From: Matias Elo <matias....@nokia.com> During sock_mmap_send() when all packets are not transmitted by sendto() call the used ring slots have to be marked as TP_STATUS_AVAILABLE or they will be transmitted during the next sock_mmap_send() call, thus breaking the function logic.
To reduce the number of failed TX operations (visible to applications) the pktio implementation can call sendto() up to TX_RETRIES times during a single sock_mmap_recv() call. Signed-off-by: Matias Elo <matias....@nokia.com> --- /** Email created from pull request 397 (matiaselo:dev/pktio_fixes) ** https://github.com/Linaro/odp/pull/397 ** Patch: https://github.com/Linaro/odp/pull/397.patch ** Base sha: 520c170d758f2d37554631bf1467ec50e027cd3e ** Merge commit sha: 926e7716387610a8be67fbc051845450640f8df6 **/ platform/linux-generic/pktio/socket_mmap.c | 98 +++++++++++++++++++----------- 1 file changed, 63 insertions(+), 35 deletions(-) diff --git a/platform/linux-generic/pktio/socket_mmap.c b/platform/linux-generic/pktio/socket_mmap.c index 13707e8f2..231c29698 100644 --- a/platform/linux-generic/pktio/socket_mmap.c +++ b/platform/linux-generic/pktio/socket_mmap.c @@ -23,6 +23,7 @@ #include <poll.h> #include <sys/ioctl.h> #include <errno.h> +#include <time.h> #include <odp_api.h> #include <odp_packet_socket.h> @@ -37,6 +38,12 @@ #include <protocols/eth.h> #include <protocols/ip.h> +/* Maximum number of retries per sock_mmap_send() call */ +#define TX_RETRIES 10 + +/* Number of nanoseconds to wait between TX retries */ +#define TX_RETRY_NSEC 1000 + /* Maximum number of packets to store in each RX/TX block */ #define MAX_PKTS_PER_BLOCK 512 @@ -148,6 +155,11 @@ static uint8_t *pkt_mmap_vlan_insert(uint8_t *l2_hdr_ptr, return l2_hdr_ptr; } +static inline unsigned next_frame(unsigned cur_frame, unsigned frame_count) +{ + return odp_unlikely(cur_frame + 1 >= frame_count) ? 0 : cur_frame + 1; +} + static inline unsigned pkt_mmap_v2_rx(pktio_entry_t *pktio_entry, pkt_sock_mmap_t *pkt_sock, odp_packet_t pkt_table[], unsigned len, @@ -185,7 +197,7 @@ static inline unsigned pkt_mmap_v2_rx(pktio_entry_t *pktio_entry, ts_val = odp_time_global(); ppd.raw = ring->rd[frame_num].iov_base; - next_frame_num = (frame_num + 1) % ring->rd_num; + next_frame_num = next_frame(frame_num, ring->rd_num); pkt_buf = (uint8_t *)ppd.raw + ppd.v2->tp_h.tp_mac; pkt_len = ppd.v2->tp_h.tp_snaplen; @@ -251,6 +263,50 @@ static inline unsigned pkt_mmap_v2_rx(pktio_entry_t *pktio_entry, return nb_rx; } +static unsigned handle_pending_frames(int sock, struct ring *ring, int frames) +{ + int i; + int retry = 0; + unsigned nb_tx = 0; + unsigned frame_num; + unsigned frame_count = ring->rd_num; + unsigned first_frame_num = ring->frame_num; + + for (frame_num = first_frame_num, i = 0; i < frames; i++) { + struct tpacket2_hdr *hdr = ring->rd[frame_num].iov_base; + + if (odp_likely(hdr->tp_status == TP_STATUS_AVAILABLE || + hdr->tp_status == TP_STATUS_SENDING)) { + nb_tx++; + } else if (hdr->tp_status == TP_STATUS_SEND_REQUEST) { + if (retry++ < TX_RETRIES) { + struct timespec ts = { .tv_nsec = TX_RETRY_NSEC, + .tv_sec = 0 }; + + sendto(sock, NULL, 0, MSG_DONTWAIT, NULL, 0); + nanosleep(&ts, NULL); + i--; + continue; + } else { + hdr->tp_status = TP_STATUS_AVAILABLE; + } + } else { /* TP_STATUS_WRONG_FORMAT */ + /* Don't try re-sending frames after failure */ + for (; i < frames; i++) { + hdr = ring->rd[frame_num].iov_base; + hdr->tp_status = TP_STATUS_AVAILABLE; + frame_num = next_frame(frame_num, frame_count); + } + break; + } + frame_num = next_frame(frame_num, frame_count); + } + + ring->frame_num = (first_frame_num + nb_tx) % frame_count; + + return nb_tx; +} + static inline unsigned pkt_mmap_v2_tx(int sock, struct ring *ring, const odp_packet_t pkt_table[], unsigned len) @@ -260,7 +316,7 @@ static inline unsigned pkt_mmap_v2_tx(int sock, struct ring *ring, unsigned first_frame_num, frame_num, frame_count; int ret; uint8_t *buf; - unsigned n, i = 0; + unsigned i = 0; unsigned nb_tx = 0; int send_errno; int total_len = 0; @@ -285,9 +341,7 @@ static inline unsigned pkt_mmap_v2_tx(int sock, struct ring *ring, mmap_tx_user_ready(ppd.raw); - if (++frame_num >= frame_count) - frame_num = 0; - + frame_num = next_frame(frame_num, frame_count); i++; } @@ -301,27 +355,11 @@ static inline unsigned pkt_mmap_v2_tx(int sock, struct ring *ring, if (odp_likely(ret == total_len)) { nb_tx = i; ring->frame_num = frame_num; - } else if (ret == -1) { - for (frame_num = first_frame_num, n = 0; n < i; ++n) { - struct tpacket2_hdr *hdr = ring->rd[frame_num].iov_base; - - if (odp_likely(hdr->tp_status == TP_STATUS_AVAILABLE || - hdr->tp_status == TP_STATUS_SENDING)) { - nb_tx++; - } else { - /* The remaining frames weren't sent, clear - * their status to indicate we're not waiting - * for the kernel to process them. */ - hdr->tp_status = TP_STATUS_AVAILABLE; - } - - if (++frame_num >= frame_count) - frame_num = 0; - } - - ring->frame_num = (first_frame_num + nb_tx) % frame_count; + } else { + nb_tx = handle_pending_frames(sock, ring, i); - if (nb_tx == 0 && SOCK_ERR_REPORT(send_errno)) { + if (odp_unlikely(ret == -1 && nb_tx == 0 && + SOCK_ERR_REPORT(send_errno))) { __odp_errno = send_errno; /* ENOBUFS indicates that the transmit queue is full, * which will happen regularly when overloaded so don't @@ -331,16 +369,6 @@ static inline unsigned pkt_mmap_v2_tx(int sock, struct ring *ring, strerror(send_errno)); return -1; } - } else { - /* Short send, return value is number of bytes sent so use this - * to determine number of complete frames sent. */ - for (n = 0; n < i && ret > 0; ++n) { - ret -= odp_packet_len(pkt_table[n]); - if (ret >= 0) - nb_tx++; - } - - ring->frame_num = (first_frame_num + nb_tx) % frame_count; } for (i = 0; i < nb_tx; ++i)
