kernel/bpf/cpumap.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* bpf/cpumap.c
   3  *
   4  * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
   5  */
   6
   7 /**
   8  * DOC: cpu map
   9  * The 'cpumap' is primarily used as a backend map for XDP BPF helper
  10  * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
  11  *
  12  * Unlike devmap which redirects XDP frames out to another NIC device,
  13  * this map type redirects raw XDP frames to another CPU.  The remote
  14  * CPU will do SKB-allocation and call the normal network stack.
  15  */
  16 /*
  17  * This is a scalability and isolation mechanism, that allow
  18  * separating the early driver network XDP layer, from the rest of the
  19  * netstack, and assigning dedicated CPUs for this stage.  This
  20  * basically allows for 10G wirespeed pre-filtering via bpf.
  21  */
  22 #include <linux/bitops.h>
  23 #include <linux/bpf.h>
  24 #include <linux/filter.h>
  25 #include <linux/ptr_ring.h>
  26 #include <net/xdp.h>
  27 #include <net/hotdata.h>
  28
  29 #include <linux/sched.h>
  30 #include <linux/workqueue.h>
  31 #include <linux/kthread.h>
  32 #include <linux/completion.h>
  33 #include <trace/events/xdp.h>
  34 #include <linux/btf_ids.h>
  35
  36 #include <linux/netdevice.h>   /* netif_receive_skb_list */
  37 #include <linux/etherdevice.h> /* eth_type_trans */
  38
  39 /* General idea: XDP packets getting XDP redirected to another CPU,
  40  * will maximum be stored/queued for one driver ->poll() call.  It is
  41  * guaranteed that queueing the frame and the flush operation happen on
  42  * same CPU.  Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
  43  * which queue in bpf_cpu_map_entry contains packets.
  44  */
  45
  46 #define CPU_MAP_BULK_SIZE 8  /* 8 == one cacheline on 64-bit archs */
  47 struct bpf_cpu_map_entry;
  48 struct bpf_cpu_map;
  49
  50 struct xdp_bulk_queue {
  51         void *q[CPU_MAP_BULK_SIZE];
  52         struct list_head flush_node;
  53         struct bpf_cpu_map_entry *obj;
  54         unsigned int count;
  55 };
  56
  57 /* Struct for every remote "destination" CPU in map */
  58 struct bpf_cpu_map_entry {
  59         u32 cpu;    /* kthread CPU and map index */
  60         int map_id; /* Back reference to map */
  61
  62         /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
  63         struct xdp_bulk_queue __percpu *bulkq;
  64
  65         /* Queue with potential multi-producers, and single-consumer kthread */
  66         struct ptr_ring *queue;
  67         struct task_struct *kthread;
  68
  69         struct bpf_cpumap_val value;
  70         struct bpf_prog *prog;
  71
  72         struct completion kthread_running;
  73         struct rcu_work free_work;
  74 };
  75
  76 struct bpf_cpu_map {
  77         struct bpf_map map;
  78         /* Below members specific for map type */
  79         struct bpf_cpu_map_entry __rcu **cpu_map;
  80 };
  81
  82 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
  83 {
  84         u32 value_size = attr->value_size;
  85         struct bpf_cpu_map *cmap;
  86
  87         /* check sanity of attributes */
  88         if (attr->max_entries == 0 || attr->key_size != 4 ||
  89             (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
  90              value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
  91             attr->map_flags & ~BPF_F_NUMA_NODE)
  92                 return ERR_PTR(-EINVAL);
  93
  94         /* Pre-limit array size based on NR_CPUS, not final CPU check */
  95         if (attr->max_entries > NR_CPUS)
  96                 return ERR_PTR(-E2BIG);
  97
  98         cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
  99         if (!cmap)
 100                 return ERR_PTR(-ENOMEM);
 101
 102         bpf_map_init_from_attr(&cmap->map, attr);
 103
 104         /* Alloc array for possible remote "destination" CPUs */
 105         cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
 106                                            sizeof(struct bpf_cpu_map_entry *),
 107                                            cmap->map.numa_node);
 108         if (!cmap->cpu_map) {
 109                 bpf_map_area_free(cmap);
 110                 return ERR_PTR(-ENOMEM);
 111         }
 112
 113         return &cmap->map;
 114 }
 115
 116 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
 117 {
 118         /* The tear-down procedure should have made sure that queue is
 119          * empty.  See __cpu_map_entry_replace() and work-queue
 120          * invoked cpu_map_kthread_stop(). Catch any broken behaviour
 121          * gracefully and warn once.
 122          */
 123         void *ptr;
 124
 125         while ((ptr = ptr_ring_consume(ring))) {
 126                 WARN_ON_ONCE(1);
 127                 if (unlikely(__ptr_test_bit(0, &ptr))) {
 128                         __ptr_clear_bit(0, &ptr);
 129                         kfree_skb(ptr);
 130                         continue;
 131                 }
 132                 xdp_return_frame(ptr);
 133         }
 134 }
 135
 136 static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
 137                                      struct list_head *listp,
 138                                      struct xdp_cpumap_stats *stats)
 139 {
 140         struct sk_buff *skb, *tmp;
 141         struct xdp_buff xdp;
 142         u32 act;
 143         int err;
 144
 145         list_for_each_entry_safe(skb, tmp, listp, list) {
 146                 act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
 147                 switch (act) {
 148                 case XDP_PASS:
 149                         break;
 150                 case XDP_REDIRECT:
 151                         skb_list_del_init(skb);
 152                         err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
 153                                                       rcpu->prog);
 154                         if (unlikely(err)) {
 155                                 kfree_skb(skb);
 156                                 stats->drop++;
 157                         } else {
 158                                 stats->redirect++;
 159                         }
 160                         return;
 161                 default:
 162                         bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
 163                         fallthrough;
 164                 case XDP_ABORTED:
 165                         trace_xdp_exception(skb->dev, rcpu->prog, act);
 166                         fallthrough;
 167                 case XDP_DROP:
 168                         skb_list_del_init(skb);
 169                         kfree_skb(skb);
 170                         stats->drop++;
 171                         return;
 172                 }
 173         }
 174 }
 175
 176 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 177                                     void **frames, int n,
 178                                     struct xdp_cpumap_stats *stats)
 179 {
 180         struct xdp_rxq_info rxq = {};
 181         struct xdp_buff xdp;
 182         int i, nframes = 0;
 183
 184         xdp_set_return_frame_no_direct();
 185         xdp.rxq = &rxq;
 186
 187         for (i = 0; i < n; i++) {
 188                 struct xdp_frame *xdpf = frames[i];
 189                 u32 act;
 190                 int err;
 191
 192                 rxq.dev = xdpf->dev_rx;
 193                 rxq.mem.type = xdpf->mem_type;
 194                 /* TODO: report queue_index to xdp_rxq_info */
 195
 196                 xdp_convert_frame_to_buff(xdpf, &xdp);
 197
 198                 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
 199                 switch (act) {
 200                 case XDP_PASS:
 201                         err = xdp_update_frame_from_buff(&xdp, xdpf);
 202                         if (err < 0) {
 203                                 xdp_return_frame(xdpf);
 204                                 stats->drop++;
 205                         } else {
 206                                 frames[nframes++] = xdpf;
 207                                 stats->pass++;
 208                         }
 209                         break;
 210                 case XDP_REDIRECT:
 211                         err = xdp_do_redirect(xdpf->dev_rx, &xdp,
 212                                               rcpu->prog);
 213                         if (unlikely(err)) {
 214                                 xdp_return_frame(xdpf);
 215                                 stats->drop++;
 216                         } else {
 217                                 stats->redirect++;
 218                         }
 219                         break;
 220                 default:
 221                         bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
 222                         fallthrough;
 223                 case XDP_DROP:
 224                         xdp_return_frame(xdpf);
 225                         stats->drop++;
 226                         break;
 227                 }
 228         }
 229
 230         xdp_clear_return_frame_no_direct();
 231
 232         return nframes;
 233 }
 234
 235 #define CPUMAP_BATCH 8
 236
 237 static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
 238                                 int xdp_n, struct xdp_cpumap_stats *stats,
 239                                 struct list_head *list)
 240 {
 241         struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
 242         int nframes;
 243
 244         if (!rcpu->prog)
 245                 return xdp_n;
 246
 247         rcu_read_lock_bh();
 248         bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
 249
 250         nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
 251
 252         if (stats->redirect)
 253                 xdp_do_flush();
 254
 255         if (unlikely(!list_empty(list)))
 256                 cpu_map_bpf_prog_run_skb(rcpu, list, stats);
 257
 258         bpf_net_ctx_clear(bpf_net_ctx);
 259         rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
 260
 261         return nframes;
 262 }
 263
 264 static int cpu_map_kthread_run(void *data)
 265 {
 266         struct bpf_cpu_map_entry *rcpu = data;
 267         unsigned long last_qs = jiffies;
 268
 269         complete(&rcpu->kthread_running);
 270         set_current_state(TASK_INTERRUPTIBLE);
 271
 272         /* When kthread gives stop order, then rcpu have been disconnected
 273          * from map, thus no new packets can enter. Remaining in-flight
 274          * per CPU stored packets are flushed to this queue.  Wait honoring
 275          * kthread_stop signal until queue is empty.
 276          */
 277         while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
 278                 struct xdp_cpumap_stats stats = {}; /* zero stats */
 279                 unsigned int kmem_alloc_drops = 0, sched = 0;
 280                 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
 281                 int i, n, m, nframes, xdp_n;
 282                 void *frames[CPUMAP_BATCH];
 283                 void *skbs[CPUMAP_BATCH];
 284                 LIST_HEAD(list);
 285
 286                 /* Release CPU reschedule checks */
 287                 if (__ptr_ring_empty(rcpu->queue)) {
 288                         set_current_state(TASK_INTERRUPTIBLE);
 289                         /* Recheck to avoid lost wake-up */
 290                         if (__ptr_ring_empty(rcpu->queue)) {
 291                                 schedule();
 292                                 sched = 1;
 293                                 last_qs = jiffies;
 294                         } else {
 295                                 __set_current_state(TASK_RUNNING);
 296                         }
 297                 } else {
 298                         rcu_softirq_qs_periodic(last_qs);
 299                         sched = cond_resched();
 300                 }
 301
 302                 /*
 303                  * The bpf_cpu_map_entry is single consumer, with this
 304                  * kthread CPU pinned. Lockless access to ptr_ring
 305                  * consume side valid as no-resize allowed of queue.
 306                  */
 307                 n = __ptr_ring_consume_batched(rcpu->queue, frames,
 308                                                CPUMAP_BATCH);
 309                 for (i = 0, xdp_n = 0; i < n; i++) {
 310                         void *f = frames[i];
 311                         struct page *page;
 312
 313                         if (unlikely(__ptr_test_bit(0, &f))) {
 314                                 struct sk_buff *skb = f;
 315
 316                                 __ptr_clear_bit(0, &skb);
 317                                 list_add_tail(&skb->list, &list);
 318                                 continue;
 319                         }
 320
 321                         frames[xdp_n++] = f;
 322                         page = virt_to_page(f);
 323
 324                         /* Bring struct page memory area to curr CPU. Read by
 325                          * build_skb_around via page_is_pfmemalloc(), and when
 326                          * freed written by page_frag_free call.
 327                          */
 328                         prefetchw(page);
 329                 }
 330
 331                 /* Support running another XDP prog on this CPU */
 332                 nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
 333                 if (nframes) {
 334                         m = kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
 335                                                   gfp, nframes, skbs);
 336                         if (unlikely(m == 0)) {
 337                                 for (i = 0; i < nframes; i++)
 338                                         skbs[i] = NULL; /* effect: xdp_return_frame */
 339                                 kmem_alloc_drops += nframes;
 340                         }
 341                 }
 342
 343                 local_bh_disable();
 344                 for (i = 0; i < nframes; i++) {
 345                         struct xdp_frame *xdpf = frames[i];
 346                         struct sk_buff *skb = skbs[i];
 347
 348                         skb = __xdp_build_skb_from_frame(xdpf, skb,
 349                                                          xdpf->dev_rx);
 350                         if (!skb) {
 351                                 xdp_return_frame(xdpf);
 352                                 continue;
 353                         }
 354
 355                         list_add_tail(&skb->list, &list);
 356                 }
 357
 358                 /* Feedback loop via tracepoint.
 359                  * NB: keep before recv to allow measuring enqueue/dequeue latency.
 360                  */
 361                 trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
 362                                          sched, &stats);
 363
 364                 netif_receive_skb_list(&list);
 365                 local_bh_enable(); /* resched point, may call do_softirq() */
 366         }
 367         __set_current_state(TASK_RUNNING);
 368
 369         return 0;
 370 }
 371
 372 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
 373                                       struct bpf_map *map, int fd)
 374 {
 375         struct bpf_prog *prog;
 376
 377         prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
 378         if (IS_ERR(prog))
 379                 return PTR_ERR(prog);
 380
 381         if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
 382             !bpf_prog_map_compatible(map, prog)) {
 383                 bpf_prog_put(prog);
 384                 return -EINVAL;
 385         }
 386
 387         rcpu->value.bpf_prog.id = prog->aux->id;
 388         rcpu->prog = prog;
 389
 390         return 0;
 391 }
 392
 393 static struct bpf_cpu_map_entry *
 394 __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
 395                       u32 cpu)
 396 {
 397         int numa, err, i, fd = value->bpf_prog.fd;
 398         gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
 399         struct bpf_cpu_map_entry *rcpu;
 400         struct xdp_bulk_queue *bq;
 401
 402         /* Have map->numa_node, but choose node of redirect target CPU */
 403         numa = cpu_to_node(cpu);
 404
 405         rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
 406         if (!rcpu)
 407                 return NULL;
 408
 409         /* Alloc percpu bulkq */
 410         rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
 411                                            sizeof(void *), gfp);
 412         if (!rcpu->bulkq)
 413                 goto free_rcu;
 414
 415         for_each_possible_cpu(i) {
 416                 bq = per_cpu_ptr(rcpu->bulkq, i);
 417                 bq->obj = rcpu;
 418         }
 419
 420         /* Alloc queue */
 421         rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
 422                                            numa);
 423         if (!rcpu->queue)
 424                 goto free_bulkq;
 425
 426         err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
 427         if (err)
 428                 goto free_queue;
 429
 430         rcpu->cpu    = cpu;
 431         rcpu->map_id = map->id;
 432         rcpu->value.qsize  = value->qsize;
 433
 434         if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
 435                 goto free_ptr_ring;
 436
 437         /* Setup kthread */
 438         init_completion(&rcpu->kthread_running);
 439         rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
 440                                                "cpumap/%d/map:%d", cpu,
 441                                                map->id);
 442         if (IS_ERR(rcpu->kthread))
 443                 goto free_prog;
 444
 445         /* Make sure kthread runs on a single CPU */
 446         kthread_bind(rcpu->kthread, cpu);
 447         wake_up_process(rcpu->kthread);
 448
 449         /* Make sure kthread has been running, so kthread_stop() will not
 450          * stop the kthread prematurely and all pending frames or skbs
 451          * will be handled by the kthread before kthread_stop() returns.
 452          */
 453         wait_for_completion(&rcpu->kthread_running);
 454
 455         return rcpu;
 456
 457 free_prog:
 458         if (rcpu->prog)
 459                 bpf_prog_put(rcpu->prog);
 460 free_ptr_ring:
 461         ptr_ring_cleanup(rcpu->queue, NULL);
 462 free_queue:
 463         kfree(rcpu->queue);
 464 free_bulkq:
 465         free_percpu(rcpu->bulkq);
 466 free_rcu:
 467         kfree(rcpu);
 468         return NULL;
 469 }
 470
 471 static void __cpu_map_entry_free(struct work_struct *work)
 472 {
 473         struct bpf_cpu_map_entry *rcpu;
 474
 475         /* This cpu_map_entry have been disconnected from map and one
 476          * RCU grace-period have elapsed. Thus, XDP cannot queue any
 477          * new packets and cannot change/set flush_needed that can
 478          * find this entry.
 479          */
 480         rcpu = container_of(to_rcu_work(work), struct bpf_cpu_map_entry, free_work);
 481
 482         /* kthread_stop will wake_up_process and wait for it to complete.
 483          * cpu_map_kthread_run() makes sure the pointer ring is empty
 484          * before exiting.
 485          */
 486         kthread_stop(rcpu->kthread);
 487
 488         if (rcpu->prog)
 489                 bpf_prog_put(rcpu->prog);
 490         /* The queue should be empty at this point */
 491         __cpu_map_ring_cleanup(rcpu->queue);
 492         ptr_ring_cleanup(rcpu->queue, NULL);
 493         kfree(rcpu->queue);
 494         free_percpu(rcpu->bulkq);
 495         kfree(rcpu);
 496 }
 497
 498 /* After the xchg of the bpf_cpu_map_entry pointer, we need to make sure the old
 499  * entry is no longer in use before freeing. We use queue_rcu_work() to call
 500  * __cpu_map_entry_free() in a separate workqueue after waiting for an RCU grace
 501  * period. This means that (a) all pending enqueue and flush operations have
 502  * completed (because of the RCU callback), and (b) we are in a workqueue
 503  * context where we can stop the kthread and wait for it to exit before freeing
 504  * everything.
 505  */
 506 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
 507                                     u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
 508 {
 509         struct bpf_cpu_map_entry *old_rcpu;
 510
 511         old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
 512         if (old_rcpu) {
 513                 INIT_RCU_WORK(&old_rcpu->free_work, __cpu_map_entry_free);
 514                 queue_rcu_work(system_wq, &old_rcpu->free_work);
 515         }
 516 }
 517
 518 static long cpu_map_delete_elem(struct bpf_map *map, void *key)
 519 {
 520         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 521         u32 key_cpu = *(u32 *)key;
 522
 523         if (key_cpu >= map->max_entries)
 524                 return -EINVAL;
 525
 526         /* notice caller map_delete_elem() uses rcu_read_lock() */
 527         __cpu_map_entry_replace(cmap, key_cpu, NULL);
 528         return 0;
 529 }
 530
 531 static long cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
 532                                 u64 map_flags)
 533 {
 534         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 535         struct bpf_cpumap_val cpumap_value = {};
 536         struct bpf_cpu_map_entry *rcpu;
 537         /* Array index key correspond to CPU number */
 538         u32 key_cpu = *(u32 *)key;
 539
 540         memcpy(&cpumap_value, value, map->value_size);
 541
 542         if (unlikely(map_flags > BPF_EXIST))
 543                 return -EINVAL;
 544         if (unlikely(key_cpu >= cmap->map.max_entries))
 545                 return -E2BIG;
 546         if (unlikely(map_flags == BPF_NOEXIST))
 547                 return -EEXIST;
 548         if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
 549                 return -EOVERFLOW;
 550
 551         /* Make sure CPU is a valid possible cpu */
 552         if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
 553                 return -ENODEV;
 554
 555         if (cpumap_value.qsize == 0) {
 556                 rcpu = NULL; /* Same as deleting */
 557         } else {
 558                 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
 559                 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
 560                 if (!rcpu)
 561                         return -ENOMEM;
 562         }
 563         rcu_read_lock();
 564         __cpu_map_entry_replace(cmap, key_cpu, rcpu);
 565         rcu_read_unlock();
 566         return 0;
 567 }
 568
 569 static void cpu_map_free(struct bpf_map *map)
 570 {
 571         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 572         u32 i;
 573
 574         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 575          * so the bpf programs (can be more than one that used this map) were
 576          * disconnected from events. Wait for outstanding critical sections in
 577          * these programs to complete. synchronize_rcu() below not only
 578          * guarantees no further "XDP/bpf-side" reads against
 579          * bpf_cpu_map->cpu_map, but also ensure pending flush operations
 580          * (if any) are completed.
 581          */
 582         synchronize_rcu();
 583
 584         /* The only possible user of bpf_cpu_map_entry is
 585          * cpu_map_kthread_run().
 586          */
 587         for (i = 0; i < cmap->map.max_entries; i++) {
 588                 struct bpf_cpu_map_entry *rcpu;
 589
 590                 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
 591                 if (!rcpu)
 592                         continue;
 593
 594                 /* Stop kthread and cleanup entry directly */
 595                 __cpu_map_entry_free(&rcpu->free_work.work);
 596         }
 597         bpf_map_area_free(cmap->cpu_map);
 598         bpf_map_area_free(cmap);
 599 }
 600
 601 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 602  * by local_bh_disable() (from XDP calls inside NAPI). The
 603  * rcu_read_lock_bh_held() below makes lockdep accept both.
 604  */
 605 static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
 606 {
 607         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 608         struct bpf_cpu_map_entry *rcpu;
 609
 610         if (key >= map->max_entries)
 611                 return NULL;
 612
 613         rcpu = rcu_dereference_check(cmap->cpu_map[key],
 614                                      rcu_read_lock_bh_held());
 615         return rcpu;
 616 }
 617
 618 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
 619 {
 620         struct bpf_cpu_map_entry *rcpu =
 621                 __cpu_map_lookup_elem(map, *(u32 *)key);
 622
 623         return rcpu ? &rcpu->value : NULL;
 624 }
 625
 626 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 627 {
 628         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 629         u32 index = key ? *(u32 *)key : U32_MAX;
 630         u32 *next = next_key;
 631
 632         if (index >= cmap->map.max_entries) {
 633                 *next = 0;
 634                 return 0;
 635         }
 636
 637         if (index == cmap->map.max_entries - 1)
 638                 return -ENOENT;
 639         *next = index + 1;
 640         return 0;
 641 }
 642
 643 static long cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
 644 {
 645         return __bpf_xdp_redirect_map(map, index, flags, 0,
 646                                       __cpu_map_lookup_elem);
 647 }
 648
 649 static u64 cpu_map_mem_usage(const struct bpf_map *map)
 650 {
 651         u64 usage = sizeof(struct bpf_cpu_map);
 652
 653         /* Currently the dynamically allocated elements are not counted */
 654         usage += (u64)map->max_entries * sizeof(struct bpf_cpu_map_entry *);
 655         return usage;
 656 }
 657
 658 BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
 659 const struct bpf_map_ops cpu_map_ops = {
 660         .map_meta_equal         = bpf_map_meta_equal,
 661         .map_alloc              = cpu_map_alloc,
 662         .map_free               = cpu_map_free,
 663         .map_delete_elem        = cpu_map_delete_elem,
 664         .map_update_elem        = cpu_map_update_elem,
 665         .map_lookup_elem        = cpu_map_lookup_elem,
 666         .map_get_next_key       = cpu_map_get_next_key,
 667         .map_check_btf          = map_check_no_btf,
 668         .map_mem_usage          = cpu_map_mem_usage,
 669         .map_btf_id             = &cpu_map_btf_ids[0],
 670         .map_redirect           = cpu_map_redirect,
 671 };
 672
 673 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
 674 {
 675         struct bpf_cpu_map_entry *rcpu = bq->obj;
 676         unsigned int processed = 0, drops = 0;
 677         const int to_cpu = rcpu->cpu;
 678         struct ptr_ring *q;
 679         int i;
 680
 681         if (unlikely(!bq->count))
 682                 return;
 683
 684         q = rcpu->queue;
 685         spin_lock(&q->producer_lock);
 686
 687         for (i = 0; i < bq->count; i++) {
 688                 struct xdp_frame *xdpf = bq->q[i];
 689                 int err;
 690
 691                 err = __ptr_ring_produce(q, xdpf);
 692                 if (err) {
 693                         drops++;
 694                         xdp_return_frame_rx_napi(xdpf);
 695                 }
 696                 processed++;
 697         }
 698         bq->count = 0;
 699         spin_unlock(&q->producer_lock);
 700
 701         __list_del_clearprev(&bq->flush_node);
 702
 703         /* Feedback loop via tracepoints */
 704         trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
 705 }
 706
 707 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
 708  * Thus, safe percpu variable access.
 709  */
 710 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
 711 {
 712         struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
 713
 714         if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
 715                 bq_flush_to_queue(bq);
 716
 717         /* Notice, xdp_buff/page MUST be queued here, long enough for
 718          * driver to code invoking us to finished, due to driver
 719          * (e.g. ixgbe) recycle tricks based on page-refcnt.
 720          *
 721          * Thus, incoming xdp_frame is always queued here (else we race
 722          * with another CPU on page-refcnt and remaining driver code).
 723          * Queue time is very short, as driver will invoke flush
 724          * operation, when completing napi->poll call.
 725          */
 726         bq->q[bq->count++] = xdpf;
 727
 728         if (!bq->flush_node.prev) {
 729                 struct list_head *flush_list = bpf_net_ctx_get_cpu_map_flush_list();
 730
 731                 list_add(&bq->flush_node, flush_list);
 732         }
 733 }
 734
 735 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
 736                     struct net_device *dev_rx)
 737 {
 738         /* Info needed when constructing SKB on remote CPU */
 739         xdpf->dev_rx = dev_rx;
 740
 741         bq_enqueue(rcpu, xdpf);
 742         return 0;
 743 }
 744
 745 int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
 746                              struct sk_buff *skb)
 747 {
 748         int ret;
 749
 750         __skb_pull(skb, skb->mac_len);
 751         skb_set_redirected(skb, false);
 752         __ptr_set_bit(0, &skb);
 753
 754         ret = ptr_ring_produce(rcpu->queue, skb);
 755         if (ret < 0)
 756                 goto trace;
 757
 758         wake_up_process(rcpu->kthread);
 759 trace:
 760         trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
 761         return ret;
 762 }
 763
 764 void __cpu_map_flush(struct list_head *flush_list)
 765 {
 766         struct xdp_bulk_queue *bq, *tmp;
 767
 768         list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 769                 bq_flush_to_queue(bq);
 770
 771                 /* If already running, costs spin_lock_irqsave + smb_mb */
 772                 wake_up_process(bq->obj->kthread);
 773         }
 774 }