From 3a8f32be9f67546d51de110efa2ada6bcceb454f Mon Sep 17 00:00:00 2001 From: Pierre Pfister Date: Mon, 21 Mar 2016 12:21:30 +0000 Subject: tapcli: Receive vector of packets and memory leak fix tapcli interfaces were creating single-packet frames. It now calls readv until the frame is full, or readv returns error EAGAIN. This is usefull when a significant amount of traffic flows through tap interfaces. This patch also fixes a memory leak by correctly initializing b->clone_count to zero. Change-Id: I15e435ba76d542be2f263274e76297425cd10243 Signed-off-by: Pierre Pfister --- vnet/vnet/unix/tapcli.c | 319 ++++++++++++++++++++++++++---------------------- vnet/vnet/unix/tapcli.h | 12 ++ 2 files changed, 183 insertions(+), 148 deletions(-) (limited to 'vnet') diff --git a/vnet/vnet/unix/tapcli.c b/vnet/vnet/unix/tapcli.c index 4421ffb918d..5b0ac9378ee 100644 --- a/vnet/vnet/unix/tapcli.c +++ b/vnet/vnet/unix/tapcli.c @@ -43,6 +43,7 @@ static vnet_device_class_t tapcli_dev_class; static vnet_hw_interface_class_t tapcli_interface_class; +static vlib_node_registration_t tapcli_rx_node; static void tapcli_nopunt_frame (vlib_main_t * vm, vlib_node_runtime_t * node, @@ -59,6 +60,21 @@ typedef struct { u8 active; /* for delete */ } tapcli_interface_t; +typedef struct { + u16 sw_if_index; +} tapcli_rx_trace_t; + +u8 * format_tapcli_rx_trace (u8 * s, va_list * va) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *); + vnet_main_t * vnm = vnet_get_main(); + tapcli_rx_trace_t * t = va_arg (*va, tapcli_rx_trace_t *); + s = format (s, "%U", format_vnet_sw_if_index_name, + vnm, t->sw_if_index); + return s; +} + typedef struct { /* Vector of iovecs for readv/writev calls. */ struct iovec * iovecs; @@ -197,24 +213,157 @@ enum { TAPCLI_RX_N_NEXT, }; + + +static uword tapcli_rx_iface(vlib_main_t * vm, + vlib_node_runtime_t * node, + tapcli_interface_t * ti) +{ + tapcli_main_t * tm = &tapcli_main; + const uword buffer_size = VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES; + u32 n_trace = vlib_get_trace_count (vm, node); + u8 set_trace = 0; + + vnet_main_t *vnm; + vnet_sw_interface_t * si; + u8 admin_down; + u32 next = node->cached_next_index; + u32 n_left_to_next, next_index; + u32 *to_next; + + vnm = vnet_get_main(); + si = vnet_get_sw_interface (vnm, ti->sw_if_index); + admin_down = !(si->flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP); + + vlib_get_next_frame(vm, node, next, to_next, n_left_to_next); + + while (n_left_to_next) { // Fill at most one vector + vlib_buffer_t *b_first, *b, *prev; + u32 bi_first, bi; + word n_bytes_in_packet; + int j, n_bytes_left; + + if (PREDICT_FALSE(vec_len(tm->rx_buffers) < tm->mtu_buffers)) { + uword len = vec_len(tm->rx_buffers); + _vec_len(tm->rx_buffers) += + vlib_buffer_alloc_from_free_list(vm, &tm->rx_buffers[len], + VLIB_FRAME_SIZE - len, VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + if (PREDICT_FALSE(vec_len(tm->rx_buffers) < tm->mtu_buffers)) { + clib_warning("vlib_buffer_alloc failed"); + break; + } + } + + uword i_rx = vec_len (tm->rx_buffers) - 1; + + /* Allocate RX buffers from end of rx_buffers. + Turn them into iovecs to pass to readv. */ + vec_validate (tm->iovecs, tm->mtu_buffers - 1); + for (j = 0; j < tm->mtu_buffers; j++) { + b = vlib_get_buffer (vm, tm->rx_buffers[i_rx - j]); + b->clone_count = 0; + tm->iovecs[j].iov_base = b->data; + tm->iovecs[j].iov_len = buffer_size; + } + + n_bytes_left = readv (ti->unix_fd, tm->iovecs, tm->mtu_buffers); + n_bytes_in_packet = n_bytes_left; + if (n_bytes_left <= 0) { + if (errno != EAGAIN) { + vlib_node_increment_counter(vm, tapcli_rx_node.index, + TAPCLI_ERROR_READ, 1); + } + break; + } + + bi_first = tm->rx_buffers[i_rx]; + b = b_first = vlib_get_buffer (vm, tm->rx_buffers[i_rx]); + prev = NULL; + + while (1) { + b->current_length = n_bytes_left < buffer_size ? n_bytes_left : buffer_size; + n_bytes_left -= buffer_size; + + if (prev) { + prev->next_buffer = bi; + prev->flags |= VLIB_BUFFER_NEXT_PRESENT; + } + prev = b; + + /* last segment */ + if (n_bytes_left <= 0) + break; + + i_rx--; + bi = tm->rx_buffers[i_rx]; + b = vlib_get_buffer (vm, bi); + } + + _vec_len (tm->rx_buffers) = i_rx; + + b_first->total_length_not_including_first_buffer = + (n_bytes_in_packet > buffer_size) ? n_bytes_in_packet - buffer_size : 0; + b_first->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID; + + /* Ensure mbufs are updated */ + vlib_buffer_chain_validate(vm, b_first); + + VLIB_BUFFER_TRACE_TRAJECTORY_INIT(b_first); + + vnet_buffer (b_first)->sw_if_index[VLIB_RX] = ti->sw_if_index; + vnet_buffer (b_first)->sw_if_index[VLIB_TX] = (u32)~0; + + b_first->error = node->errors[TAPCLI_ERROR_NONE]; + next_index = TAPCLI_RX_NEXT_ETHERNET_INPUT; + next_index = (ti->per_interface_next_index != ~0) ? + ti->per_interface_next_index : next_index; + next_index = admin_down ? TAPCLI_RX_NEXT_DROP : next_index; + + to_next[0] = bi_first; + to_next++; + n_left_to_next--; + + vlib_validate_buffer_enqueue_x1 (vm, node, next, + to_next, n_left_to_next, + bi_first, next_index); + + /* Interface counters for tapcli interface. */ + if (PREDICT_TRUE(!admin_down)) { + vlib_increment_combined_counter ( + vnet_main.interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, + os_get_cpu_number(), ti->sw_if_index, + 1, n_bytes_in_packet); + + if (PREDICT_FALSE(n_trace > 0)) { + vlib_trace_buffer (vm, node, next_index, + b_first, /* follow_chain */ 1); + n_trace--; + set_trace = 1; + tapcli_rx_trace_t *t0 = vlib_add_trace (vm, node, b_first, sizeof (*t0)); + t0->sw_if_index = si->sw_if_index; + } + } + } + vlib_put_next_frame (vm, node, next, n_left_to_next); + if (set_trace) + vlib_set_trace_count (vm, node, n_trace); + return VLIB_FRAME_SIZE - n_left_to_next; +} + static uword tapcli_rx (vlib_main_t * vm, vlib_node_runtime_t * node, vlib_frame_t * frame) { tapcli_main_t * tm = &tapcli_main; - vlib_buffer_t *b_first; - u32 bi; - vlib_buffer_free_list_t *fl; - const uword buffer_size = VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES; static u32 * ready_interface_indices; tapcli_interface_t * ti; int i; - word n_bytes_in_packet; + u32 total_count = 0; vec_reset_length (ready_interface_indices); - - clib_bitmap_foreach (i, tm->pending_read_bitmap, + clib_bitmap_foreach (i, tm->pending_read_bitmap, ({ vec_add1 (ready_interface_indices, i); })); @@ -222,161 +371,33 @@ tapcli_rx (vlib_main_t * vm, if (vec_len (ready_interface_indices) == 0) return 1; - fl = vlib_buffer_get_free_list(vm, VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); - for (i = 0; i < vec_len(ready_interface_indices); i++) - { - /* Clear the "interrupt" bit */ - tm->pending_read_bitmap = + { + tm->pending_read_bitmap = clib_bitmap_set (tm->pending_read_bitmap, ready_interface_indices[i], 0); - - ti = vec_elt_at_index (tm->tapcli_interfaces, ready_interface_indices[i]); - - /* Make sure we have some RX buffers. */ - { - uword n_left = vec_len (tm->rx_buffers); - uword n_alloc; - if (n_left < VLIB_FRAME_SIZE / 2) { - vec_validate(tm->rx_buffers, VLIB_FRAME_SIZE + n_left - 1); - n_alloc = vlib_buffer_alloc(vm, &tm->rx_buffers[n_left], VLIB_FRAME_SIZE); - n_left += n_alloc; - _vec_len (tm->rx_buffers) = n_left; - } - } - - /* Allocate RX buffers from end of rx_buffers. - Turn them into iovecs to pass to readv. */ - { - uword i_rx = vec_len (tm->rx_buffers) - 1; - vlib_buffer_t * b, *prev = 0; - word j, n_bytes_left; - - /* We need enough buffers left for an MTU sized packet. */ - if (PREDICT_FALSE(vec_len (tm->rx_buffers) < tm->mtu_buffers)) - { - clib_bitmap_set (tm->pending_read_bitmap, - ready_interface_indices[i], 1); - clib_warning ("buffer allocation failure"); - continue; - } - - vec_validate (tm->iovecs, tm->mtu_buffers - 1); - for (j = 0; j < tm->mtu_buffers; j++) - { - b = vlib_get_buffer (vm, tm->rx_buffers[i_rx - j]); - vlib_buffer_init_for_free_list (b, fl); - tm->iovecs[j].iov_base = b->data; - tm->iovecs[j].iov_len = buffer_size; - } - - n_bytes_left = readv (ti->unix_fd, tm->iovecs, tm->mtu_buffers); - n_bytes_in_packet = n_bytes_left; - if (n_bytes_left <= 0) - { - if (errno != EAGAIN) - clib_unix_warning ("readv %d", n_bytes_left); - return 0; - } - - bi = tm->rx_buffers[i_rx]; - b = b_first = vlib_get_buffer (vm, tm->rx_buffers[i_rx]); - - while (1) { - u32 bi; - - vlib_buffer_init_for_free_list(b, fl); - - b->current_length = n_bytes_left < buffer_size ? n_bytes_left : buffer_size; - n_bytes_left -= buffer_size; - - if (prev) { - prev->next_buffer = bi; - prev->flags |= VLIB_BUFFER_NEXT_PRESENT; - } - prev = b; - - /* last segment */ - if (n_bytes_left <= 0) break; - - i_rx--; - bi = tm->rx_buffers[i_rx]; - b = vlib_get_buffer (vm, bi); - } - - /* Interface counters for tapcli interface. */ - vlib_increment_combined_counter - (vnet_main.interface_main.combined_sw_if_counters - + VNET_INTERFACE_COUNTER_RX, - os_get_cpu_number(), - ti->sw_if_index, - 1, n_bytes_in_packet); - - _vec_len (tm->rx_buffers) = i_rx; - } - - b_first->total_length_not_including_first_buffer = (n_bytes_in_packet > buffer_size) ? n_bytes_in_packet - buffer_size : 0; - b_first->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID; - /* Ensure mbufs are updated */ - vlib_buffer_chain_validate(vm, b_first); - - /* - * Turn this on if you run into - * "bad monkey" contexts, and you want to know exactly - * which nodes they've visited... See .../vlib/vlib/buffer.h - */ - VLIB_BUFFER_TRACE_TRAJECTORY_INIT(b_first); - - { - u32 next_index; - uword n_trace = vlib_get_trace_count (vm, node); - - vnet_buffer (b_first)->sw_if_index[VLIB_RX] = ti->sw_if_index; - vnet_buffer (b_first)->sw_if_index[VLIB_TX] = (u32)~0; - - b_first->error = node->errors[0]; - - { - next_index = TAPCLI_RX_NEXT_ETHERNET_INPUT; - - next_index = (ti->per_interface_next_index != ~0) ? - ti->per_interface_next_index : next_index; - } - { - vnet_main_t *vnm = vnet_get_main(); - vnet_sw_interface_t * si; - si = vnet_get_sw_interface (vnm, ti->sw_if_index); - if (!(si->flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP)) - next_index = TAPCLI_RX_NEXT_DROP; - } - - vlib_set_next_frame_buffer (vm, node, next_index, bi); - - if (n_trace > 0) - { - vlib_trace_buffer (vm, node, next_index, - b_first, /* follow_chain */ 1); - vlib_set_trace_count (vm, node, n_trace - 1); - } - } - } - - return 1; + ti = vec_elt_at_index (tm->tapcli_interfaces, ready_interface_indices[i]); + total_count += tapcli_rx_iface(vm, node, ti); + } + return total_count; //This might return more than 256. } static char * tapcli_rx_error_strings[] = { - "Interface down", +#define _(sym,string) string, + foreach_tapcli_error +#undef _ }; -VLIB_REGISTER_NODE (tapcli_rx_node,static) = { +VLIB_REGISTER_NODE (tapcli_rx_node, static) = { .function = tapcli_rx, .name = "tapcli-rx", .type = VLIB_NODE_TYPE_INPUT, .state = VLIB_NODE_STATE_INTERRUPT, .vector_size = 4, - .n_errors = 1, + .n_errors = TAPCLI_N_ERROR, .error_strings = tapcli_rx_error_strings, + .format_trace = format_tapcli_rx_trace, .n_next_nodes = TAPCLI_RX_N_NEXT, .next_nodes = { @@ -1169,8 +1190,10 @@ tapcli_init (vlib_main_t * vm) tm->mtu_bytes = 4096 + 256; tm->tapcli_interface_index_by_sw_if_index = hash_create (0, sizeof(uword)); tm->tapcli_interface_index_by_unix_fd = hash_create (0, sizeof (uword)); + tm->rx_buffers = 0; + vec_alloc(tm->rx_buffers, VLIB_FRAME_SIZE); + vec_reset_length(tm->rx_buffers); vm->os_punt_frame = tapcli_nopunt_frame; - return 0; } diff --git a/vnet/vnet/unix/tapcli.h b/vnet/vnet/unix/tapcli.h index 1f5f4c3ee73..00a96c6b7bf 100644 --- a/vnet/vnet/unix/tapcli.h +++ b/vnet/vnet/unix/tapcli.h @@ -18,6 +18,18 @@ #ifndef __included_tapcli_h__ #define __included_tapcli_h__ +#define foreach_tapcli_error \ + /* Must be first. */ \ + _(NONE, "no error") \ + _(READ, "read error") \ + _(UNKNOWN, "unknown error") + +typedef enum { +#define _(sym,str) TAPCLI_ERROR_##sym, + foreach_tapcli_error +#undef _ + TAPCLI_N_ERROR, + } tapcli_error_t; typedef struct { u32 sw_if_index; -- cgit 1.2.3-korg