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
  28 #include <linux/sched.h>
  29 #include <linux/workqueue.h>
  30 #include <linux/kthread.h>
  31 #include <trace/events/xdp.h>
  32 #include <linux/btf_ids.h>
  33
  34 #include <linux/netdevice.h>   /* netif_receive_skb_list */
  35 #include <linux/etherdevice.h> /* eth_type_trans */
  36
  37 /* General idea: XDP packets getting XDP redirected to another CPU,
  38  * will maximum be stored/queued for one driver ->poll() call.  It is
  39  * guaranteed that queueing the frame and the flush operation happen on
  40  * same CPU.  Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
  41  * which queue in bpf_cpu_map_entry contains packets.
  42  */
  43
  44 #define CPU_MAP_BULK_SIZE 8  /* 8 == one cacheline on 64-bit archs */
  45 struct bpf_cpu_map_entry;
  46 struct bpf_cpu_map;
  47
  48 struct xdp_bulk_queue {
  49         void *q[CPU_MAP_BULK_SIZE];
  50         struct list_head flush_node;
  51         struct bpf_cpu_map_entry *obj;
  52         unsigned int count;
  53 };
  54
  55 /* Struct for every remote "destination" CPU in map */
  56 struct bpf_cpu_map_entry {
  57         u32 cpu;    /* kthread CPU and map index */
  58         int map_id; /* Back reference to map */
  59
  60         /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
  61         struct xdp_bulk_queue __percpu *bulkq;
  62
  63         struct bpf_cpu_map *cmap;
  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         atomic_t refcnt; /* Control when this struct can be free'ed */
  73         struct rcu_head rcu;
  74
  75         struct work_struct kthread_stop_wq;
  76 };
  77
  78 struct bpf_cpu_map {
  79         struct bpf_map map;
  80         /* Below members specific for map type */
  81         struct bpf_cpu_map_entry __rcu **cpu_map;
  82 };
  83
  84 static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
  85
  86 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
  87 {
  88         u32 value_size = attr->value_size;
  89         struct bpf_cpu_map *cmap;
  90
  91         /* check sanity of attributes */
  92         if (attr->max_entries == 0 || attr->key_size != 4 ||
  93             (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
  94              value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
  95             attr->map_flags & ~BPF_F_NUMA_NODE)
  96                 return ERR_PTR(-EINVAL);
  97
  98         /* Pre-limit array size based on NR_CPUS, not final CPU check */
  99         if (attr->max_entries > NR_CPUS)
 100                 return ERR_PTR(-E2BIG);
 101
 102         cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
 103         if (!cmap)
 104                 return ERR_PTR(-ENOMEM);
 105
 106         bpf_map_init_from_attr(&cmap->map, attr);
 107
 108         /* Alloc array for possible remote "destination" CPUs */
 109         cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
 110                                            sizeof(struct bpf_cpu_map_entry *),
 111                                            cmap->map.numa_node);
 112         if (!cmap->cpu_map) {
 113                 bpf_map_area_free(cmap);
 114                 return ERR_PTR(-ENOMEM);
 115         }
 116
 117         return &cmap->map;
 118 }
 119
 120 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
 121 {
 122         atomic_inc(&rcpu->refcnt);
 123 }
 124
 125 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
 126 {
 127         /* The tear-down procedure should have made sure that queue is
 128          * empty.  See __cpu_map_entry_replace() and work-queue
 129          * invoked cpu_map_kthread_stop(). Catch any broken behaviour
 130          * gracefully and warn once.
 131          */
 132         struct xdp_frame *xdpf;
 133
 134         while ((xdpf = ptr_ring_consume(ring)))
 135                 if (WARN_ON_ONCE(xdpf))
 136                         xdp_return_frame(xdpf);
 137 }
 138
 139 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
 140 {
 141         if (atomic_dec_and_test(&rcpu->refcnt)) {
 142                 if (rcpu->prog)
 143                         bpf_prog_put(rcpu->prog);
 144                 /* The queue should be empty at this point */
 145                 __cpu_map_ring_cleanup(rcpu->queue);
 146                 ptr_ring_cleanup(rcpu->queue, NULL);
 147                 kfree(rcpu->queue);
 148                 kfree(rcpu);
 149         }
 150 }
 151
 152 /* called from workqueue, to workaround syscall using preempt_disable */
 153 static void cpu_map_kthread_stop(struct work_struct *work)
 154 {
 155         struct bpf_cpu_map_entry *rcpu;
 156         int err;
 157
 158         rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
 159
 160         /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
 161          * as it waits until all in-flight call_rcu() callbacks complete.
 162          */
 163         rcu_barrier();
 164
 165         /* kthread_stop will wake_up_process and wait for it to complete */
 166         err = kthread_stop(rcpu->kthread);
 167         if (err) {
 168                 /* kthread_stop may be called before cpu_map_kthread_run
 169                  * is executed, so we need to release the memory related
 170                  * to rcpu.
 171                  */
 172                 put_cpu_map_entry(rcpu);
 173         }
 174 }
 175
 176 static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
 177                                      struct list_head *listp,
 178                                      struct xdp_cpumap_stats *stats)
 179 {
 180         struct sk_buff *skb, *tmp;
 181         struct xdp_buff xdp;
 182         u32 act;
 183         int err;
 184
 185         list_for_each_entry_safe(skb, tmp, listp, list) {
 186                 act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
 187                 switch (act) {
 188                 case XDP_PASS:
 189                         break;
 190                 case XDP_REDIRECT:
 191                         skb_list_del_init(skb);
 192                         err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
 193                                                       rcpu->prog);
 194                         if (unlikely(err)) {
 195                                 kfree_skb(skb);
 196                                 stats->drop++;
 197                         } else {
 198                                 stats->redirect++;
 199                         }
 200                         return;
 201                 default:
 202                         bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
 203                         fallthrough;
 204                 case XDP_ABORTED:
 205                         trace_xdp_exception(skb->dev, rcpu->prog, act);
 206                         fallthrough;
 207                 case XDP_DROP:
 208                         skb_list_del_init(skb);
 209                         kfree_skb(skb);
 210                         stats->drop++;
 211                         return;
 212                 }
 213         }
 214 }
 215
 216 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 217                                     void **frames, int n,
 218                                     struct xdp_cpumap_stats *stats)
 219 {
 220         struct xdp_rxq_info rxq;
 221         struct xdp_buff xdp;
 222         int i, nframes = 0;
 223
 224         xdp_set_return_frame_no_direct();
 225         xdp.rxq = &rxq;
 226
 227         for (i = 0; i < n; i++) {
 228                 struct xdp_frame *xdpf = frames[i];
 229                 u32 act;
 230                 int err;
 231
 232                 rxq.dev = xdpf->dev_rx;
 233                 rxq.mem = xdpf->mem;
 234                 /* TODO: report queue_index to xdp_rxq_info */
 235
 236                 xdp_convert_frame_to_buff(xdpf, &xdp);
 237
 238                 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
 239                 switch (act) {
 240                 case XDP_PASS:
 241                         err = xdp_update_frame_from_buff(&xdp, xdpf);
 242                         if (err < 0) {
 243                                 xdp_return_frame(xdpf);
 244                                 stats->drop++;
 245                         } else {
 246                                 frames[nframes++] = xdpf;
 247                                 stats->pass++;
 248                         }
 249                         break;
 250                 case XDP_REDIRECT:
 251                         err = xdp_do_redirect(xdpf->dev_rx, &xdp,
 252                                               rcpu->prog);
 253                         if (unlikely(err)) {
 254                                 xdp_return_frame(xdpf);
 255                                 stats->drop++;
 256                         } else {
 257                                 stats->redirect++;
 258                         }
 259                         break;
 260                 default:
 261                         bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
 262                         fallthrough;
 263                 case XDP_DROP:
 264                         xdp_return_frame(xdpf);
 265                         stats->drop++;
 266                         break;
 267                 }
 268         }
 269
 270         xdp_clear_return_frame_no_direct();
 271
 272         return nframes;
 273 }
 274
 275 #define CPUMAP_BATCH 8
 276
 277 static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
 278                                 int xdp_n, struct xdp_cpumap_stats *stats,
 279                                 struct list_head *list)
 280 {
 281         int nframes;
 282
 283         if (!rcpu->prog)
 284                 return xdp_n;
 285
 286         rcu_read_lock_bh();
 287
 288         nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
 289
 290         if (stats->redirect)
 291                 xdp_do_flush();
 292
 293         if (unlikely(!list_empty(list)))
 294                 cpu_map_bpf_prog_run_skb(rcpu, list, stats);
 295
 296         rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
 297
 298         return nframes;
 299 }
 300
 301
 302 static int cpu_map_kthread_run(void *data)
 303 {
 304         struct bpf_cpu_map_entry *rcpu = data;
 305
 306         set_current_state(TASK_INTERRUPTIBLE);
 307
 308         /* When kthread gives stop order, then rcpu have been disconnected
 309          * from map, thus no new packets can enter. Remaining in-flight
 310          * per CPU stored packets are flushed to this queue.  Wait honoring
 311          * kthread_stop signal until queue is empty.
 312          */
 313         while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
 314                 struct xdp_cpumap_stats stats = {}; /* zero stats */
 315                 unsigned int kmem_alloc_drops = 0, sched = 0;
 316                 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
 317                 int i, n, m, nframes, xdp_n;
 318                 void *frames[CPUMAP_BATCH];
 319                 void *skbs[CPUMAP_BATCH];
 320                 LIST_HEAD(list);
 321
 322                 /* Release CPU reschedule checks */
 323                 if (__ptr_ring_empty(rcpu->queue)) {
 324                         set_current_state(TASK_INTERRUPTIBLE);
 325                         /* Recheck to avoid lost wake-up */
 326                         if (__ptr_ring_empty(rcpu->queue)) {
 327                                 schedule();
 328                                 sched = 1;
 329                         } else {
 330                                 __set_current_state(TASK_RUNNING);
 331                         }
 332                 } else {
 333                         sched = cond_resched();
 334                 }
 335
 336                 /*
 337                  * The bpf_cpu_map_entry is single consumer, with this
 338                  * kthread CPU pinned. Lockless access to ptr_ring
 339                  * consume side valid as no-resize allowed of queue.
 340                  */
 341                 n = __ptr_ring_consume_batched(rcpu->queue, frames,
 342                                                CPUMAP_BATCH);
 343                 for (i = 0, xdp_n = 0; i < n; i++) {
 344                         void *f = frames[i];
 345                         struct page *page;
 346
 347                         if (unlikely(__ptr_test_bit(0, &f))) {
 348                                 struct sk_buff *skb = f;
 349
 350                                 __ptr_clear_bit(0, &skb);
 351                                 list_add_tail(&skb->list, &list);
 352                                 continue;
 353                         }
 354
 355                         frames[xdp_n++] = f;
 356                         page = virt_to_page(f);
 357
 358                         /* Bring struct page memory area to curr CPU. Read by
 359                          * build_skb_around via page_is_pfmemalloc(), and when
 360                          * freed written by page_frag_free call.
 361                          */
 362                         prefetchw(page);
 363                 }
 364
 365                 /* Support running another XDP prog on this CPU */
 366                 nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
 367                 if (nframes) {
 368                         m = kmem_cache_alloc_bulk(skbuff_cache, gfp, nframes, skbs);
 369                         if (unlikely(m == 0)) {
 370                                 for (i = 0; i < nframes; i++)
 371                                         skbs[i] = NULL; /* effect: xdp_return_frame */
 372                                 kmem_alloc_drops += nframes;
 373                         }
 374                 }
 375
 376                 local_bh_disable();
 377                 for (i = 0; i < nframes; i++) {
 378                         struct xdp_frame *xdpf = frames[i];
 379                         struct sk_buff *skb = skbs[i];
 380
 381                         skb = __xdp_build_skb_from_frame(xdpf, skb,
 382                                                          xdpf->dev_rx);
 383                         if (!skb) {
 384                                 xdp_return_frame(xdpf);
 385                                 continue;
 386                         }
 387
 388                         list_add_tail(&skb->list, &list);
 389                 }
 390                 netif_receive_skb_list(&list);
 391
 392                 /* Feedback loop via tracepoint */
 393                 trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
 394                                          sched, &stats);
 395
 396                 local_bh_enable(); /* resched point, may call do_softirq() */
 397         }
 398         __set_current_state(TASK_RUNNING);
 399
 400         put_cpu_map_entry(rcpu);
 401         return 0;
 402 }
 403
 404 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
 405                                       struct bpf_map *map, int fd)
 406 {
 407         struct bpf_prog *prog;
 408
 409         prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
 410         if (IS_ERR(prog))
 411                 return PTR_ERR(prog);
 412
 413         if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
 414             !bpf_prog_map_compatible(map, prog)) {
 415                 bpf_prog_put(prog);
 416                 return -EINVAL;
 417         }
 418
 419         rcpu->value.bpf_prog.id = prog->aux->id;
 420         rcpu->prog = prog;
 421
 422         return 0;
 423 }
 424
 425 static struct bpf_cpu_map_entry *
 426 __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
 427                       u32 cpu)
 428 {
 429         int numa, err, i, fd = value->bpf_prog.fd;
 430         gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
 431         struct bpf_cpu_map_entry *rcpu;
 432         struct xdp_bulk_queue *bq;
 433
 434         /* Have map->numa_node, but choose node of redirect target CPU */
 435         numa = cpu_to_node(cpu);
 436
 437         rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
 438         if (!rcpu)
 439                 return NULL;
 440
 441         /* Alloc percpu bulkq */
 442         rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
 443                                            sizeof(void *), gfp);
 444         if (!rcpu->bulkq)
 445                 goto free_rcu;
 446
 447         for_each_possible_cpu(i) {
 448                 bq = per_cpu_ptr(rcpu->bulkq, i);
 449                 bq->obj = rcpu;
 450         }
 451
 452         /* Alloc queue */
 453         rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
 454                                            numa);
 455         if (!rcpu->queue)
 456                 goto free_bulkq;
 457
 458         err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
 459         if (err)
 460                 goto free_queue;
 461
 462         rcpu->cpu    = cpu;
 463         rcpu->map_id = map->id;
 464         rcpu->value.qsize  = value->qsize;
 465
 466         if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
 467                 goto free_ptr_ring;
 468
 469         /* Setup kthread */
 470         rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
 471                                                "cpumap/%d/map:%d", cpu,
 472                                                map->id);
 473         if (IS_ERR(rcpu->kthread))
 474                 goto free_prog;
 475
 476         get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
 477         get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
 478
 479         /* Make sure kthread runs on a single CPU */
 480         kthread_bind(rcpu->kthread, cpu);
 481         wake_up_process(rcpu->kthread);
 482
 483         return rcpu;
 484
 485 free_prog:
 486         if (rcpu->prog)
 487                 bpf_prog_put(rcpu->prog);
 488 free_ptr_ring:
 489         ptr_ring_cleanup(rcpu->queue, NULL);
 490 free_queue:
 491         kfree(rcpu->queue);
 492 free_bulkq:
 493         free_percpu(rcpu->bulkq);
 494 free_rcu:
 495         kfree(rcpu);
 496         return NULL;
 497 }
 498
 499 static void __cpu_map_entry_free(struct rcu_head *rcu)
 500 {
 501         struct bpf_cpu_map_entry *rcpu;
 502
 503         /* This cpu_map_entry have been disconnected from map and one
 504          * RCU grace-period have elapsed.  Thus, XDP cannot queue any
 505          * new packets and cannot change/set flush_needed that can
 506          * find this entry.
 507          */
 508         rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
 509
 510         free_percpu(rcpu->bulkq);
 511         /* Cannot kthread_stop() here, last put free rcpu resources */
 512         put_cpu_map_entry(rcpu);
 513 }
 514
 515 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
 516  * ensure any driver rcu critical sections have completed, but this
 517  * does not guarantee a flush has happened yet. Because driver side
 518  * rcu_read_lock/unlock only protects the running XDP program.  The
 519  * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
 520  * pending flush op doesn't fail.
 521  *
 522  * The bpf_cpu_map_entry is still used by the kthread, and there can
 523  * still be pending packets (in queue and percpu bulkq).  A refcnt
 524  * makes sure to last user (kthread_stop vs. call_rcu) free memory
 525  * resources.
 526  *
 527  * The rcu callback __cpu_map_entry_free flush remaining packets in
 528  * percpu bulkq to queue.  Due to caller map_delete_elem() disable
 529  * preemption, cannot call kthread_stop() to make sure queue is empty.
 530  * Instead a work_queue is started for stopping kthread,
 531  * cpu_map_kthread_stop, which waits for an RCU grace period before
 532  * stopping kthread, emptying the queue.
 533  */
 534 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
 535                                     u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
 536 {
 537         struct bpf_cpu_map_entry *old_rcpu;
 538
 539         old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
 540         if (old_rcpu) {
 541                 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
 542                 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
 543                 schedule_work(&old_rcpu->kthread_stop_wq);
 544         }
 545 }
 546
 547 static long cpu_map_delete_elem(struct bpf_map *map, void *key)
 548 {
 549         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 550         u32 key_cpu = *(u32 *)key;
 551
 552         if (key_cpu >= map->max_entries)
 553                 return -EINVAL;
 554
 555         /* notice caller map_delete_elem() use preempt_disable() */
 556         __cpu_map_entry_replace(cmap, key_cpu, NULL);
 557         return 0;
 558 }
 559
 560 static long cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
 561                                 u64 map_flags)
 562 {
 563         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 564         struct bpf_cpumap_val cpumap_value = {};
 565         struct bpf_cpu_map_entry *rcpu;
 566         /* Array index key correspond to CPU number */
 567         u32 key_cpu = *(u32 *)key;
 568
 569         memcpy(&cpumap_value, value, map->value_size);
 570
 571         if (unlikely(map_flags > BPF_EXIST))
 572                 return -EINVAL;
 573         if (unlikely(key_cpu >= cmap->map.max_entries))
 574                 return -E2BIG;
 575         if (unlikely(map_flags == BPF_NOEXIST))
 576                 return -EEXIST;
 577         if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
 578                 return -EOVERFLOW;
 579
 580         /* Make sure CPU is a valid possible cpu */
 581         if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
 582                 return -ENODEV;
 583
 584         if (cpumap_value.qsize == 0) {
 585                 rcpu = NULL; /* Same as deleting */
 586         } else {
 587                 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
 588                 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
 589                 if (!rcpu)
 590                         return -ENOMEM;
 591                 rcpu->cmap = cmap;
 592         }
 593         rcu_read_lock();
 594         __cpu_map_entry_replace(cmap, key_cpu, rcpu);
 595         rcu_read_unlock();
 596         return 0;
 597 }
 598
 599 static void cpu_map_free(struct bpf_map *map)
 600 {
 601         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 602         u32 i;
 603
 604         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 605          * so the bpf programs (can be more than one that used this map) were
 606          * disconnected from events. Wait for outstanding critical sections in
 607          * these programs to complete. The rcu critical section only guarantees
 608          * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
 609          * It does __not__ ensure pending flush operations (if any) are
 610          * complete.
 611          */
 612
 613         synchronize_rcu();
 614
 615         /* For cpu_map the remote CPUs can still be using the entries
 616          * (struct bpf_cpu_map_entry).
 617          */
 618         for (i = 0; i < cmap->map.max_entries; i++) {
 619                 struct bpf_cpu_map_entry *rcpu;
 620
 621                 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
 622                 if (!rcpu)
 623                         continue;
 624
 625                 /* bq flush and cleanup happens after RCU grace-period */
 626                 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
 627         }
 628         bpf_map_area_free(cmap->cpu_map);
 629         bpf_map_area_free(cmap);
 630 }
 631
 632 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 633  * by local_bh_disable() (from XDP calls inside NAPI). The
 634  * rcu_read_lock_bh_held() below makes lockdep accept both.
 635  */
 636 static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
 637 {
 638         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 639         struct bpf_cpu_map_entry *rcpu;
 640
 641         if (key >= map->max_entries)
 642                 return NULL;
 643
 644         rcpu = rcu_dereference_check(cmap->cpu_map[key],
 645                                      rcu_read_lock_bh_held());
 646         return rcpu;
 647 }
 648
 649 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
 650 {
 651         struct bpf_cpu_map_entry *rcpu =
 652                 __cpu_map_lookup_elem(map, *(u32 *)key);
 653
 654         return rcpu ? &rcpu->value : NULL;
 655 }
 656
 657 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 658 {
 659         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 660         u32 index = key ? *(u32 *)key : U32_MAX;
 661         u32 *next = next_key;
 662
 663         if (index >= cmap->map.max_entries) {
 664                 *next = 0;
 665                 return 0;
 666         }
 667
 668         if (index == cmap->map.max_entries - 1)
 669                 return -ENOENT;
 670         *next = index + 1;
 671         return 0;
 672 }
 673
 674 static long cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
 675 {
 676         return __bpf_xdp_redirect_map(map, index, flags, 0,
 677                                       __cpu_map_lookup_elem);
 678 }
 679
 680 static u64 cpu_map_mem_usage(const struct bpf_map *map)
 681 {
 682         u64 usage = sizeof(struct bpf_cpu_map);
 683
 684         /* Currently the dynamically allocated elements are not counted */
 685         usage += (u64)map->max_entries * sizeof(struct bpf_cpu_map_entry *);
 686         return usage;
 687 }
 688
 689 BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
 690 const struct bpf_map_ops cpu_map_ops = {
 691         .map_meta_equal         = bpf_map_meta_equal,
 692         .map_alloc              = cpu_map_alloc,
 693         .map_free               = cpu_map_free,
 694         .map_delete_elem        = cpu_map_delete_elem,
 695         .map_update_elem        = cpu_map_update_elem,
 696         .map_lookup_elem        = cpu_map_lookup_elem,
 697         .map_get_next_key       = cpu_map_get_next_key,
 698         .map_check_btf          = map_check_no_btf,
 699         .map_mem_usage          = cpu_map_mem_usage,
 700         .map_btf_id             = &cpu_map_btf_ids[0],
 701         .map_redirect           = cpu_map_redirect,
 702 };
 703
 704 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
 705 {
 706         struct bpf_cpu_map_entry *rcpu = bq->obj;
 707         unsigned int processed = 0, drops = 0;
 708         const int to_cpu = rcpu->cpu;
 709         struct ptr_ring *q;
 710         int i;
 711
 712         if (unlikely(!bq->count))
 713                 return;
 714
 715         q = rcpu->queue;
 716         spin_lock(&q->producer_lock);
 717
 718         for (i = 0; i < bq->count; i++) {
 719                 struct xdp_frame *xdpf = bq->q[i];
 720                 int err;
 721
 722                 err = __ptr_ring_produce(q, xdpf);
 723                 if (err) {
 724                         drops++;
 725                         xdp_return_frame_rx_napi(xdpf);
 726                 }
 727                 processed++;
 728         }
 729         bq->count = 0;
 730         spin_unlock(&q->producer_lock);
 731
 732         __list_del_clearprev(&bq->flush_node);
 733
 734         /* Feedback loop via tracepoints */
 735         trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
 736 }
 737
 738 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
 739  * Thus, safe percpu variable access.
 740  */
 741 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
 742 {
 743         struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
 744         struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
 745
 746         if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
 747                 bq_flush_to_queue(bq);
 748
 749         /* Notice, xdp_buff/page MUST be queued here, long enough for
 750          * driver to code invoking us to finished, due to driver
 751          * (e.g. ixgbe) recycle tricks based on page-refcnt.
 752          *
 753          * Thus, incoming xdp_frame is always queued here (else we race
 754          * with another CPU on page-refcnt and remaining driver code).
 755          * Queue time is very short, as driver will invoke flush
 756          * operation, when completing napi->poll call.
 757          */
 758         bq->q[bq->count++] = xdpf;
 759
 760         if (!bq->flush_node.prev)
 761                 list_add(&bq->flush_node, flush_list);
 762 }
 763
 764 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
 765                     struct net_device *dev_rx)
 766 {
 767         /* Info needed when constructing SKB on remote CPU */
 768         xdpf->dev_rx = dev_rx;
 769
 770         bq_enqueue(rcpu, xdpf);
 771         return 0;
 772 }
 773
 774 int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
 775                              struct sk_buff *skb)
 776 {
 777         int ret;
 778
 779         __skb_pull(skb, skb->mac_len);
 780         skb_set_redirected(skb, false);
 781         __ptr_set_bit(0, &skb);
 782
 783         ret = ptr_ring_produce(rcpu->queue, skb);
 784         if (ret < 0)
 785                 goto trace;
 786
 787         wake_up_process(rcpu->kthread);
 788 trace:
 789         trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
 790         return ret;
 791 }
 792
 793 void __cpu_map_flush(void)
 794 {
 795         struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
 796         struct xdp_bulk_queue *bq, *tmp;
 797
 798         list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 799                 bq_flush_to_queue(bq);
 800
 801                 /* If already running, costs spin_lock_irqsave + smb_mb */
 802                 wake_up_process(bq->obj->kthread);
 803         }
 804 }
 805
 806 static int __init cpu_map_init(void)
 807 {
 808         int cpu;
 809
 810         for_each_possible_cpu(cpu)
 811                 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
 812         return 0;
 813 }
 814
 815 subsys_initcall(cpu_map_init);