Merge branch 'packetmill'

tbarbette · Apr 21, 2021 · e0a81f3 · e0a81f3
2 parents 828d726 + 8d02cd9
commit e0a81f3
Show file tree

Hide file tree

Showing 14 changed files with 226 additions and 57 deletions.
diff --git a/README.md b/README.md
@@ -1,38 +1,75 @@
-[FastClick](https://www.fastclick.dev) ![CI](https://github.com/tbarbette/fastclick/workflows/C/C++%20CI/badge.svg)
-=========
-This is an extended version of the Click Modular Router featuring an
-improved Netmap support and a new DPDK support. It was the result of
-our ANCS paper available at http://hdl.handle.net/2268/181954, but received
-multiple contributions and improvements since then.
+# FastClick (PacketMill)
 
-The [Wiki](https://github.com/tbarbette/fastclick/wiki) provides documentation about the elements and how to use some FastClick features
-such as batching.
+This repo is a modified FastClick that comes with some additional source-code optimization techniques to improve the performance of high-speed packet processing. Additionally, it implements different metadata management models, i.e., copying, overlaying, and X-Change.
 
-Announcements
--------------
-Be sure to watch the repository and check out the [GitHub Discussions](https://github.com/tbarbette/fastclick/discussions) to stay up to date!
+For more information, please refer to PacketMill's [paper][packetmill-paper] and [repo][packetmill-repo].
 
-Quick start for DPDK
---------------------
 
- * Install DPDK's dependencies (`sudo apt install libelf-dev build-essential pkg-config zlib1g-dev libnuma-dev`)
- * Install DPDK (http://core.dpdk.org/doc/quick-start/). Since 20.11 you have to use meson : `meson build && cd build && ninja && sudo ninja install`
- * Build FastClick, with support for DPDK using the following command:
+## Source-code Modifications
 
-```
-./configure --enable-dpdk --enable-intel-cpu --verbose --enable-select=poll CFLAGS="-O3" CXXFLAGS="-std=c++11 -O3"  --disable-dynamic-linking --enable-poll --enable-bound-port-transfer --enable-local --enable-flow --disable-task-stats --disable-cpu-load
+PacketMill performs multiple source-code optimizations that exploit the availabe information in a given NF configuration file. We have implemented these optimizations on top of `click-devirtualize`. 
+
+To use these optimization, you need to build and install FastClick as follows:
+
+
+```bash
+git clone --branch packetmill git@github.com:tbarbette/fastclick.git
+cd fastclick
+./configure --disable-linuxmodule --enable-userlevel --enable-user-multithread --enable-etherswitch --disable-dynamic-linking --enable-local --enable-dpdk --enable-research --enable-flow --disable-task-stats --enable-cpu-load --prefix $(pwd)/build/ --enable-intel-cpu CXX="clang++ -fno-access-control" CC="clang" CXXFLAGS="-std=gnu++14 -O3" --disable-bound-port-transfer --enable-dpdk-pool --disable-dpdk-packet
 make
+sudo make uninstall
+sudo make install
 ```
-  * Since DPDK is using Meson and pkg-config, to compile against various, or non-globally installed DPDK versions, one can prepend `PKG_CONFIG_PATH=path/to/libpdpdk.pc/../` to both configure and make.
 
-*You will find more information in the [High-Speed I/O wiki page](https://github.com/tbarbette/fastclick/wiki/High-speed-I-O).*
+You need to define `RTE_SDK` and `RTE_TARGET` before configuring FastClick.
+
+**Note: PacketMill's [repo][packetmill-repo] offers a automated pipeline/workflow for building/installing PacketMill (FastClick + X-Change) and performing some experiments related to PacketMill's PacketMill's [paper][packetmill-paper].**
+
+### Devirtualize Pass
 
-FastClick "Light"
------------------
-FastClick, like Click comes with a lot of features that you may not use. The following options will improve performance further :
+This optimization pass removes virtual function calls from elements' source code based on the input configuration file. More specifically, it duplicates the source code, override some methods, and defines the type of called pointers & the called functions. The following code snippet shows the source code of `FromDPDKDevice::output_push` defined by `click-devirtualize` after applying the pass.
+
+```cpp
+inline void
+FromDPDKDevice_a_afd0::output_push(int i, Packet *p) const
+{
+  if (i == 0) { ((Classifier_a_aFNT_a3_sc0 *)output(i).element())->Classifier_a_aFNT_a3_sc0::push(0, p); return; }
+  output(i).push(p);
+}
 ```
-./configure --enable-dpdk --enable-intel-cpu --verbose --enable-select=poll CFLAGS="-O3" CXXFLAGS="-std=c++11 -O3"  --disable-dynamic-linking --enable-poll --enable-bound-port-transfer --enable-local --enable-flow --disable-task-stats --disable-cpu-load --enable-dpdk-packet --disable-clone --disable-dpdk-softqueue
-make
+
+The new source code replaces the actual call as opposed to the normal FastClick code that uses `output(port).push_batch(batch);`. This pass has been originally introduced by `click-devirtualize`. We have adopted it and modified it to work with FastClick. Fore more information, please check click-devirtualize [paper][devirtualize-paper].
+
+To use this pass, run the following commands:
+
+**Note that you need to compile & install FastClick before applying any pass.**
+
+```bash
+sudo bin/click-devirtualize CONFIG  > package.uo
+ar x package.uo config
+cd userlevel
+sudo ../bin/click-mkmindriver -V -C .. -p embed --ship -u ../package.uo
+make embedclick MINDRIVER=embed STATIC=0
+```
+
+After building `embedclick`, you can run the new click binary with the new configuration file. For instance, you can run the following command:
+
+```bash
+sudo userlevel/embedclick --dpdk -l 0-35 -n 6 -w 0000:17:00.0 -v -- config
+```
+
+To see the changes to the source code, you can check `clickdv*.cc` and `clickdv*.hh` files in `userlevel/`. 
+
+### Replace Pass
+
+This pass replaces the variables with their available value in the configuration file. For instance, the following code snippet shows a part `FromDPDKDevice::run_task` function. Replace pass substitutes `_burst` variable by its value, `32`, which is specified in the input configuration file.
+
+```cpp
+// FastClick
+unsigned n = rte_eth_rx_burst(_dev->port_id, iqueue, pkts, _burst);
+
+// click-devirtualize + replace pass
+unsigned n = rte_eth_rx_burst(_dev->port_id, iqueue, pkts, 32);
 ```
  * Disable task stats suppress statistics tracking for advanced task scheduling with e.g. BalancedThreadSched. With DPDK, it's polling anyway... And as far as scheduling is concerned, RSS++ has a better solution.
  * Disable CPU load will remove load tracking. That is accounting for a CPU percentage while using DPDK by counting cycles spent in empty runs vs all runs. Accessible with the "load" handler.

diff --git a/click-buildtool.in b/click-buildtool.in
@@ -34,6 +34,10 @@ elem2=""
 default_provisions="@provisions@"
 driver_provisions="@DRIVERS@"
 
+if [ -z "$CLICK_ELEM_RAND_MAX" ] ; then
+    CLICK_ELEM_RAND_MAX=0
+fi
+
 export LC_COLLATE=C
 trap "exit 1" HUP
 

diff --git a/config-userlevel.h.in b/config-userlevel.h.in
@@ -274,6 +274,9 @@
 /* Define if a Click user-level driver manages Intel DPDK packet pools. */
 #undef CLICK_PACKET_USE_DPDK
 
+/* Define if a Click packet is inside DPDK packet. */
+#undef CLICK_PACKET_INSIDE_DPDK
+
 /* Define if Click should use Valgrind client requests. */
 #undef HAVE_VALGRIND
 

diff --git a/config.h.in b/config.h.in
@@ -144,6 +144,9 @@
 /* Define if the RE2 library is present. */
 #undef HAVE_RE2
 
+/* Define if elements have random alignment. */
+#undef HAVE_RAND_ALIGN
+
 /* Define if you want to use the stride scheduler. */
 #undef HAVE_STRIDE_SCHED
 

diff --git a/configure b/configure
@@ -725,6 +725,7 @@ minios_dir
 xen_dir
 INCLUDE_KSYMS
 LINUXMODULE_FIXINCLUDES
+DO_RAND_ALIGN
 PTHREAD_LIBS
 HAVE_BATCH
 AR_CREATEFLAGS
@@ -820,6 +821,7 @@ enable_batch
 enable_verbose_batch
 enable_auto_batch
 enable_netmap_pool
+enable_rand_align
 enable_select
 enable_poll
 enable_kqueue
@@ -1553,6 +1555,7 @@ Optional Features:
     --enable-auto-batch=[list|jump|port]
                           make vanilla elements batch-compatible automatically
     --enable-netmap-pool  use netmap buffers instead of standard Click buffers
+    --enable-rand-align   enable random alignment of element
     --enable-select=[select|poll|kqueue]
                           set file descriptor wait mechanism
     --disable-select      do not use select()
@@ -8473,6 +8476,22 @@ else
     have_pci=no
 fi
 
+# Check whether --enable-rand-align was given.
+if test "${enable_rand_align+set}" = set; then :
+  enableval=$enable_rand_align; :
+else
+  enable_rand_align=no
+fi
+
+if test "x$enable_rand_align" = "xyes"; then
+    $as_echo "#define HAVE_RAND_ALIGN 1" >>confdefs.h
+
+    DO_RAND_ALIGN=1
+
+else
+    DO_RAND_ALIGN=0
+
+fi
 
 # Check whether --enable-select was given.
 if test "${enable_select+set}" = set; then :

diff --git a/configure.in b/configure.in
@@ -278,6 +278,15 @@ else
     have_pci=no
 fi
 
+AC_ARG_ENABLE([rand-align],
+    [AS_HELP_STRING([  --enable-rand-align], [enable random alignment of element])],
+    [:], [enable_rand_align=no])
+if test "x$enable_rand_align" = "xyes"; then
+    AC_DEFINE([HAVE_RAND_ALIGN])
+    AC_SUBST(DO_RAND_ALIGN,1)
+else
+    AC_SUBST(DO_RAND_ALIGN,0)
+fi
 
 AC_ARG_ENABLE([select],
     [AS_HELP_STRING([  --enable-select=[[select|poll|kqueue]]], [set file descriptor wait mechanism])

diff --git a/elements/userlevel/fromdpdkdevice.cc b/elements/userlevel/fromdpdkdevice.cc
@@ -322,36 +322,55 @@ void FromDPDKDevice::clear_buffers() {
         click_chatter("Cleared %d buffers for queue %d",tot,q);
     }
 }
+#ifdef DPDK_USE_XCHG
+extern "C" {
+#include <mlx5_xchg.h>
+}
+#endif
 
-bool FromDPDKDevice::run_task(Task *t)
-{
-    struct rte_mbuf *pkts[_burst];
-    int ret = 0;
+bool FromDPDKDevice::run_task(Task *t) {
+  struct rte_mbuf *pkts[_burst];
+  int ret = 0;
 
-    for (int iqueue = queue_for_thisthread_begin();
-            iqueue<=queue_for_thisthread_end(); iqueue++) {
+  int iqueue = queue_for_thisthread_begin();
+  { //This version differs from multi by having support for one queue per thread only, which is extremly usual
 #if HAVE_BATCH
-         PacketBatch* head = 0;
-         WritablePacket *last;
+  PacketBatch *head = 0;
+  WritablePacket *last;
 #endif
-        unsigned n = rte_eth_rx_burst(_dev->port_id, iqueue, pkts, _burst);
-        for (unsigned i = 0; i < n; ++i) {
-            unsigned char* data = rte_pktmbuf_mtod(pkts[i], unsigned char *);
-            rte_prefetch0(data);
+
+#ifdef DPDK_USE_XCHG
+ unsigned n = rte_mlx5_rx_burst_xchg(_dev->port_id, iqueue, (struct xchg**)pkts, _burst);
+#else
+ unsigned n = rte_eth_rx_burst(_dev->port_id, iqueue, pkts, _burst);
+#endif
+
+for (unsigned i = 0; i < n; ++i) {
+    unsigned char *data = rte_pktmbuf_mtod(pkts[i], unsigned char *);
+    rte_prefetch0(data);
 #if CLICK_PACKET_USE_DPDK
     WritablePacket *p = static_cast<WritablePacket *>(Packet::make(pkts[i], _clear));
 #elif HAVE_ZEROCOPY
-            WritablePacket *p = Packet::make(data,
-                     rte_pktmbuf_data_len(pkts[i]),
-                     DPDKDevice::free_pkt,
-                     pkts[i],
-                     rte_pktmbuf_headroom(pkts[i]),
-                     rte_pktmbuf_tailroom(pkts[i]),
-                     _clear
-                     );
+
+# if CLICK_PACKET_INSIDE_DPDK
+    WritablePacket *p =(WritablePacket*)( pkts[i] + 1);
+    new (p) WritablePacket();
+
+    p->initialize(_clear);
+    p->set_buffer((unsigned char*)(pkts[i]->buf_addr), DPDKDevice::MBUF_DATA_SIZE);
+    p->set_data(data);
+    p->set_data_length(rte_pktmbuf_data_len(pkts[i]));
+    p->set_buffer_destructor(DPDKDevice::free_pkt);
+
+    p->set_destructor_argument(pkts[i]);
+# else
+    WritablePacket *p = Packet::make(
+        data, rte_pktmbuf_data_len(pkts[i]), DPDKDevice::free_pkt, pkts[i],
+        rte_pktmbuf_headroom(pkts[i]), rte_pktmbuf_tailroom(pkts[i]), _clear);
+# endif
 #else
             WritablePacket *p = Packet::make(data,
-                                     (uint32_t)rte_pktmbuf_pkt_len(pkts[i]), _clear);
+                                     (uint32_t)rte_pktmbuf_pkt_len(pkts[i]));
             rte_pktmbuf_free(pkts[i]);
             data = p->data();
 #endif

diff --git a/elements/userlevel/todpdkdevice.cc b/elements/userlevel/todpdkdevice.cc
@@ -223,6 +223,13 @@ void ToDPDKDevice::run_timer(Timer *)
     flush_internal_tx_queue(_iqueues.get());
 }
 
+# ifdef DPDK_USE_XCHG
+extern "C" {
+#  include <mlx5_xchg.h>
+}
+# endif
+
+
 /* Flush as much as possible packets from a given internal queue to the DPDK
  * device. */
 void ToDPDKDevice::flush_internal_tx_queue(DPDKDevice::TXInternalQueue &iqueue) {
@@ -245,8 +252,15 @@ void ToDPDKDevice::flush_internal_tx_queue(DPDKDevice::TXInternalQueue &iqueue)
             // The sub_burst wraps around the ring
             sub_burst = _internal_tx_queue_size - iqueue.index;
         //Todo : if there is multiple queue assigned to this thread, send on all of them
+
+
+# ifdef DPDK_USE_XCHG
+        r = rte_mlx5_tx_burst_xchg(_dev->port_id, queue_for_thisthread_begin(),(struct xchg**) &iqueue.pkts[iqueue.index], sub_burst);
+# else
         r = rte_eth_tx_burst(_dev->port_id, queue_for_thisthread_begin(), &iqueue.pkts[iqueue.index],
-                             sub_burst);
+                            sub_burst);
+# endif
+
         iqueue.nr_pending -= r;
         iqueue.index += r;
 
@@ -302,12 +316,12 @@ void ToDPDKDevice::push(int, Packet *p)
         // If we're in blocking mode, we loop until we can put p in the iqueue
     } while (unlikely(_blocking && congestioned));
 
-#if !CLICK_PACKET_USE_DPDK
+# if !CLICK_PACKET_USE_DPDK && !CLICK_PACKET_INSIDE_DPDK
     if (likely(is_fullpush()))
         p->kill_nonatomic();
     else
         p->kill();
-#endif
+# endif
 }
 
 /**
@@ -327,7 +341,7 @@ void ToDPDKDevice::push_batch(int, PacketBatch *head)
 
     //No recycling through click if we have DPDK-backed packets
     bool congestioned;
-#  if !CLICK_PACKET_USE_DPDK
+#  if !CLICK_PACKET_USE_DPDK && !CLICK_PACKET_INSIDE_DPDK
     BATCH_RECYCLE_START();
 #  endif
     do {
@@ -344,7 +358,8 @@ void ToDPDKDevice::push_batch(int, PacketBatch *head)
                 abort();
             }
             next = p->next();
-#  if !CLICK_PACKET_USE_DPDK
+
+#  if !CLICK_PACKET_USE_DPDK && !CLICK_PACKET_INSIDE_DPDK
             BATCH_RECYCLE_PACKET_CONTEXT(p);
 #  endif
             p = next;
@@ -373,7 +388,7 @@ void ToDPDKDevice::push_batch(int, PacketBatch *head)
         // If we're in blocking mode, we loop until we can put p in the iqueue
     } while (unlikely(_blocking && congestioned));
 
-#  if !CLICK_PACKET_USE_DPDK
+#  if !CLICK_PACKET_USE_DPDK && !CLICK_PACKET_INSIDE_DPDK
     //If non-blocking, drop all packets that could not be sent
     while (p) {
         next = p->next();
@@ -383,7 +398,7 @@ void ToDPDKDevice::push_batch(int, PacketBatch *head)
     }
 #  endif
 
-#  if !CLICK_PACKET_USE_DPDK
+#  if !CLICK_PACKET_USE_DPDK && !CLICK_PACKET_INSIDE_DPDK
     BATCH_RECYCLE_END();
 #  endif
 }