target/linux/generic/pending-5.4/690-net-add-support-for-threaded-NAPI-polling.patch

   1 From: Felix Fietkau <nbd@nbd.name>
   2 Date: Sun, 26 Jul 2020 14:03:21 +0200
   3 Subject: [PATCH] net: add support for threaded NAPI polling
   4
   5 For some drivers (especially 802.11 drivers), doing a lot of work in the NAPI
   6 poll function does not perform well. Since NAPI poll is bound to the CPU it
   7 was scheduled from, we can easily end up with a few very busy CPUs spending
   8 most of their time in softirq/ksoftirqd and some idle ones.
   9
  10 Introduce threaded NAPI for such drivers based on a workqueue. The API is the
  11 same except for using netif_threaded_napi_add instead of netif_napi_add.
  12
  13 In my tests with mt76 on MT7621 using threaded NAPI + a thread for tx scheduling
  14 improves LAN->WLAN bridging throughput by 10-50%. Throughput without threaded
  15 NAPI is wildly inconsistent, depending on the CPU that runs the tx scheduling
  16 thread.
  17
  18 With threaded NAPI it seems stable and consistent (and higher than the best
  19 results I got without it).
  20
  21 Based on a patch by Hillf Danton
  22
  23 Cc: Hillf Danton <hdanton@sina.com>
  24 Signed-off-by: Felix Fietkau <nbd@nbd.name>
  25 ---
  26
  27 --- a/include/linux/netdevice.h
  28 +++ b/include/linux/netdevice.h
  29 @@ -340,6 +340,7 @@ struct napi_struct {
  30         struct list_head        dev_list;
  31         struct hlist_node       napi_hash_node;
  32         unsigned int            napi_id;
  33 +       struct work_struct      work;
  34  };
  35
  36  enum {
  37 @@ -350,6 +351,7 @@ enum {
  38         NAPI_STATE_HASHED,      /* In NAPI hash (busy polling possible) */
  39         NAPI_STATE_NO_BUSY_POLL,/* Do not add in napi_hash, no busy polling */
  40         NAPI_STATE_IN_BUSY_POLL,/* sk_busy_loop() owns this NAPI */
  41 +       NAPI_STATE_THREADED,    /* Use threaded NAPI */
  42  };
  43
  44  enum {
  45 @@ -360,6 +362,7 @@ enum {
  46         NAPIF_STATE_HASHED       = BIT(NAPI_STATE_HASHED),
  47         NAPIF_STATE_NO_BUSY_POLL = BIT(NAPI_STATE_NO_BUSY_POLL),
  48         NAPIF_STATE_IN_BUSY_POLL = BIT(NAPI_STATE_IN_BUSY_POLL),
  49 +       NAPIF_STATE_THREADED     = BIT(NAPI_STATE_THREADED),
  50  };
  51
  52  enum gro_result {
  53 @@ -2101,6 +2104,7 @@ struct net_device {
  54         struct lock_class_key   addr_list_lock_key;
  55         bool                    proto_down;
  56         unsigned                wol_enabled:1;
  57 +       unsigned                threaded:1;
  58  };
  59  #define to_net_dev(d) container_of(d, struct net_device, dev)
  60
  61 @@ -2281,6 +2285,26 @@ void netif_napi_add(struct net_device *d
  62                     int (*poll)(struct napi_struct *, int), int weight);
  63
  64  /**
  65 + *     netif_threaded_napi_add - initialize a NAPI context
  66 + *     @dev:  network device
  67 + *     @napi: NAPI context
  68 + *     @poll: polling function
  69 + *     @weight: default weight
  70 + *
  71 + * This variant of netif_napi_add() should be used from drivers using NAPI
  72 + * with CPU intensive poll functions.
  73 + * This will schedule polling from a high priority workqueue
  74 + */
  75 +static inline void netif_threaded_napi_add(struct net_device *dev,
  76 +                                          struct napi_struct *napi,
  77 +                                          int (*poll)(struct napi_struct *, int),
  78 +                                          int weight)
  79 +{
  80 +       set_bit(NAPI_STATE_THREADED, &napi->state);
  81 +       netif_napi_add(dev, napi, poll, weight);
  82 +}
  83 +
  84 +/**
  85   *     netif_tx_napi_add - initialize a NAPI context
  86   *     @dev:  network device
  87   *     @napi: NAPI context
  88 --- a/net/core/dev.c
  89 +++ b/net/core/dev.c
  90 @@ -156,6 +156,7 @@ static DEFINE_SPINLOCK(offload_lock);
  91  struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
  92  struct list_head ptype_all __read_mostly;      /* Taps */
  93  static struct list_head offload_base __read_mostly;
  94 +static struct workqueue_struct *napi_workq __read_mostly;
  95
  96  static int netif_rx_internal(struct sk_buff *skb);
  97  static int call_netdevice_notifiers_info(unsigned long val,
  98 @@ -5912,6 +5913,11 @@ void __napi_schedule(struct napi_struct
  99  {
 100         unsigned long flags;
 101
 102 +       if (test_bit(NAPI_STATE_THREADED, &n->state)) {
 103 +               queue_work(napi_workq, &n->work);
 104 +               return;
 105 +       }
 106 +
 107         local_irq_save(flags);
 108         ____napi_schedule(this_cpu_ptr(&softnet_data), n);
 109         local_irq_restore(flags);
 110 @@ -5959,6 +5965,11 @@ EXPORT_SYMBOL(napi_schedule_prep);
 111   */
 112  void __napi_schedule_irqoff(struct napi_struct *n)
 113  {
 114 +       if (test_bit(NAPI_STATE_THREADED, &n->state)) {
 115 +               queue_work(napi_workq, &n->work);
 116 +               return;
 117 +       }
 118 +
 119         ____napi_schedule(this_cpu_ptr(&softnet_data), n);
 120  }
 121  EXPORT_SYMBOL(__napi_schedule_irqoff);
 122 @@ -6220,9 +6231,89 @@ static void init_gro_hash(struct napi_st
 123         napi->gro_bitmask = 0;
 124  }
 125
 126 +static int __napi_poll(struct napi_struct *n, bool *repoll)
 127 +{
 128 +       int work, weight;
 129 +
 130 +       weight = n->weight;
 131 +
 132 +       /* This NAPI_STATE_SCHED test is for avoiding a race
 133 +        * with netpoll's poll_napi().  Only the entity which
 134 +        * obtains the lock and sees NAPI_STATE_SCHED set will
 135 +        * actually make the ->poll() call.  Therefore we avoid
 136 +        * accidentally calling ->poll() when NAPI is not scheduled.
 137 +        */
 138 +       work = 0;
 139 +       if (test_bit(NAPI_STATE_SCHED, &n->state)) {
 140 +               work = n->poll(n, weight);
 141 +               trace_napi_poll(n, work, weight);
 142 +       }
 143 +
 144 +       WARN_ON_ONCE(work > weight);
 145 +
 146 +       if (likely(work < weight))
 147 +               return work;
 148 +
 149 +       /* Drivers must not modify the NAPI state if they
 150 +        * consume the entire weight.  In such cases this code
 151 +        * still "owns" the NAPI instance and therefore can
 152 +        * move the instance around on the list at-will.
 153 +        */
 154 +       if (unlikely(napi_disable_pending(n))) {
 155 +               napi_complete(n);
 156 +               return work;
 157 +       }
 158 +
 159 +       if (n->gro_bitmask) {
 160 +               /* flush too old packets
 161 +                * If HZ < 1000, flush all packets.
 162 +                */
 163 +               napi_gro_flush(n, HZ >= 1000);
 164 +       }
 165 +
 166 +       gro_normal_list(n);
 167 +
 168 +       *repoll = true;
 169 +
 170 +       return work;
 171 +}
 172 +
 173 +static void napi_workfn(struct work_struct *work)
 174 +{
 175 +       struct napi_struct *n = container_of(work, struct napi_struct, work);
 176 +       void *have;
 177 +
 178 +       for (;;) {
 179 +               bool repoll = false;
 180 +
 181 +               local_bh_disable();
 182 +
 183 +               have = netpoll_poll_lock(n);
 184 +               __napi_poll(n, &repoll);
 185 +               netpoll_poll_unlock(have);
 186 +
 187 +               local_bh_enable();
 188 +
 189 +               if (!repoll)
 190 +                       return;
 191 +
 192 +               if (!need_resched())
 193 +                       continue;
 194 +
 195 +               /*
 196 +                * have to pay for the latency of task switch even if
 197 +                * napi is scheduled
 198 +                */
 199 +               queue_work(napi_workq, work);
 200 +               return;
 201 +       }
 202 +}
 203 +
 204  void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
 205                     int (*poll)(struct napi_struct *, int), int weight)
 206  {
 207 +       if (dev->threaded)
 208 +               set_bit(NAPI_STATE_THREADED, &napi->state);
 209         INIT_LIST_HEAD(&napi->poll_list);
 210         hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
 211         napi->timer.function = napi_watchdog;
 212 @@ -6239,6 +6330,7 @@ void netif_napi_add(struct net_device *d
 213  #ifdef CONFIG_NETPOLL
 214         napi->poll_owner = -1;
 215  #endif
 216 +       INIT_WORK(&napi->work, napi_workfn);
 217         set_bit(NAPI_STATE_SCHED, &napi->state);
 218         set_bit(NAPI_STATE_NPSVC, &napi->state);
 219         list_add_rcu(&napi->dev_list, &dev->napi_list);
 220 @@ -6279,6 +6371,7 @@ static void flush_gro_hash(struct napi_s
 221  void netif_napi_del(struct napi_struct *napi)
 222  {
 223         might_sleep();
 224 +       cancel_work_sync(&napi->work);
 225         if (napi_hash_del(napi))
 226                 synchronize_net();
 227         list_del_init(&napi->dev_list);
 228 @@ -6291,50 +6384,18 @@ EXPORT_SYMBOL(netif_napi_del);
 229
 230  static int napi_poll(struct napi_struct *n, struct list_head *repoll)
 231  {
 232 +       bool do_repoll = false;
 233         void *have;
 234 -       int work, weight;
 235 +       int work;
 236
 237         list_del_init(&n->poll_list);
 238
 239         have = netpoll_poll_lock(n);
 240
 241 -       weight = n->weight;
 242 +       work = __napi_poll(n, &do_repoll);
 243
 244 -       /* This NAPI_STATE_SCHED test is for avoiding a race
 245 -        * with netpoll's poll_napi().  Only the entity which
 246 -        * obtains the lock and sees NAPI_STATE_SCHED set will
 247 -        * actually make the ->poll() call.  Therefore we avoid
 248 -        * accidentally calling ->poll() when NAPI is not scheduled.
 249 -        */
 250 -       work = 0;
 251 -       if (test_bit(NAPI_STATE_SCHED, &n->state)) {
 252 -               work = n->poll(n, weight);
 253 -               trace_napi_poll(n, work, weight);
 254 -       }
 255 -
 256 -       WARN_ON_ONCE(work > weight);
 257 -
 258 -       if (likely(work < weight))
 259 -               goto out_unlock;
 260 -
 261 -       /* Drivers must not modify the NAPI state if they
 262 -        * consume the entire weight.  In such cases this code
 263 -        * still "owns" the NAPI instance and therefore can
 264 -        * move the instance around on the list at-will.
 265 -        */
 266 -       if (unlikely(napi_disable_pending(n))) {
 267 -               napi_complete(n);
 268 +       if (!do_repoll)
 269                 goto out_unlock;
 270 -       }
 271 -
 272 -       if (n->gro_bitmask) {
 273 -               /* flush too old packets
 274 -                * If HZ < 1000, flush all packets.
 275 -                */
 276 -               napi_gro_flush(n, HZ >= 1000);
 277 -       }
 278 -
 279 -       gro_normal_list(n);
 280
 281         /* Some drivers may have called napi_schedule
 282          * prior to exhausting their budget.
 283 @@ -10314,6 +10375,10 @@ static int __init net_dev_init(void)
 284                 sd->backlog.weight = weight_p;
 285         }
 286
 287 +       napi_workq = alloc_workqueue("napi_workq", WQ_UNBOUND | WQ_HIGHPRI,
 288 +                                    WQ_UNBOUND_MAX_ACTIVE | WQ_SYSFS);
 289 +       BUG_ON(!napi_workq);
 290 +
 291         dev_boot_phase = 0;
 292
 293         /* The loopback device is special if any other network devices
 294 --- a/net/core/net-sysfs.c
 295 +++ b/net/core/net-sysfs.c
 296 @@ -442,6 +442,52 @@ static ssize_t proto_down_store(struct d
 297  }
 298  NETDEVICE_SHOW_RW(proto_down, fmt_dec);
 299
 300 +static int change_napi_threaded(struct net_device *dev, unsigned long val)
 301 +{
 302 +       struct napi_struct *napi;
 303 +
 304 +       if (list_empty(&dev->napi_list))
 305 +               return -EOPNOTSUPP;
 306 +
 307 +       list_for_each_entry(napi, &dev->napi_list, dev_list) {
 308 +               if (val)
 309 +                       set_bit(NAPI_STATE_THREADED, &napi->state);
 310 +               else
 311 +                       clear_bit(NAPI_STATE_THREADED, &napi->state);
 312 +       }
 313 +
 314 +       return 0;
 315 +}
 316 +
 317 +static ssize_t napi_threaded_store(struct device *dev,
 318 +                               struct device_attribute *attr,
 319 +                               const char *buf, size_t len)
 320 +{
 321 +       return netdev_store(dev, attr, buf, len, change_napi_threaded);
 322 +}
 323 +
 324 +static ssize_t napi_threaded_show(struct device *dev,
 325 +                                 struct device_attribute *attr,
 326 +                                 char *buf)
 327 +{
 328 +       struct net_device *netdev = to_net_dev(dev);
 329 +       struct napi_struct *napi;
 330 +       bool enabled = false;
 331 +
 332 +       if (!rtnl_trylock())
 333 +               return restart_syscall();
 334 +
 335 +       list_for_each_entry(napi, &netdev->napi_list, dev_list) {
 336 +               if (test_bit(NAPI_STATE_THREADED, &napi->state))
 337 +                       enabled = true;
 338 +       }
 339 +
 340 +       rtnl_unlock();
 341 +
 342 +       return sprintf(buf, fmt_dec, enabled);
 343 +}
 344 +static DEVICE_ATTR_RW(napi_threaded);
 345 +
 346  static ssize_t phys_port_id_show(struct device *dev,
 347                                  struct device_attribute *attr, char *buf)
 348  {
 349 @@ -532,6 +578,7 @@ static struct attribute *net_class_attrs
 350         &dev_attr_flags.attr,
 351         &dev_attr_tx_queue_len.attr,
 352         &dev_attr_gro_flush_timeout.attr,
 353 +       &dev_attr_napi_threaded.attr,
 354         &dev_attr_phys_port_id.attr,
 355         &dev_attr_phys_port_name.attr,
 356         &dev_attr_phys_switch_id.attr,