ - n_buffers = clib_min (len, n_buffers); + if (CLIB_DEBUG > 0) + vlib_buffer_validate_alloc_free (vm, buffers, n_buffers, + VLIB_BUFFER_KNOWN_FREE); + return n_buffers; + } - /* following code is intentionaly duplicated to allow compiler - to optimize fast path when n_buffers is constant value */ - src = fl->buffers + len - n_buffers; - vlib_buffer_copy_indices (buffers, src, n_buffers); - _vec_len (fl->buffers) -= n_buffers; + /* take everything available in the cache */ + if (len) + { + vlib_buffer_copy_indices (dst, bpt->cached_buffers, len); + _vec_len (bpt->cached_buffers) = 0; + dst += len; + n_left -= len; + } - /* Verify that buffers are known free. */ - vlib_buffer_validate_alloc_free (vm, buffers, n_buffers, - VLIB_BUFFER_KNOWN_FREE); + len = round_pow2 (n_left, 32); + vec_validate_aligned (bpt->cached_buffers, len - 1, CLIB_CACHE_LINE_BYTES); + len = vlib_buffer_pool_get (vm, buffer_pool_index, bpt->cached_buffers, + len); + _vec_len (bpt->cached_buffers) = len; - return n_buffers; + if (len) + { + u32 n_copy = clib_min (len, n_left); + src = bpt->cached_buffers + len - n_copy; + vlib_buffer_copy_indices (dst, src, n_copy); + _vec_len (bpt->cached_buffers) -= n_copy; + n_left -= n_copy; } - src = fl->buffers + len - n_buffers; - vlib_buffer_copy_indices (buffers, src, n_buffers); - _vec_len (fl->buffers) -= n_buffers; + n_buffers -= n_left; /* Verify that buffers are known free. */ - vlib_buffer_validate_alloc_free (vm, buffers, n_buffers, - VLIB_BUFFER_KNOWN_FREE); + if (CLIB_DEBUG > 0) + vlib_buffer_validate_alloc_free (vm, buffers, n_buffers, + VLIB_BUFFER_KNOWN_FREE); return n_buffers; } +/** \brief Allocate buffers from specific numa node into supplied array + + @param vm - (vlib_main_t *) vlib main data structure pointer + @param buffers - (u32 * ) buffer index array + @param n_buffers - (u32) number of buffers requested + @param numa_node - (u32) numa node + @return - (u32) number of buffers actually allocated, may be + less than the number requested or zero +*/ +always_inline u32 +vlib_buffer_alloc_on_numa (vlib_main_t * vm, u32 * buffers, u32 n_buffers, + u32 numa_node) +{ + u8 index = vlib_buffer_pool_get_default_for_numa (vm, numa_node); + return vlib_buffer_alloc_from_pool (vm, buffers, n_buffers, index); +} + /** \brief Allocate buffers into supplied array @param vm - (vlib_main_t *) vlib main data structure pointer @@ -483,11 +539,11 @@ vlib_buffer_alloc_from_free_list (vlib_main_t * vm, @return - (u32) number of buffers actually allocated, may be less than the number requested or zero */ + always_inline u32 vlib_buffer_alloc (vlib_main_t * vm, u32 * buffers, u32 n_buffers) { - return vlib_buffer_alloc_from_free_list (vm, buffers, n_buffers, - VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + return vlib_buffer_alloc_on_numa (vm, buffers, n_buffers, vm->numa_node); } /** \brief Allocate buffers into ring @@ -519,6 +575,215 @@ vlib_buffer_alloc_to_ring (vlib_main_t * vm, u32 * ring, u32 start, return n_alloc; } +/** \brief Allocate buffers into ring from specific buffer pool + + @param vm - (vlib_main_t *) vlib main data structure pointer + @param buffers - (u32 * ) buffer index ring + @param start - (u32) first slot in the ring + @param ring_size - (u32) ring size + @param n_buffers - (u32) number of buffers requested + @return - (u32) number of buffers actually allocated, may be + less than the number requested or zero +*/ +always_inline u32 +vlib_buffer_alloc_to_ring_from_pool (vlib_main_t * vm, u32 * ring, u32 start, + u32 ring_size, u32 n_buffers, + u8 buffer_pool_index) +{ + u32 n_alloc; + + ASSERT (n_buffers <= ring_size); + + if (PREDICT_TRUE (start + n_buffers <= ring_size)) + return vlib_buffer_alloc_from_pool (vm, ring + start, n_buffers, + buffer_pool_index); + + n_alloc = vlib_buffer_alloc_from_pool (vm, ring + start, ring_size - start, + buffer_pool_index); + + if (PREDICT_TRUE (n_alloc == ring_size - start)) + n_alloc += vlib_buffer_alloc_from_pool (vm, ring, n_buffers - n_alloc, + buffer_pool_index); + + return n_alloc; +} + +static void +vlib_buffer_pool_put (vlib_main_t * vm, u8 buffer_pool_index, + u32 * buffers, u32 n_buffers) +{ + vlib_buffer_pool_t *bp = vlib_get_buffer_pool (vm, buffer_pool_index); + vlib_buffer_pool_thread_t *bpt = + vec_elt_at_index (bp->threads, vm->thread_index); + + vec_add_aligned (bpt->cached_buffers, buffers, n_buffers, + CLIB_CACHE_LINE_BYTES); + + if (vec_len (bpt->cached_buffers) > 4 * VLIB_FRAME_SIZE) + { + clib_spinlock_lock (&bp->lock); + /* keep last stored buffers, as they are more likely hot in the cache */ + vec_add_aligned (bp->buffers, bpt->cached_buffers, VLIB_FRAME_SIZE, + CLIB_CACHE_LINE_BYTES); + vec_delete (bpt->cached_buffers, VLIB_FRAME_SIZE, 0); + bpt->n_alloc -= VLIB_FRAME_SIZE; + clib_spinlock_unlock (&bp->lock); + } +} + +static_always_inline void +vlib_buffer_free_inline (vlib_main_t * vm, u32 * buffers, u32 n_buffers, + int maybe_next) +{ + const int queue_size = 128; + vlib_buffer_pool_t *bp = 0; + u8 buffer_pool_index = ~0; + u32 n_queue = 0, queue[queue_size + 4]; + vlib_buffer_t bt = { }; +#if defined(CLIB_HAVE_VEC128) && !__aarch64__ + vlib_buffer_t bpi_mask = {.buffer_pool_index = ~0 }; + vlib_buffer_t bpi_vec = {.buffer_pool_index = ~0 }; + vlib_buffer_t flags_refs_mask = { + .flags = VLIB_BUFFER_NEXT_PRESENT, + .ref_count = ~0 + }; +#endif + + while (n_buffers) + { + vlib_buffer_t *b[8]; + u32 bi, sum = 0, flags, next; + + if (n_buffers < 12) + goto one_by_one; + + vlib_get_buffers (vm, buffers, b, 4); + vlib_get_buffers (vm, buffers + 8, b + 4, 4); + + vlib_prefetch_buffer_header (b[4], LOAD); + vlib_prefetch_buffer_header (b[5], LOAD); + vlib_prefetch_buffer_header (b[6], LOAD); + vlib_prefetch_buffer_header (b[7], LOAD); + +#if defined(CLIB_HAVE_VEC128) && !__aarch64__ + u8x16 p0, p1, p2, p3, r; + p0 = u8x16_load_unaligned (b[0]); + p1 = u8x16_load_unaligned (b[1]); + p2 = u8x16_load_unaligned (b[2]); + p3 = u8x16_load_unaligned (b[3]); + + r = p0 ^ bpi_vec.as_u8x16[0]; + r |= p1 ^ bpi_vec.as_u8x16[0]; + r |= p2 ^ bpi_vec.as_u8x16[0]; + r |= p3 ^ bpi_vec.as_u8x16[0]; + r &= bpi_mask.as_u8x16[0]; + r |= (p0 | p1 | p2 | p3) & flags_refs_mask.as_u8x16[0]; + + sum = !u8x16_is_all_zero (r); +#else + sum |= b[0]->flags; + sum |= b[1]->flags; + sum |= b[2]->flags; + sum |= b[3]->flags; + sum &= VLIB_BUFFER_NEXT_PRESENT; + sum += b[0]->ref_count - 1; + sum += b[1]->ref_count - 1; + sum += b[2]->ref_count - 1; + sum += b[3]->ref_count - 1; + sum |= b[0]->buffer_pool_index ^ buffer_pool_index; + sum |= b[1]->buffer_pool_index ^ buffer_pool_index; + sum |= b[2]->buffer_pool_index ^ buffer_pool_index; + sum |= b[3]->buffer_pool_index ^ buffer_pool_index; +#endif + + if (sum) + goto one_by_one; + + vlib_buffer_copy_indices (queue + n_queue, buffers, 4); + vlib_buffer_copy_template (b[0], &bt); + vlib_buffer_copy_template (b[1], &bt); + vlib_buffer_copy_template (b[2], &bt); + vlib_buffer_copy_template (b[3], &bt); + n_queue += 4; + + if (CLIB_DEBUG > 0) + vlib_buffer_validate_alloc_free (vm, buffers, 4, + VLIB_BUFFER_KNOWN_ALLOCATED); + + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[0]); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[1]); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[2]); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[3]); + + if (n_queue >= queue_size) + { + vlib_buffer_pool_put (vm, buffer_pool_index, queue, n_queue); + n_queue = 0; + } + buffers += 4; + n_buffers -= 4; + continue; + + one_by_one: + bi = buffers[0]; + + next_in_chain: + b[0] = vlib_get_buffer (vm, bi); + flags = b[0]->flags; + next = b[0]->next_buffer; + + if (PREDICT_FALSE (buffer_pool_index != b[0]->buffer_pool_index)) + { + buffer_pool_index = b[0]->buffer_pool_index; +#if defined(CLIB_HAVE_VEC128) && !__aarch64__ + bpi_vec.buffer_pool_index = buffer_pool_index; +#endif + bp = vlib_get_buffer_pool (vm, buffer_pool_index); + vlib_buffer_copy_template (&bt, &bp->buffer_template); + + if (n_queue) + { + vlib_buffer_pool_put (vm, buffer_pool_index, queue, n_queue); + n_queue = 0; + } + } + + ASSERT (pointer_to_uword (b[0]) >= bp->start && + pointer_to_uword (b[0]) < + bp->start + bp->size - (bp->data_size + sizeof (*b[0]))); + + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[0]); + + if (clib_atomic_sub_fetch (&b[0]->ref_count, 1) == 0) + { + if (CLIB_DEBUG > 0) + vlib_buffer_validate_alloc_free (vm, &bi, 1, + VLIB_BUFFER_KNOWN_ALLOCATED); + vlib_buffer_copy_template (b[0], &bt); + queue[n_queue++] = bi; + } + + if (n_queue == queue_size) + { + vlib_buffer_pool_put (vm, buffer_pool_index, queue, queue_size); + n_queue = 0; + } + + if (flags & VLIB_BUFFER_NEXT_PRESENT) + { + bi = next; + goto next_in_chain; + } + + buffers++; + n_buffers--; + } + + if (n_queue) + vlib_buffer_pool_put (vm, buffer_pool_index, queue, n_queue); +} + + /** \brief Free buffers Frees the entire buffer chain for each buffer @@ -534,11 +799,7 @@ vlib_buffer_free (vlib_main_t * vm, /* number of buffers to free */ u32 n_buffers) { - vlib_buffer_main_t *bm = vm->buffer_main; - - ASSERT (bm->cb.vlib_buffer_free_cb); - - return bm->cb.vlib_buffer_free_cb (vm, buffers, n_buffers); + vlib_buffer_free_inline (vm, buffers, n_buffers, /* maybe next */ 1); } /** \brief Free buffers, does not free the buffer chain for each buffer @@ -555,11 +816,7 @@ vlib_buffer_free_no_next (vlib_main_t * vm, /* number of buffers to free */ u32 n_buffers) { - vlib_buffer_main_t *bm = vm->buffer_main; - - ASSERT (bm->cb.vlib_buffer_free_no_next_cb); - - return bm->cb.vlib_buffer_free_no_next_cb (vm, buffers, n_buffers); + vlib_buffer_free_inline (vm, buffers, n_buffers, /* maybe next */ 0); } /** \brief Free one buffer @@ -571,7 +828,7 @@ vlib_buffer_free_no_next (vlib_main_t * vm, always_inline void vlib_buffer_free_one (vlib_main_t * vm, u32 buffer_index) { - vlib_buffer_free (vm, &buffer_index, /* n_buffers */ 1); + vlib_buffer_free_inline (vm, &buffer_index, 1, /* maybe next */ 1); } /** \brief Free buffers from ring @@ -624,11 +881,6 @@ vlib_buffer_free_from_ring_no_next (vlib_main_t * vm, u32 * ring, u32 start, } } -/* Add/delete buffer free lists. */ -vlib_buffer_free_list_index_t vlib_buffer_create_free_list (vlib_main_t * vm, - u32 n_data_bytes, - char *fmt, ...); - /* Append given data to end of buffer, possibly allocating new buffers. */ int vlib_buffer_add_data (vlib_main_t * vm, u32 * buffer_index, void *data, u32 n_data_bytes); @@ -710,7 +962,7 @@ vlib_buffer_clone_256 (vlib_main_t * vm, u32 src_buffer, u32 * buffers, u16 i; vlib_buffer_t *s = vlib_get_buffer (vm, src_buffer); - ASSERT (s->n_add_refs == 0); + ASSERT (s->ref_count == 1); ASSERT (n_buffers); ASSERT (n_buffers <= 256); @@ -735,13 +987,16 @@ vlib_buffer_clone_256 (vlib_main_t * vm, u32 src_buffer, u32 * buffers, return 1; } - n_buffers = vlib_buffer_alloc (vm, buffers, n_buffers); + n_buffers = vlib_buffer_alloc_from_pool (vm, buffers, n_buffers, + s->buffer_pool_index); for (i = 0; i < n_buffers; i++) { vlib_buffer_t *d = vlib_get_buffer (vm, buffers[i]); d->current_data = s->current_data; d->current_length = head_end_offset; + ASSERT (d->buffer_pool_index == s->buffer_pool_index); + d->total_length_not_including_first_buffer = s->current_length - head_end_offset; if (PREDICT_FALSE (s->flags & VLIB_BUFFER_NEXT_PRESENT)) @@ -758,11 +1013,11 @@ vlib_buffer_clone_256 (vlib_main_t * vm, u32 src_buffer, u32 * buffers, d->next_buffer = src_buffer; } vlib_buffer_advance (s, head_end_offset); - s->n_add_refs = n_buffers - 1; + s->ref_count = n_buffers; while (s->flags & VLIB_BUFFER_NEXT_PRESENT) { s = vlib_get_buffer (vm, s->next_buffer); - s->n_add_refs = n_buffers - 1; + s->ref_count = n_buffers; } return n_buffers; @@ -816,6 +1071,7 @@ vlib_buffer_attach_clone (vlib_main_t * vm, vlib_buffer_t * head, vlib_buffer_t * tail) { ASSERT ((head->flags & VLIB_BUFFER_NEXT_PRESENT) == 0); + ASSERT (head->buffer_pool_index == tail->buffer_pool_index); head->flags |= VLIB_BUFFER_NEXT_PRESENT; head->flags &= ~VLIB_BUFFER_TOTAL_LENGTH_VALID; @@ -826,7 +1082,7 @@ vlib_buffer_attach_clone (vlib_main_t * vm, vlib_buffer_t * head, tail->total_length_not_including_first_buffer; next_segment: - clib_atomic_add_fetch (&tail->n_add_refs, 1); + clib_atomic_add_fetch (&tail->ref_count, 1); if (tail->flags & VLIB_BUFFER_NEXT_PRESENT) { @@ -875,7 +1131,6 @@ vlib_buffer_chain_increase_length (vlib_buffer_t * first, * Returns the number of copied bytes. */ always_inline u16 vlib_buffer_chain_append_data (vlib_main_t * vm, - vlib_buffer_free_list_index_t free_list_index, vlib_buffer_t * first, vlib_buffer_t * last, void *data, u16 data_len) { @@ -897,8 +1152,6 @@ vlib_buffer_chain_append_data (vlib_main_t * vm, * chained and points to the last buffer in the chain. */ u16 vlib_buffer_chain_append_data_with_alloc (vlib_main_t * vm, - vlib_buffer_free_list_index_t - free_list_index, vlib_buffer_t * first, vlib_buffer_t ** last, void *data, u16 data_len); @@ -915,17 +1168,9 @@ typedef struct /* Number of buffers to allocate in each call to allocator. */ u32 min_n_buffers_each_alloc; - /* Buffer free list for this template. */ - vlib_buffer_free_list_index_t free_list_index; - - u32 *free_buffers; - u8 *name; } vlib_packet_template_t; -void vlib_packet_template_get_packet_helper (vlib_main_t * vm, - vlib_packet_template_t * t); - void vlib_packet_template_init (vlib_main_t * vm, vlib_packet_template_t * t, void *packet_data, @@ -943,112 +1188,9 @@ vlib_packet_template_free (vlib_main_t * vm, vlib_packet_template_t * t) vec_free (t->packet_data); } -static_always_inline vlib_buffer_pool_t * -vlib_buffer_pool_get (vlib_main_t * vm, u8 buffer_pool_index) -{ - vlib_buffer_main_t *bm = vm->buffer_main; - return vec_elt_at_index (bm->buffer_pools, buffer_pool_index); -} - -always_inline void -vlib_buffer_add_to_free_list (vlib_main_t * vm, - vlib_buffer_free_list_t * f, - u32 buffer_index, u8 do_init) -{ - vlib_buffer_pool_t *bp = vlib_buffer_pool_get (vm, f->buffer_pool_index); - vlib_buffer_t *b; - b = vlib_get_buffer (vm, buffer_index); - if (PREDICT_TRUE (do_init)) - clib_memset (b, 0, STRUCT_OFFSET_OF (vlib_buffer_t, template_end)); - vec_add1_aligned (f->buffers, buffer_index, CLIB_CACHE_LINE_BYTES); - - if (vec_len (f->buffers) > 4 * VLIB_FRAME_SIZE) - { - clib_spinlock_lock (&bp->lock); - /* keep last stored buffers, as they are more likely hot in the cache */ - vec_add_aligned (bp->buffers, f->buffers, VLIB_FRAME_SIZE, - CLIB_CACHE_LINE_BYTES); - vec_delete (f->buffers, VLIB_FRAME_SIZE, 0); - f->n_alloc -= VLIB_FRAME_SIZE; - clib_spinlock_unlock (&bp->lock); - } -} - -#if CLIB_DEBUG > 0 -extern u32 *vlib_buffer_state_validation_lock; -extern uword *vlib_buffer_state_validation_hash; -extern void *vlib_buffer_state_heap; -#endif - -static inline void -vlib_validate_buffer_in_use (vlib_buffer_t * b, u32 expected) -{ -#if CLIB_DEBUG > 0 - uword *p; - void *oldheap; - - oldheap = clib_mem_set_heap (vlib_buffer_state_heap); - - while (clib_atomic_test_and_set (vlib_buffer_state_validation_lock)) - ; - - p = hash_get (vlib_buffer_state_validation_hash, b); - - /* If we don't know about b, declare it to be in the expected state */ - if (!p) - { - hash_set (vlib_buffer_state_validation_hash, b, expected); - goto out; - } - - if (p[0] != expected) - { - void cj_stop (void); - u32 bi; - vlib_main_t *vm = &vlib_global_main; - - cj_stop (); - - bi = vlib_get_buffer_index (vm, b); - - clib_mem_set_heap (oldheap); - clib_warning ("%.6f buffer %llx (%d): %s, not %s", - vlib_time_now (vm), bi, - p[0] ? "busy" : "free", expected ? "busy" : "free"); - os_panic (); - } -out: - CLIB_MEMORY_BARRIER (); - *vlib_buffer_state_validation_lock = 0; - clib_mem_set_heap (oldheap); -#endif -} - -static inline void -vlib_validate_buffer_set_in_use (vlib_buffer_t * b, u32 expected) -{ -#if CLIB_DEBUG > 0 - void *oldheap; - - oldheap = clib_mem_set_heap (vlib_buffer_state_heap); - - while (clib_atomic_test_and_set (vlib_buffer_state_validation_lock)) - ; - - hash_set (vlib_buffer_state_validation_hash, b, expected); - - CLIB_MEMORY_BARRIER (); - *vlib_buffer_state_validation_lock = 0; - clib_mem_set_heap (oldheap); -#endif -} - -/** minimum data size of first buffer in a buffer chain */ -#define VLIB_BUFFER_CHAIN_MIN_FIRST_DATA_SIZE (256) - /** * @brief compress buffer chain in a way where the first buffer is at least - * VLIB_BUFFER_CHAIN_MIN_FIRST_DATA_SIZE long + * VLIB_BUFFER_CLONE_HEAD_SIZE long * * @param[in] vm - vlib_main * @param[in,out] first - first buffer in chain @@ -1059,14 +1201,14 @@ always_inline void vlib_buffer_chain_compress (vlib_main_t * vm, vlib_buffer_t * first, u32 ** discard_vector) { - if (first->current_length >= VLIB_BUFFER_CHAIN_MIN_FIRST_DATA_SIZE || + if (first->current_length >= VLIB_BUFFER_CLONE_HEAD_SIZE || !(first->flags & VLIB_BUFFER_NEXT_PRESENT)) { /* this is already big enough or not a chain */ return; } - u32 want_first_size = clib_min (VLIB_BUFFER_CHAIN_MIN_FIRST_DATA_SIZE, + u32 want_first_size = clib_min (VLIB_BUFFER_CLONE_HEAD_SIZE, VLIB_BUFFER_DATA_SIZE - first->current_data); do diff --git a/src/vlib/log.h b/src/vlib/log.h index d21bca9f195..9b285429b9a 100644 --- a/src/vlib/log.h +++ b/src/vlib/log.h @@ -18,8 +18,6 @@ #include <vppinfra/types.h> -typedef u32 vlib_log_class_t; - #define foreach_vlib_log_level \ _(0, EMERG, emerg) \ _(1, ALERT, alert) \ diff --git a/src/vlib/main.c b/src/vlib/main.c index a64a991d438..c0ab8e129b9 100644 --- a/src/vlib/main.c +++ b/src/vlib/main.c @@ -463,7 +463,7 @@ vlib_put_next_frame (vlib_main_t * vm, vlib_frame_t *f; u32 n_vectors_in_frame; - if (vm->buffer_main->callbacks_registered == 0 && CLIB_DEBUG > 0) + if (CLIB_DEBUG > 0) vlib_put_next_frame_validate (vm, r, next_index, n_vectors_left); nf = vlib_node_runtime_get_next_frame (vm, r, next_index); @@ -987,8 +987,8 @@ format_buffer_metadata (u8 * s, va_list * args) (i32) (b->current_data), (i32) (b->current_length)); s = format (s, "current_config_index: %d, flow_id: %x, next_buffer: %x\n", b->current_config_index, b->flow_id, b->next_buffer); - s = format (s, "error: %d, n_add_refs: %d, buffer_pool_index: %d\n", - (u32) (b->error), (u32) (b->n_add_refs), + s = format (s, "error: %d, ref_count: %d, buffer_pool_index: %d\n", + (u32) (b->error), (u32) (b->ref_count), (u32) (b->buffer_pool_index)); s = format (s, "trace_index: %d, len_not_first_buf: %d\n", @@ -1993,9 +1993,6 @@ vlib_main (vlib_main_t * volatile vm, unformat_input_t * input) if ((error = vlib_call_all_init_functions (vm))) goto done; - /* Create default buffer free list. */ - vlib_buffer_create_free_list (vm, VLIB_BUFFER_DATA_SIZE, "default"); - nm->timing_wheel = clib_mem_alloc_aligned (sizeof (TWT (tw_timer_wheel)), CLIB_CACHE_LINE_BYTES); diff --git a/src/vlib/main.h b/src/vlib/main.h index c6786d8f788..920e5f3f95a 100644 --- a/src/vlib/main.h +++ b/src/vlib/main.h @@ -115,9 +115,6 @@ typedef struct vlib_main_t /* Size of the heap */ uword heap_size; - /* Pool of buffer free lists. */ - vlib_buffer_free_list_t *buffer_free_list_pool; - /* buffer main structure. */ vlib_buffer_main_t *buffer_main; diff --git a/src/vlib/threads.c b/src/vlib/threads.c index ed1ba61a0cc..cdb4eb082b5 100644 --- a/src/vlib/threads.c +++ b/src/vlib/threads.c @@ -719,8 +719,6 @@ start_workers (vlib_main_t * vm) for (i = 0; i < vec_len (tm->registrations); i++) { vlib_node_main_t *nm, *nm_clone; - vlib_buffer_free_list_t *fl_clone, *fl_orig; - vlib_buffer_free_list_t *orig_freelist_pool; int k; tr = tm->registrations[i]; @@ -883,26 +881,6 @@ start_workers (vlib_main_t * vm) (vlib_mains[0]->error_main.counters_last_clear, CLIB_CACHE_LINE_BYTES); - /* Fork the vlib_buffer_main_t free lists, etc. */ - orig_freelist_pool = vm_clone->buffer_free_list_pool; - vm_clone->buffer_free_list_pool = 0; - - /* *INDENT-OFF* */ - pool_foreach (fl_orig, orig_freelist_pool, - ({ - pool_get_aligned (vm_clone->buffer_free_list_pool, - fl_clone, CLIB_CACHE_LINE_BYTES); - ASSERT (fl_orig - orig_freelist_pool - == fl_clone - vm_clone->buffer_free_list_pool); - - fl_clone[0] = fl_orig[0]; - fl_clone->buffers = 0; - vec_validate(fl_clone->buffers, 0); - vec_reset_length(fl_clone->buffers); - fl_clone->n_alloc = 0; - })); -/* *INDENT-ON* */ - worker_thread_index++; } } diff --git a/src/vlib/vlib.h b/src/vlib/vlib.h index 49bb51bfbe3..2149c4763ab 100644 --- a/src/vlib/vlib.h +++ b/src/vlib/vlib.h @@ -48,6 +48,7 @@ /* Forward declarations of structs to avoid circular dependencies. */ struct vlib_main_t; +typedef u32 vlib_log_class_t; /* All includes in alphabetical order. */ #include <vlib/physmem.h> |