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 /* The 'cpumap' is primarily used as a backend map for XDP BPF helper
   8  * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
   9  *
  10  * Unlike devmap which redirects XDP frames out another NIC device,
  11  * this map type redirects raw XDP frames to another CPU.  The remote
  12  * CPU will do SKB-allocation and call the normal network stack.
  13  *
  14  * This is a scalability and isolation mechanism, that allow
  15  * separating the early driver network XDP layer, from the rest of the
  16  * netstack, and assigning dedicated CPUs for this stage.  This
  17  * basically allows for 10G wirespeed pre-filtering via bpf.
  18  */
  19 #include <linux/bitops.h>
  20 #include <linux/bpf.h>
  21 #include <linux/filter.h>
  22 #include <linux/ptr_ring.h>
  23 #include <net/xdp.h>
  24
  25 #include <linux/sched.h>
  26 #include <linux/workqueue.h>
  27 #include <linux/kthread.h>
  28 #include <linux/capability.h>
  29 #include <trace/events/xdp.h>
  30 #include <linux/btf_ids.h>
  31
  32 #include <linux/netdevice.h>   /* netif_receive_skb_list */
  33 #include <linux/etherdevice.h> /* eth_type_trans */
  34
  35 /* General idea: XDP packets getting XDP redirected to another CPU,
  36  * will maximum be stored/queued for one driver ->poll() call.  It is
  37  * guaranteed that queueing the frame and the flush operation happen on
  38  * same CPU.  Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
  39  * which queue in bpf_cpu_map_entry contains packets.
  40  */
  41
  42 #define CPU_MAP_BULK_SIZE 8  /* 8 == one cacheline on 64-bit archs */
  43 struct bpf_cpu_map_entry;
  44 struct bpf_cpu_map;
  45
  46 struct xdp_bulk_queue {
  47         void *q[CPU_MAP_BULK_SIZE];
  48         struct list_head flush_node;
  49         struct bpf_cpu_map_entry *obj;
  50         unsigned int count;
  51 };
  52
  53 /* Struct for every remote "destination" CPU in map */
  54 struct bpf_cpu_map_entry {
  55         u32 cpu;    /* kthread CPU and map index */
  56         int map_id; /* Back reference to map */
  57
  58         /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
  59         struct xdp_bulk_queue __percpu *bulkq;
  60
  61         struct bpf_cpu_map *cmap;
  62
  63         /* Queue with potential multi-producers, and single-consumer kthread */
  64         struct ptr_ring *queue;
  65         struct task_struct *kthread;
  66
  67         struct bpf_cpumap_val value;
  68         struct bpf_prog *prog;
  69
  70         atomic_t refcnt; /* Control when this struct can be free'ed */
  71         struct rcu_head rcu;
  72
  73         struct work_struct kthread_stop_wq;
  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 DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
  83
  84 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
  85 {
  86         u32 value_size = attr->value_size;
  87         struct bpf_cpu_map *cmap;
  88         int err = -ENOMEM;
  89
  90         if (!bpf_capable())
  91                 return ERR_PTR(-EPERM);
  92
  93         /* check sanity of attributes */
  94         if (attr->max_entries == 0 || attr->key_size != 4 ||
  95             (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
  96              value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
  97             attr->map_flags & ~BPF_F_NUMA_NODE)
  98                 return ERR_PTR(-EINVAL);
  99
 100         cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
 101         if (!cmap)
 102                 return ERR_PTR(-ENOMEM);
 103
 104         bpf_map_init_from_attr(&cmap->map, attr);
 105
 106         /* Pre-limit array size based on NR_CPUS, not final CPU check */
 107         if (cmap->map.max_entries > NR_CPUS) {
 108                 err = -E2BIG;
 109                 goto free_cmap;
 110         }
 111
 112         /* Alloc array for possible remote "destination" CPUs */
 113         cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
 114                                            sizeof(struct bpf_cpu_map_entry *),
 115                                            cmap->map.numa_node);
 116         if (!cmap->cpu_map)
 117                 goto free_cmap;
 118
 119         return &cmap->map;
 120 free_cmap:
 121         bpf_map_area_free(cmap);
 122         return ERR_PTR(err);
 123 }
 124
 125 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
 126 {
 127         atomic_inc(&rcpu->refcnt);
 128 }
 129
 130 /* called from workqueue, to workaround syscall using preempt_disable */
 131 static void cpu_map_kthread_stop(struct work_struct *work)
 132 {
 133         struct bpf_cpu_map_entry *rcpu;
 134
 135         rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
 136
 137         /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
 138          * as it waits until all in-flight call_rcu() callbacks complete.
 139          */
 140         rcu_barrier();
 141
 142         /* kthread_stop will wake_up_process and wait for it to complete */
 143         kthread_stop(rcpu->kthread);
 144 }
 145
 146 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
 147 {
 148         /* The tear-down procedure should have made sure that queue is
 149          * empty.  See __cpu_map_entry_replace() and work-queue
 150          * invoked cpu_map_kthread_stop(). Catch any broken behaviour
 151          * gracefully and warn once.
 152          */
 153         struct xdp_frame *xdpf;
 154
 155         while ((xdpf = ptr_ring_consume(ring)))
 156                 if (WARN_ON_ONCE(xdpf))
 157                         xdp_return_frame(xdpf);
 158 }
 159
 160 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
 161 {
 162         if (atomic_dec_and_test(&rcpu->refcnt)) {
 163                 if (rcpu->prog)
 164                         bpf_prog_put(rcpu->prog);
 165                 /* The queue should be empty at this point */
 166                 __cpu_map_ring_cleanup(rcpu->queue);
 167                 ptr_ring_cleanup(rcpu->queue, NULL);
 168                 kfree(rcpu->queue);
 169                 kfree(rcpu);
 170         }
 171 }
 172
 173 static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
 174                                      struct list_head *listp,
 175                                      struct xdp_cpumap_stats *stats)
 176 {
 177         struct sk_buff *skb, *tmp;
 178         struct xdp_buff xdp;
 179         u32 act;
 180         int err;
 181
 182         list_for_each_entry_safe(skb, tmp, listp, list) {
 183                 act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
 184                 switch (act) {
 185                 case XDP_PASS:
 186                         break;
 187                 case XDP_REDIRECT:
 188                         skb_list_del_init(skb);
 189                         err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
 190                                                       rcpu->prog);
 191                         if (unlikely(err)) {
 192                                 kfree_skb(skb);
 193                                 stats->drop++;
 194                         } else {
 195                                 stats->redirect++;
 196                         }
 197                         return;
 198                 default:
 199                         bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
 200                         fallthrough;
 201                 case XDP_ABORTED:
 202                         trace_xdp_exception(skb->dev, rcpu->prog, act);
 203                         fallthrough;
 204                 case XDP_DROP:
 205                         skb_list_del_init(skb);
 206                         kfree_skb(skb);
 207                         stats->drop++;
 208                         return;
 209                 }
 210         }
 211 }
 212
 213 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 214                                     void **frames, int n,
 215                                     struct xdp_cpumap_stats *stats)
 216 {
 217         struct xdp_rxq_info rxq;
 218         struct xdp_buff xdp;
 219         int i, nframes = 0;
 220
 221         xdp_set_return_frame_no_direct();
 222         xdp.rxq = &rxq;
 223
 224         for (i = 0; i < n; i++) {
 225                 struct xdp_frame *xdpf = frames[i];
 226                 u32 act;
 227                 int err;
 228
 229                 rxq.dev = xdpf->dev_rx;
 230                 rxq.mem = xdpf->mem;
 231                 /* TODO: report queue_index to xdp_rxq_info */
 232
 233                 xdp_convert_frame_to_buff(xdpf, &xdp);
 234
 235                 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
 236                 switch (act) {
 237                 case XDP_PASS:
 238                         err = xdp_update_frame_from_buff(&xdp, xdpf);
 239                         if (err < 0) {
 240                                 xdp_return_frame(xdpf);
 241                                 stats->drop++;
 242                         } else {
 243                                 frames[nframes++] = xdpf;
 244                                 stats->pass++;
 245                         }
 246                         break;
 247                 case XDP_REDIRECT:
 248                         err = xdp_do_redirect(xdpf->dev_rx, &xdp,
 249                                               rcpu->prog);
 250                         if (unlikely(err)) {
 251                                 xdp_return_frame(xdpf);
 252                                 stats->drop++;
 253                         } else {
 254                                 stats->redirect++;
 255                         }
 256                         break;
 257                 default:
 258                         bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
 259                         fallthrough;
 260                 case XDP_DROP:
 261                         xdp_return_frame(xdpf);
 262                         stats->drop++;
 263                         break;
 264                 }
 265         }
 266
 267         xdp_clear_return_frame_no_direct();
 268
 269         return nframes;
 270 }
 271
 272 #define CPUMAP_BATCH 8
 273
 274 static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
 275                                 int xdp_n, struct xdp_cpumap_stats *stats,
 276                                 struct list_head *list)
 277 {
 278         int nframes;
 279
 280         if (!rcpu->prog)
 281                 return xdp_n;
 282
 283         rcu_read_lock_bh();
 284
 285         nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
 286
 287         if (stats->redirect)
 288                 xdp_do_flush();
 289
 290         if (unlikely(!list_empty(list)))
 291                 cpu_map_bpf_prog_run_skb(rcpu, list, stats);
 292
 293         rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
 294
 295         return nframes;
 296 }
 297
 298
 299 static int cpu_map_kthread_run(void *data)
 300 {
 301         struct bpf_cpu_map_entry *rcpu = data;
 302
 303         set_current_state(TASK_INTERRUPTIBLE);
 304
 305         /* When kthread gives stop order, then rcpu have been disconnected
 306          * from map, thus no new packets can enter. Remaining in-flight
 307          * per CPU stored packets are flushed to this queue.  Wait honoring
 308          * kthread_stop signal until queue is empty.
 309          */
 310         while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
 311                 struct xdp_cpumap_stats stats = {}; /* zero stats */
 312                 unsigned int kmem_alloc_drops = 0, sched = 0;
 313                 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
 314                 int i, n, m, nframes, xdp_n;
 315                 void *frames[CPUMAP_BATCH];
 316                 void *skbs[CPUMAP_BATCH];
 317                 LIST_HEAD(list);
 318
 319                 /* Release CPU reschedule checks */
 320                 if (__ptr_ring_empty(rcpu->queue)) {
 321                         set_current_state(TASK_INTERRUPTIBLE);
 322                         /* Recheck to avoid lost wake-up */
 323                         if (__ptr_ring_empty(rcpu->queue)) {
 324                                 schedule();
 325                                 sched = 1;
 326                         } else {
 327                                 __set_current_state(TASK_RUNNING);
 328                         }
 329                 } else {
 330                         sched = cond_resched();
 331                 }
 332
 333                 /*
 334                  * The bpf_cpu_map_entry is single consumer, with this
 335                  * kthread CPU pinned. Lockless access to ptr_ring
 336                  * consume side valid as no-resize allowed of queue.
 337                  */
 338                 n = __ptr_ring_consume_batched(rcpu->queue, frames,
 339                                                CPUMAP_BATCH);
 340                 for (i = 0, xdp_n = 0; i < n; i++) {
 341                         void *f = frames[i];
 342                         struct page *page;
 343
 344                         if (unlikely(__ptr_test_bit(0, &f))) {
 345                                 struct sk_buff *skb = f;
 346
 347                                 __ptr_clear_bit(0, &skb);
 348                                 list_add_tail(&skb->list, &list);
 349                                 continue;
 350                         }
 351
 352                         frames[xdp_n++] = f;
 353                         page = virt_to_page(f);
 354
 355                         /* Bring struct page memory area to curr CPU. Read by
 356                          * build_skb_around via page_is_pfmemalloc(), and when
 357                          * freed written by page_frag_free call.
 358                          */
 359                         prefetchw(page);
 360                 }
 361
 362                 /* Support running another XDP prog on this CPU */
 363                 nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
 364                 if (nframes) {
 365                         m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
 366                         if (unlikely(m == 0)) {
 367                                 for (i = 0; i < nframes; i++)
 368                                         skbs[i] = NULL; /* effect: xdp_return_frame */
 369                                 kmem_alloc_drops += nframes;
 370                         }
 371                 }
 372
 373                 local_bh_disable();
 374                 for (i = 0; i < nframes; i++) {
 375                         struct xdp_frame *xdpf = frames[i];
 376                         struct sk_buff *skb = skbs[i];
 377
 378                         skb = __xdp_build_skb_from_frame(xdpf, skb,
 379                                                          xdpf->dev_rx);
 380                         if (!skb) {
 381                                 xdp_return_frame(xdpf);
 382                                 continue;
 383                         }
 384
 385                         list_add_tail(&skb->list, &list);
 386                 }
 387                 netif_receive_skb_list(&list);
 388
 389                 /* Feedback loop via tracepoint */
 390                 trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
 391                                          sched, &stats);
 392
 393                 local_bh_enable(); /* resched point, may call do_softirq() */
 394         }
 395         __set_current_state(TASK_RUNNING);
 396
 397         put_cpu_map_entry(rcpu);
 398         return 0;
 399 }
 400
 401 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
 402                                       struct bpf_map *map, int fd)
 403 {
 404         struct bpf_prog *prog;
 405
 406         prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
 407         if (IS_ERR(prog))
 408                 return PTR_ERR(prog);
 409
 410         if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
 411             !bpf_prog_map_compatible(map, prog)) {
 412                 bpf_prog_put(prog);
 413                 return -EINVAL;
 414         }
 415
 416         rcpu->value.bpf_prog.id = prog->aux->id;
 417         rcpu->prog = prog;
 418
 419         return 0;
 420 }
 421
 422 static struct bpf_cpu_map_entry *
 423 __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
 424                       u32 cpu)
 425 {
 426         int numa, err, i, fd = value->bpf_prog.fd;
 427         gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
 428         struct bpf_cpu_map_entry *rcpu;
 429         struct xdp_bulk_queue *bq;
 430
 431         /* Have map->numa_node, but choose node of redirect target CPU */
 432         numa = cpu_to_node(cpu);
 433
 434         rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
 435         if (!rcpu)
 436                 return NULL;
 437
 438         /* Alloc percpu bulkq */
 439         rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
 440                                            sizeof(void *), gfp);
 441         if (!rcpu->bulkq)
 442                 goto free_rcu;
 443
 444         for_each_possible_cpu(i) {
 445                 bq = per_cpu_ptr(rcpu->bulkq, i);
 446                 bq->obj = rcpu;
 447         }
 448
 449         /* Alloc queue */
 450         rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
 451                                            numa);
 452         if (!rcpu->queue)
 453                 goto free_bulkq;
 454
 455         err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
 456         if (err)
 457                 goto free_queue;
 458
 459         rcpu->cpu    = cpu;
 460         rcpu->map_id = map->id;
 461         rcpu->value.qsize  = value->qsize;
 462
 463         if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
 464                 goto free_ptr_ring;
 465
 466         /* Setup kthread */
 467         rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
 468                                                "cpumap/%d/map:%d", cpu,
 469                                                map->id);
 470         if (IS_ERR(rcpu->kthread))
 471                 goto free_prog;
 472
 473         get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
 474         get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
 475
 476         /* Make sure kthread runs on a single CPU */
 477         kthread_bind(rcpu->kthread, cpu);
 478         wake_up_process(rcpu->kthread);
 479
 480         return rcpu;
 481
 482 free_prog:
 483         if (rcpu->prog)
 484                 bpf_prog_put(rcpu->prog);
 485 free_ptr_ring:
 486         ptr_ring_cleanup(rcpu->queue, NULL);
 487 free_queue:
 488         kfree(rcpu->queue);
 489 free_bulkq:
 490         free_percpu(rcpu->bulkq);
 491 free_rcu:
 492         kfree(rcpu);
 493         return NULL;
 494 }
 495
 496 static void __cpu_map_entry_free(struct rcu_head *rcu)
 497 {
 498         struct bpf_cpu_map_entry *rcpu;
 499
 500         /* This cpu_map_entry have been disconnected from map and one
 501          * RCU grace-period have elapsed.  Thus, XDP cannot queue any
 502          * new packets and cannot change/set flush_needed that can
 503          * find this entry.
 504          */
 505         rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
 506
 507         free_percpu(rcpu->bulkq);
 508         /* Cannot kthread_stop() here, last put free rcpu resources */
 509         put_cpu_map_entry(rcpu);
 510 }
 511
 512 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
 513  * ensure any driver rcu critical sections have completed, but this
 514  * does not guarantee a flush has happened yet. Because driver side
 515  * rcu_read_lock/unlock only protects the running XDP program.  The
 516  * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
 517  * pending flush op doesn't fail.
 518  *
 519  * The bpf_cpu_map_entry is still used by the kthread, and there can
 520  * still be pending packets (in queue and percpu bulkq).  A refcnt
 521  * makes sure to last user (kthread_stop vs. call_rcu) free memory
 522  * resources.
 523  *
 524  * The rcu callback __cpu_map_entry_free flush remaining packets in
 525  * percpu bulkq to queue.  Due to caller map_delete_elem() disable
 526  * preemption, cannot call kthread_stop() to make sure queue is empty.
 527  * Instead a work_queue is started for stopping kthread,
 528  * cpu_map_kthread_stop, which waits for an RCU grace period before
 529  * stopping kthread, emptying the queue.
 530  */
 531 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
 532                                     u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
 533 {
 534         struct bpf_cpu_map_entry *old_rcpu;
 535
 536         old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
 537         if (old_rcpu) {
 538                 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
 539                 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
 540                 schedule_work(&old_rcpu->kthread_stop_wq);
 541         }
 542 }
 543
 544 static int cpu_map_delete_elem(struct bpf_map *map, void *key)
 545 {
 546         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 547         u32 key_cpu = *(u32 *)key;
 548
 549         if (key_cpu >= map->max_entries)
 550                 return -EINVAL;
 551
 552         /* notice caller map_delete_elem() use preempt_disable() */
 553         __cpu_map_entry_replace(cmap, key_cpu, NULL);
 554         return 0;
 555 }
 556
 557 static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
 558                                u64 map_flags)
 559 {
 560         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 561         struct bpf_cpumap_val cpumap_value = {};
 562         struct bpf_cpu_map_entry *rcpu;
 563         /* Array index key correspond to CPU number */
 564         u32 key_cpu = *(u32 *)key;
 565
 566         memcpy(&cpumap_value, value, map->value_size);
 567
 568         if (unlikely(map_flags > BPF_EXIST))
 569                 return -EINVAL;
 570         if (unlikely(key_cpu >= cmap->map.max_entries))
 571                 return -E2BIG;
 572         if (unlikely(map_flags == BPF_NOEXIST))
 573                 return -EEXIST;
 574         if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
 575                 return -EOVERFLOW;
 576
 577         /* Make sure CPU is a valid possible cpu */
 578         if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
 579                 return -ENODEV;
 580
 581         if (cpumap_value.qsize == 0) {
 582                 rcpu = NULL; /* Same as deleting */
 583         } else {
 584                 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
 585                 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
 586                 if (!rcpu)
 587                         return -ENOMEM;
 588                 rcpu->cmap = cmap;
 589         }
 590         rcu_read_lock();
 591         __cpu_map_entry_replace(cmap, key_cpu, rcpu);
 592         rcu_read_unlock();
 593         return 0;
 594 }
 595
 596 static void cpu_map_free(struct bpf_map *map)
 597 {
 598         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 599         u32 i;
 600
 601         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 602          * so the bpf programs (can be more than one that used this map) were
 603          * disconnected from events. Wait for outstanding critical sections in
 604          * these programs to complete. The rcu critical section only guarantees
 605          * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
 606          * It does __not__ ensure pending flush operations (if any) are
 607          * complete.
 608          */
 609
 610         synchronize_rcu();
 611
 612         /* For cpu_map the remote CPUs can still be using the entries
 613          * (struct bpf_cpu_map_entry).
 614          */
 615         for (i = 0; i < cmap->map.max_entries; i++) {
 616                 struct bpf_cpu_map_entry *rcpu;
 617
 618                 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
 619                 if (!rcpu)
 620                         continue;
 621
 622                 /* bq flush and cleanup happens after RCU grace-period */
 623                 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
 624         }
 625         bpf_map_area_free(cmap->cpu_map);
 626         bpf_map_area_free(cmap);
 627 }
 628
 629 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 630  * by local_bh_disable() (from XDP calls inside NAPI). The
 631  * rcu_read_lock_bh_held() below makes lockdep accept both.
 632  */
 633 static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
 634 {
 635         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 636         struct bpf_cpu_map_entry *rcpu;
 637
 638         if (key >= map->max_entries)
 639                 return NULL;
 640
 641         rcpu = rcu_dereference_check(cmap->cpu_map[key],
 642                                      rcu_read_lock_bh_held());
 643         return rcpu;
 644 }
 645
 646 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
 647 {
 648         struct bpf_cpu_map_entry *rcpu =
 649                 __cpu_map_lookup_elem(map, *(u32 *)key);
 650
 651         return rcpu ? &rcpu->value : NULL;
 652 }
 653
 654 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 655 {
 656         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 657         u32 index = key ? *(u32 *)key : U32_MAX;
 658         u32 *next = next_key;
 659
 660         if (index >= cmap->map.max_entries) {
 661                 *next = 0;
 662                 return 0;
 663         }
 664
 665         if (index == cmap->map.max_entries - 1)
 666                 return -ENOENT;
 667         *next = index + 1;
 668         return 0;
 669 }
 670
 671 static int cpu_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
 672 {
 673         return __bpf_xdp_redirect_map(map, ifindex, flags, 0,
 674                                       __cpu_map_lookup_elem);
 675 }
 676
 677 BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
 678 const struct bpf_map_ops cpu_map_ops = {
 679         .map_meta_equal         = bpf_map_meta_equal,
 680         .map_alloc              = cpu_map_alloc,
 681         .map_free               = cpu_map_free,
 682         .map_delete_elem        = cpu_map_delete_elem,
 683         .map_update_elem        = cpu_map_update_elem,
 684         .map_lookup_elem        = cpu_map_lookup_elem,
 685         .map_get_next_key       = cpu_map_get_next_key,
 686         .map_check_btf          = map_check_no_btf,
 687         .map_btf_id             = &cpu_map_btf_ids[0],
 688         .map_redirect           = cpu_map_redirect,
 689 };
 690
 691 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
 692 {
 693         struct bpf_cpu_map_entry *rcpu = bq->obj;
 694         unsigned int processed = 0, drops = 0;
 695         const int to_cpu = rcpu->cpu;
 696         struct ptr_ring *q;
 697         int i;
 698
 699         if (unlikely(!bq->count))
 700                 return;
 701
 702         q = rcpu->queue;
 703         spin_lock(&q->producer_lock);
 704
 705         for (i = 0; i < bq->count; i++) {
 706                 struct xdp_frame *xdpf = bq->q[i];
 707                 int err;
 708
 709                 err = __ptr_ring_produce(q, xdpf);
 710                 if (err) {
 711                         drops++;
 712                         xdp_return_frame_rx_napi(xdpf);
 713                 }
 714                 processed++;
 715         }
 716         bq->count = 0;
 717         spin_unlock(&q->producer_lock);
 718
 719         __list_del_clearprev(&bq->flush_node);
 720
 721         /* Feedback loop via tracepoints */
 722         trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
 723 }
 724
 725 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
 726  * Thus, safe percpu variable access.
 727  */
 728 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
 729 {
 730         struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
 731         struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
 732
 733         if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
 734                 bq_flush_to_queue(bq);
 735
 736         /* Notice, xdp_buff/page MUST be queued here, long enough for
 737          * driver to code invoking us to finished, due to driver
 738          * (e.g. ixgbe) recycle tricks based on page-refcnt.
 739          *
 740          * Thus, incoming xdp_frame is always queued here (else we race
 741          * with another CPU on page-refcnt and remaining driver code).
 742          * Queue time is very short, as driver will invoke flush
 743          * operation, when completing napi->poll call.
 744          */
 745         bq->q[bq->count++] = xdpf;
 746
 747         if (!bq->flush_node.prev)
 748                 list_add(&bq->flush_node, flush_list);
 749 }
 750
 751 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
 752                     struct net_device *dev_rx)
 753 {
 754         /* Info needed when constructing SKB on remote CPU */
 755         xdpf->dev_rx = dev_rx;
 756
 757         bq_enqueue(rcpu, xdpf);
 758         return 0;
 759 }
 760
 761 int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
 762                              struct sk_buff *skb)
 763 {
 764         int ret;
 765
 766         __skb_pull(skb, skb->mac_len);
 767         skb_set_redirected(skb, false);
 768         __ptr_set_bit(0, &skb);
 769
 770         ret = ptr_ring_produce(rcpu->queue, skb);
 771         if (ret < 0)
 772                 goto trace;
 773
 774         wake_up_process(rcpu->kthread);
 775 trace:
 776         trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
 777         return ret;
 778 }
 779
 780 void __cpu_map_flush(void)
 781 {
 782         struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
 783         struct xdp_bulk_queue *bq, *tmp;
 784
 785         list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 786                 bq_flush_to_queue(bq);
 787
 788                 /* If already running, costs spin_lock_irqsave + smb_mb */
 789                 wake_up_process(bq->obj->kthread);
 790         }
 791 }
 792
 793 static int __init cpu_map_init(void)
 794 {
 795         int cpu;
 796
 797         for_each_possible_cpu(cpu)
 798                 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
 799         return 0;
 800 }
 801
 802 subsys_initcall(cpu_map_init);