Changeset b148e3b for lib

Timestamp:

11/25/16 10:13:11 (4 years ago)

Author:

Richard Sanger <rsanger@…>

Branches:

4.0.1-hotfixes, cachetimestamps, develop, dpdk-ndag, etsilive, master, ndag_format, rc-4.0.1, rc-4.0.2, rc-4.0.3, rc-4.0.4, ringdecrementfix, ringperformance, ringtimestampfixes

Children:

caf7841

Parents:

571e2f9

git-author:

Richard Sanger <rsanger@…> (11/25/16 00:22:43)

git-committer:

Richard Sanger <rsanger@…> (11/25/16 10:13:11)

Message:

Updates DPDK to latest release and improves performance

Thanks to Richard Cziva for supplying an intial patch for this.

We now recommend using the latest release of DPDK, ideally 16.04 or newer

To support newer releases

• Fixes RSS hashing renames
• Fixes deprecated rte_mempool_count
• Fixes ETH_LINK_SPEED_X rename
• Fixes TX minimum memory requirement
• Fixes dropped vs errored counting in recent versions (for best results use 16.04 or newer)

Tuned to allow DPDK's SSE vector in supporting drivers mode for better performance.
Bumps default internal batch size up to 32 to matches DPDK in SSE vector mode.

Location:

lib

Files:

• format_dpdk.c (modified) (14 diffs)
• libtrace.h.in (modified) (16 diffs)
• trace_parallel.c (modified) (1 diff)

lib/format_dpdk.c

@@ -23 +23 @@
  *
  *
- */
+ *
  * Kit capture format.
  *
@@ -69 +69 @@
  * code (that we still attempt to support).
  *
- * DPDK v1.7.1 is recommended.
- * However 1.5 to 1.8 are likely supported.
+ * DPDK 16.04 or newer is recommended.
+ * However 1.6 and newer are still likely supported.
  */
 #include <rte_eal.h>
@@ -145 +145 @@
 #else
 #       define DPDK_USE_NULL_QUEUE_CONFIG 0
+#endif
+
+/* 2.0.0-rc1
+ * Unifies RSS hash between cards
+ */
+#if RTE_VERSION >= RTE_VERSION_NUM(2, 0, 0, 1)
+#       define RX_RSS_FLAGS (ETH_RSS_IP | ETH_RSS_UDP | ETH_RSS_TCP | \
+                             ETH_RSS_SCTP)
+#else
+#       define RX_RSS_FLAGS (ETH_RSS_IPV4_UDP | ETH_RSS_IPV6 | ETH_RSS_IPV4 | \
+                             ETH_RSS_IPV4_TCP | ETH_RSS_IPV6_TCP |\
+                             ETH_RSS_IPV6_UDP)
+#endif
+
+/* v16.07-rc1 - deprecated
+ * rte_mempool_avail_count to replace rte_mempool_count
+ * rte_mempool_in_use_count to replace rte_mempool_free_count
+ */
+#if RTE_VERSION < RTE_VERSION_NUM(16, 7, 0, 1)
+#define rte_mempool_avail_count rte_mempool_count
+#define rte_mempool_in_use_count rte_mempool_free_count
 #endif
 
@@ -170 +191 @@
 #endif
 
+/* 16.04-rc3 ETH_LINK_SPEED_X are replaced with ETH_SPEED_NUM_X.
+ * ETH_LINK_SPEED_ are reused as flags, ugly.
+ * We use the new way in this code.
+ */
+#ifndef ETH_SPEED_NUM_1G
+        #define ETH_SPEED_NUM_1G ETH_LINK_SPEED_1000
+        #define ETH_SPEED_NUM_10G ETH_LINK_SPEED_10G
+        #define ETH_SPEED_NUM_20G ETH_LINK_SPEED_20G
+        #define ETH_SPEED_NUM_40G ETH_LINK_SPEED_40G
+#endif
 
 /* The default size of memory buffers to use - This is the max size of standard
@@ -175 +206 @@
 #define RX_MBUF_SIZE 1514
 
-/* The minimum number of memory buffers per queue tx or rx. Search for
- * _MIN_RING_DESC in DPDK. The largest minimum is 64 for 10GBit cards.
- */
-#define MIN_NB_BUF 64
+/* The minimum number of memory buffers per queue tx or rx. Based on
+ * the requirement of the memory pool with 128 per thread buffers, needing
+ * at least 128*1.5 = 192 buffers. Our code allocates 128*2 to be safe.
+ */
+#define MIN_NB_BUF 128
 
 /* Number of receive memory buffers to use
@@ -185 +217 @@
  * This can be increased in the driver and here.
  * Should be at least MIN_NB_BUF.
- */
-#define NB_RX_MBUF 4096
+ * We choose 2K rather than 4K because it enables the usage of sse vector
+ * drivers which are significantly faster than using the larger buffer.
+ */
+#define NB_RX_MBUF (4096/2)
 
 /* Number of send memory buffers to use.
@@ -229 +263 @@
  *
  * Make sure you understand what these are doing before enabling them.
- * They might make traces incompatable with other builds etc.
+ * They might make traces incompatible with other builds etc.
  *
  * These are also included to show how to do somethings which aren't
@@ -236 +270 @@
 
 /* Print verbose messages to stderr */
-#define DEBUG 0
+#define DEBUG 1
 
 /* Use clock_gettime() for nanosecond resolution rather than gettimeofday()
  * only turn on if you know clock_gettime is a vsyscall on your system
- * overwise could be a large overhead. Again gettimeofday() should be
+ * otherwise could be a large overhead. Again gettimeofday() should be
  * vsyscall also if it's not you should seriously consider updating your
  * kernel.
@@ -297 +331 @@
         int lcore;
 #if HAS_HW_TIMESTAMPS_82580
-        /* Timestamping only relevent to RX */
+        /* Timestamping only relevant to RX */
         uint64_t ts_first_sys; /* Sytem timestamp of the first packet in nanoseconds */
         uint32_t wrap_count; /* Number of times the NIC clock has wrapped around completely */
@@ -959 +993 @@
                 .rss_conf = {
                         // .rss_key = &rss_key, // We set this per format
-                        .rss_hf = ETH_RSS_IPV4_UDP | ETH_RSS_IPV6 | ETH_RSS_IPV4 | ETH_RSS_IPV4_TCP | ETH_RSS_IPV6_TCP | ETH_RSS_IPV6_UDP,
+                        .rss_hf = RX_RSS_FLAGS,
                 },
         },
@@ -1087 +1121 @@
         int i;
         struct rte_config *cfg = rte_eal_get_configuration();
+        (void) socket;
 
         pthread_mutex_lock(&dpdk_lock);
@@ -1246 +1281 @@
                 fprintf(stderr, "DPDK memory pool not empty %d of %d, please "
                         "free all packets before finishing a trace\n",
-                        rte_mempool_count(mempool), mempool->size);
+                        rte_mempool_avail_count(mempool), mempool->size);
         }
 
@@ -1842 +1877 @@
 retry_calc_wiretime:
         switch (format_data->link_speed) {
-        case ETH_LINK_SPEED_40G:
-                wire_time /=  ETH_LINK_SPEED_40G;
+        case ETH_SPEED_NUM_40G:
+                wire_time /=  ETH_SPEED_NUM_40G;
                 break;
-        case ETH_LINK_SPEED_20G:
-                wire_time /= ETH_LINK_SPEED_20G;
+        case ETH_SPEED_NUM_20G:
+                wire_time /= ETH_SPEED_NUM_20G;
                 break;
-        case ETH_LINK_SPEED_10G:
-                wire_time /= ETH_LINK_SPEED_10G;
+        case ETH_SPEED_NUM_10G:
+                wire_time /= ETH_SPEED_NUM_10G;
                 break;
-        case ETH_LINK_SPEED_1000:
-                wire_time /= ETH_LINK_SPEED_1000;
+        case ETH_SPEED_NUM_1G:
+                wire_time /= ETH_SPEED_NUM_1G;
                 break;
         case 0:
@@ -2194 +2229 @@
         stats->captured = dev_stats.ipackets;
 
-        /* Not that we support adding filters but if we did this
-         * would work */
-        stats->filtered += dev_stats.fdirmiss;
-
         stats->dropped_valid = true;
         stats->dropped = dev_stats.imissed;
 
+#if RTE_VERSION >= RTE_VERSION_NUM(16, 4, 0, 2)
+        /* DPDK commit 86057c fixes ensures missed does not get counted as
+         * errors */
+        stats->errors_valid = true;
+        stats->errors = dev_stats.ierrors;
+#else
         /* DPDK errors includes drops */
         stats->errors_valid = true;
         stats->errors = dev_stats.ierrors - dev_stats.imissed;
-
+#endif
         stats->received_valid = true;
         stats->received = dev_stats.ipackets + dev_stats.imissed;

lib/libtrace.h.in

@@ -260 +260 @@
 /** If the packet has allocated its own memory the buffer_control should be
  * set to TRACE_CTRL_PACKET, so that the memory will be freed when the packet
- * is destroyed. If the packet has been zerocopied out of memory owned by
+ * is destroyed. If the packet has been zero-copied out of memory owned by
  * something else, e.g. a DAG card, it should be TRACE_CTRL_EXTERNAL.
  *
@@ -282 +282 @@
 /** Enumeration of error codes */
 enum {
-        /** No Error has occured.... yet. */
+        /** No Error has occurred.... yet. */
         TRACE_ERR_NOERROR       = 0,
         /** The URI passed to trace_create() is unsupported, or badly formed */
@@ -336 +336 @@
 #endif
         TRACE_DLT_PPP_SERIAL = 50,
-        TRACE_DLT_LINKTYPE_RAW = 101, /**< See TRACE_DLT_RAW for explainations of pain. */
+        TRACE_DLT_LINKTYPE_RAW = 101, /**< See TRACE_DLT_RAW for explanations of pain. */
         TRACE_DLT_C_HDLC = 104,
         TRACE_DLT_IEEE802_11 = 105,
@@ -548 +548 @@
         uint64_t hash; /**< A hash of the packet as supplied by the user */
         int error; /**< The error status of pread_packet */
-        uint64_t internalid;            /** Internal indentifier for the pkt */
+        uint64_t internalid;            /** Internal identifier for the pkt */
         void *srcbucket;
 } libtrace_packet_t;
@@ -576 +576 @@
     TRACE_RADIOTAP_LOCK_QUALITY = 7, /**< Barker Code lock quality (uint16) */
     TRACE_RADIOTAP_TX_ATTENUATION = 8, /**< TX attenuation as unitless distance from max power (uint16) */
-    TRACE_RADIOTAP_DB_TX_ATTENUATION = 9, /**< TX attenutation as dB from max power (uint16) */
+    TRACE_RADIOTAP_DB_TX_ATTENUATION = 9, /**< TX attenuation as dB from max power (uint16) */
     TRACE_RADIOTAP_DBM_TX_POWER = 10, /**< TX Power in dBm (int8) */
     TRACE_RADIOTAP_ANTENNA = 11, /**< Antenna frame was rx'd or tx'd on (uint8) */
@@ -1182 +1182 @@
  * @param [out] format  A pointer that will be updated to point to an allocated 
  *                      string holding the format component of the URI
- * @return NULL if an error occured, otherwise return a pointer to the uridata 
+ * @return NULL if an error occurred, otherwise return a pointer to the uridata 
  * component
  *
@@ -1207 +1207 @@
  *  - rt:hostname:port
  *
- *  If an error occured when attempting to open the trace file, a
+ *  If an error occurred when attempting to open the trace file, a
  *  trace is still returned so trace_is_err() should be called to find out
- *  if an error occured. The trace is created in the configuration state, you 
+ *  if an error occurred. The trace is created in the configuration state, you 
  *  must call trace_start before attempting to read packets from the trace.
  */
@@ -1238 +1238 @@
  *  - pcap:/path/to/pcap/file
  *
- *  If an error occured when attempting to open the output trace, a trace is 
+ *  If an error occurred when attempting to open the output trace, a trace is 
  *  still returned but trace_errno will be set. Use trace_is_err_out() and 
  *  trace_perror_output() to get more information.
@@ -1377 +1377 @@
 } trace_option_output_t;
 
-/* To add a new stat field update this list, and the relevent places in
+/* To add a new stat field update this list, and the relevant places in
  * libtrace_stat_t structure.
  */
@@ -1430 +1430 @@
         uint64_t filtered;
 
-        /** The total number of good packets which have been recevied. Including
+        /** The total number of good packets which have been received. Including
          * those which are dropped and filtered. This does not include errors.
          *
@@ -1610 +1610 @@
  * Returns statistic counters for a trace, for a parallel trace this is a
  * combined total.
- * Where possible these are retrived atomically, however this behaviour depends
+ * Where possible these are retrieved atomically, however this behaviour depends
  * on the underlying trace format.
  *
@@ -1631 +1631 @@
 /**
  * Returns statistic counters for a single thread of a trace.
- * Where possible these are retrived atomically, however this behaviour depends
+ * Where possible these are retrieved atomically, however this behaviour depends
  * on the underlying trace format.
  *
  * @param trace The input trace to examine.
- * @param t An optional thread to received stats for or NULL to retrive stats
+ * @param t An optional thread to received stats for or NULL to retrieve stats
  *          for the current thread
  * @param stats Filled upon return with statistics about the trace, check the
@@ -1734 +1734 @@
  * function should be avoided where possible.
  * 
- * @par The reason you would want to use this function is that a zerocopied
+ * @par The reason you would want to use this function is that a zero-copied
  * packet from a device will be stored using memory owned by the device which
  * may be a limited resource. Copying the packet will ensure that the packet
@@ -2693 +2693 @@
  *
  * @note This function only works for OSPF version 2 packets.
- * @note trace_get_next_ospf_lsa_v2() should be subequently used to process the LSAs
+ * @note trace_get_next_ospf_lsa_v2() should be subsequently used to process the LSAs
  */
 DLLEXPORT SIMPLE_FUNCTION 
@@ -2716 +2716 @@
  *
  * @note This function only works for OSPF version 2 packets.
- * @note trace_get_next_ospf_lsa_header_v2() should be subequently used to process the LSA headers
+ * @note trace_get_next_ospf_lsa_header_v2() should be subsequently used to process the LSA headers
  */
 DLLEXPORT SIMPLE_FUNCTION 
@@ -2739 +2739 @@
  *
  * @note This function only works for OSPF version 2 packets.
- * @note trace_get_next_ospf_link_v2() should be subequently used to process
+ * @note trace_get_next_ospf_link_v2() should be subsequently used to process
  * the links
  */

lib/trace_parallel.c

@@ -1471 +1471 @@
                 libtrace->config.reporter_thold = 100;
         if (libtrace->config.burst_size <= 0)
-                libtrace->config.burst_size = 10;
+                libtrace->config.burst_size = 32;
         if (libtrace->config.thread_cache_size <= 0)
-                libtrace->config.thread_cache_size = 20;
+                libtrace->config.thread_cache_size = 64;
         if (libtrace->config.cache_size <= 0)
                 libtrace->config.cache_size = (libtrace->config.hasher_queue_size + 1) * libtrace->perpkt_thread_count;