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
  31 #include <linux/netdevice.h>   /* netif_receive_skb_list */
  32 #include <linux/etherdevice.h> /* eth_type_trans */
  33
  34 /* General idea: XDP packets getting XDP redirected to another CPU,
  35  * will maximum be stored/queued for one driver ->poll() call.  It is
  36  * guaranteed that queueing the frame and the flush operation happen on
  37  * same CPU.  Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
  38  * which queue in bpf_cpu_map_entry contains packets.
  39  */
  40
  41 #define CPU_MAP_BULK_SIZE 8  /* 8 == one cacheline on 64-bit archs */
  42 struct bpf_cpu_map_entry;
  43 struct bpf_cpu_map;
  44
  45 struct xdp_bulk_queue {
  46         void *q[CPU_MAP_BULK_SIZE];
  47         struct list_head flush_node;
  48         struct bpf_cpu_map_entry *obj;
  49         unsigned int count;
  50 };
  51
  52 /* Struct for every remote "destination" CPU in map */
  53 struct bpf_cpu_map_entry {
  54         u32 cpu;    /* kthread CPU and map index */
  55         int map_id; /* Back reference to map */
  56
  57         /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
  58         struct xdp_bulk_queue __percpu *bulkq;
  59
  60         struct bpf_cpu_map *cmap;
  61
  62         /* Queue with potential multi-producers, and single-consumer kthread */
  63         struct ptr_ring *queue;
  64         struct task_struct *kthread;
  65
  66         struct bpf_cpumap_val value;
  67         struct bpf_prog *prog;
  68
  69         atomic_t refcnt; /* Control when this struct can be free'ed */
  70         struct rcu_head rcu;
  71
  72         struct work_struct kthread_stop_wq;
  73 };
  74
  75 struct bpf_cpu_map {
  76         struct bpf_map map;
  77         /* Below members specific for map type */
  78         struct bpf_cpu_map_entry __rcu **cpu_map;
  79 };
  80
  81 static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
  82
  83 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
  84 {
  85         u32 value_size = attr->value_size;
  86         struct bpf_cpu_map *cmap;
  87         int err = -ENOMEM;
  88
  89         if (!bpf_capable())
  90                 return ERR_PTR(-EPERM);
  91
  92         /* check sanity of attributes */
  93         if (attr->max_entries == 0 || attr->key_size != 4 ||
  94             (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
  95              value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
  96             attr->map_flags & ~BPF_F_NUMA_NODE)
  97                 return ERR_PTR(-EINVAL);
  98
  99         cmap = kzalloc(sizeof(*cmap), GFP_USER | __GFP_ACCOUNT);
 100         if (!cmap)
 101                 return ERR_PTR(-ENOMEM);
 102
 103         bpf_map_init_from_attr(&cmap->map, attr);
 104
 105         /* Pre-limit array size based on NR_CPUS, not final CPU check */
 106         if (cmap->map.max_entries > NR_CPUS) {
 107                 err = -E2BIG;
 108                 goto free_cmap;
 109         }
 110
 111         /* Alloc array for possible remote "destination" CPUs */
 112         cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
 113                                            sizeof(struct bpf_cpu_map_entry *),
 114                                            cmap->map.numa_node);
 115         if (!cmap->cpu_map)
 116                 goto free_cmap;
 117
 118         return &cmap->map;
 119 free_cmap:
 120         kfree(cmap);
 121         return ERR_PTR(err);
 122 }
 123
 124 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
 125 {
 126         atomic_inc(&rcpu->refcnt);
 127 }
 128
 129 /* called from workqueue, to workaround syscall using preempt_disable */
 130 static void cpu_map_kthread_stop(struct work_struct *work)
 131 {
 132         struct bpf_cpu_map_entry *rcpu;
 133
 134         rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
 135
 136         /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
 137          * as it waits until all in-flight call_rcu() callbacks complete.
 138          */
 139         rcu_barrier();
 140
 141         /* kthread_stop will wake_up_process and wait for it to complete */
 142         kthread_stop(rcpu->kthread);
 143 }
 144
 145 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
 146 {
 147         /* The tear-down procedure should have made sure that queue is
 148          * empty.  See __cpu_map_entry_replace() and work-queue
 149          * invoked cpu_map_kthread_stop(). Catch any broken behaviour
 150          * gracefully and warn once.
 151          */
 152         struct xdp_frame *xdpf;
 153
 154         while ((xdpf = ptr_ring_consume(ring)))
 155                 if (WARN_ON_ONCE(xdpf))
 156                         xdp_return_frame(xdpf);
 157 }
 158
 159 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
 160 {
 161         if (atomic_dec_and_test(&rcpu->refcnt)) {
 162                 if (rcpu->prog)
 163                         bpf_prog_put(rcpu->prog);
 164                 /* The queue should be empty at this point */
 165                 __cpu_map_ring_cleanup(rcpu->queue);
 166                 ptr_ring_cleanup(rcpu->queue, NULL);
 167                 kfree(rcpu->queue);
 168                 kfree(rcpu);
 169         }
 170 }
 171
 172 static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
 173                                      struct list_head *listp,
 174                                      struct xdp_cpumap_stats *stats)
 175 {
 176         struct sk_buff *skb, *tmp;
 177         struct xdp_buff xdp;
 178         u32 act;
 179         int err;
 180
 181         list_for_each_entry_safe(skb, tmp, listp, list) {
 182                 act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
 183                 switch (act) {
 184                 case XDP_PASS:
 185                         break;
 186                 case XDP_REDIRECT:
 187                         skb_list_del_init(skb);
 188                         err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
 189                                                       rcpu->prog);
 190                         if (unlikely(err)) {
 191                                 kfree_skb(skb);
 192                                 stats->drop++;
 193                         } else {
 194                                 stats->redirect++;
 195                         }
 196                         return;
 197                 default:
 198                         bpf_warn_invalid_xdp_action(act);
 199                         fallthrough;
 200                 case XDP_ABORTED:
 201                         trace_xdp_exception(skb->dev, rcpu->prog, act);
 202                         fallthrough;
 203                 case XDP_DROP:
 204                         skb_list_del_init(skb);
 205                         kfree_skb(skb);
 206                         stats->drop++;
 207                         return;
 208                 }
 209         }
 210 }
 211
 212 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 213                                     void **frames, int n,
 214                                     struct xdp_cpumap_stats *stats)
 215 {
 216         struct xdp_rxq_info rxq;
 217         struct xdp_buff xdp;
 218         int i, nframes = 0;
 219
 220         xdp_set_return_frame_no_direct();
 221         xdp.rxq = &rxq;
 222
 223         for (i = 0; i < n; i++) {
 224                 struct xdp_frame *xdpf = frames[i];
 225                 u32 act;
 226                 int err;
 227
 228                 rxq.dev = xdpf->dev_rx;
 229                 rxq.mem = xdpf->mem;
 230                 /* TODO: report queue_index to xdp_rxq_info */
 231
 232                 xdp_convert_frame_to_buff(xdpf, &xdp);
 233
 234                 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
 235                 switch (act) {
 236                 case XDP_PASS:
 237                         err = xdp_update_frame_from_buff(&xdp, xdpf);
 238                         if (err < 0) {
 239                                 xdp_return_frame(xdpf);
 240                                 stats->drop++;
 241                         } else {
 242                                 frames[nframes++] = xdpf;
 243                                 stats->pass++;
 244                         }
 245                         break;
 246                 case XDP_REDIRECT:
 247                         err = xdp_do_redirect(xdpf->dev_rx, &xdp,
 248                                               rcpu->prog);
 249                         if (unlikely(err)) {
 250                                 xdp_return_frame(xdpf);
 251                                 stats->drop++;
 252                         } else {
 253                                 stats->redirect++;
 254                         }
 255                         break;
 256                 default:
 257                         bpf_warn_invalid_xdp_action(act);
 258                         fallthrough;
 259                 case XDP_DROP:
 260                         xdp_return_frame(xdpf);
 261                         stats->drop++;
 262                         break;
 263                 }
 264         }
 265
 266         xdp_clear_return_frame_no_direct();
 267
 268         return nframes;
 269 }
 270
 271 #define CPUMAP_BATCH 8
 272
 273 static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
 274                                 int xdp_n, struct xdp_cpumap_stats *stats,
 275                                 struct list_head *list)
 276 {
 277         int nframes;
 278
 279         if (!rcpu->prog)
 280                 return xdp_n;
 281
 282         rcu_read_lock_bh();
 283
 284         nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
 285
 286         if (stats->redirect)
 287                 xdp_do_flush();
 288
 289         if (unlikely(!list_empty(list)))
 290                 cpu_map_bpf_prog_run_skb(rcpu, list, stats);
 291
 292         rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
 293
 294         return nframes;
 295 }
 296
 297
 298 static int cpu_map_kthread_run(void *data)
 299 {
 300         struct bpf_cpu_map_entry *rcpu = data;
 301
 302         set_current_state(TASK_INTERRUPTIBLE);
 303
 304         /* When kthread gives stop order, then rcpu have been disconnected
 305          * from map, thus no new packets can enter. Remaining in-flight
 306          * per CPU stored packets are flushed to this queue.  Wait honoring
 307          * kthread_stop signal until queue is empty.
 308          */
 309         while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
 310                 struct xdp_cpumap_stats stats = {}; /* zero stats */
 311                 unsigned int kmem_alloc_drops = 0, sched = 0;
 312                 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
 313                 int i, n, m, nframes, xdp_n;
 314                 void *frames[CPUMAP_BATCH];
 315                 void *skbs[CPUMAP_BATCH];
 316                 LIST_HEAD(list);
 317
 318                 /* Release CPU reschedule checks */
 319                 if (__ptr_ring_empty(rcpu->queue)) {
 320                         set_current_state(TASK_INTERRUPTIBLE);
 321                         /* Recheck to avoid lost wake-up */
 322                         if (__ptr_ring_empty(rcpu->queue)) {
 323                                 schedule();
 324                                 sched = 1;
 325                         } else {
 326                                 __set_current_state(TASK_RUNNING);
 327                         }
 328                 } else {
 329                         sched = cond_resched();
 330                 }
 331
 332                 /*
 333                  * The bpf_cpu_map_entry is single consumer, with this
 334                  * kthread CPU pinned. Lockless access to ptr_ring
 335                  * consume side valid as no-resize allowed of queue.
 336                  */
 337                 n = __ptr_ring_consume_batched(rcpu->queue, frames,
 338                                                CPUMAP_BATCH);
 339                 for (i = 0, xdp_n = 0; i < n; i++) {
 340                         void *f = frames[i];
 341                         struct page *page;
 342
 343                         if (unlikely(__ptr_test_bit(0, &f))) {
 344                                 struct sk_buff *skb = f;
 345
 346                                 __ptr_clear_bit(0, &skb);
 347                                 list_add_tail(&skb->list, &list);
 348                                 continue;
 349                         }
 350
 351                         frames[xdp_n++] = f;
 352                         page = virt_to_page(f);
 353
 354                         /* Bring struct page memory area to curr CPU. Read by
 355                          * build_skb_around via page_is_pfmemalloc(), and when
 356                          * freed written by page_frag_free call.
 357                          */
 358                         prefetchw(page);
 359                 }
 360
 361                 /* Support running another XDP prog on this CPU */
 362                 nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
 363                 if (nframes) {
 364                         m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
 365                         if (unlikely(m == 0)) {
 366                                 for (i = 0; i < nframes; i++)
 367                                         skbs[i] = NULL; /* effect: xdp_return_frame */
 368                                 kmem_alloc_drops += nframes;
 369                         }
 370                 }
 371
 372                 local_bh_disable();
 373                 for (i = 0; i < nframes; i++) {
 374                         struct xdp_frame *xdpf = frames[i];
 375                         struct sk_buff *skb = skbs[i];
 376
 377                         skb = __xdp_build_skb_from_frame(xdpf, skb,
 378                                                          xdpf->dev_rx);
 379                         if (!skb) {
 380                                 xdp_return_frame(xdpf);
 381                                 continue;
 382                         }
 383
 384                         list_add_tail(&skb->list, &list);
 385                 }
 386                 netif_receive_skb_list(&list);
 387
 388                 /* Feedback loop via tracepoint */
 389                 trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
 390                                          sched, &stats);
 391
 392                 local_bh_enable(); /* resched point, may call do_softirq() */
 393         }
 394         __set_current_state(TASK_RUNNING);
 395
 396         put_cpu_map_entry(rcpu);
 397         return 0;
 398 }
 399
 400 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, int fd)
 401 {
 402         struct bpf_prog *prog;
 403
 404         prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
 405         if (IS_ERR(prog))
 406                 return PTR_ERR(prog);
 407
 408         if (prog->expected_attach_type != BPF_XDP_CPUMAP) {
 409                 bpf_prog_put(prog);
 410                 return -EINVAL;
 411         }
 412
 413         rcpu->value.bpf_prog.id = prog->aux->id;
 414         rcpu->prog = prog;
 415
 416         return 0;
 417 }
 418
 419 static struct bpf_cpu_map_entry *
 420 __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
 421                       u32 cpu)
 422 {
 423         int numa, err, i, fd = value->bpf_prog.fd;
 424         gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
 425         struct bpf_cpu_map_entry *rcpu;
 426         struct xdp_bulk_queue *bq;
 427
 428         /* Have map->numa_node, but choose node of redirect target CPU */
 429         numa = cpu_to_node(cpu);
 430
 431         rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
 432         if (!rcpu)
 433                 return NULL;
 434
 435         /* Alloc percpu bulkq */
 436         rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
 437                                            sizeof(void *), gfp);
 438         if (!rcpu->bulkq)
 439                 goto free_rcu;
 440
 441         for_each_possible_cpu(i) {
 442                 bq = per_cpu_ptr(rcpu->bulkq, i);
 443                 bq->obj = rcpu;
 444         }
 445
 446         /* Alloc queue */
 447         rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
 448                                            numa);
 449         if (!rcpu->queue)
 450                 goto free_bulkq;
 451
 452         err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
 453         if (err)
 454                 goto free_queue;
 455
 456         rcpu->cpu    = cpu;
 457         rcpu->map_id = map->id;
 458         rcpu->value.qsize  = value->qsize;
 459
 460         if (fd > 0 && __cpu_map_load_bpf_program(rcpu, fd))
 461                 goto free_ptr_ring;
 462
 463         /* Setup kthread */
 464         rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
 465                                                "cpumap/%d/map:%d", cpu,
 466                                                map->id);
 467         if (IS_ERR(rcpu->kthread))
 468                 goto free_prog;
 469
 470         get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
 471         get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
 472
 473         /* Make sure kthread runs on a single CPU */
 474         kthread_bind(rcpu->kthread, cpu);
 475         wake_up_process(rcpu->kthread);
 476
 477         return rcpu;
 478
 479 free_prog:
 480         if (rcpu->prog)
 481                 bpf_prog_put(rcpu->prog);
 482 free_ptr_ring:
 483         ptr_ring_cleanup(rcpu->queue, NULL);
 484 free_queue:
 485         kfree(rcpu->queue);
 486 free_bulkq:
 487         free_percpu(rcpu->bulkq);
 488 free_rcu:
 489         kfree(rcpu);
 490         return NULL;
 491 }
 492
 493 static void __cpu_map_entry_free(struct rcu_head *rcu)
 494 {
 495         struct bpf_cpu_map_entry *rcpu;
 496
 497         /* This cpu_map_entry have been disconnected from map and one
 498          * RCU grace-period have elapsed.  Thus, XDP cannot queue any
 499          * new packets and cannot change/set flush_needed that can
 500          * find this entry.
 501          */
 502         rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
 503
 504         free_percpu(rcpu->bulkq);
 505         /* Cannot kthread_stop() here, last put free rcpu resources */
 506         put_cpu_map_entry(rcpu);
 507 }
 508
 509 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
 510  * ensure any driver rcu critical sections have completed, but this
 511  * does not guarantee a flush has happened yet. Because driver side
 512  * rcu_read_lock/unlock only protects the running XDP program.  The
 513  * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
 514  * pending flush op doesn't fail.
 515  *
 516  * The bpf_cpu_map_entry is still used by the kthread, and there can
 517  * still be pending packets (in queue and percpu bulkq).  A refcnt
 518  * makes sure to last user (kthread_stop vs. call_rcu) free memory
 519  * resources.
 520  *
 521  * The rcu callback __cpu_map_entry_free flush remaining packets in
 522  * percpu bulkq to queue.  Due to caller map_delete_elem() disable
 523  * preemption, cannot call kthread_stop() to make sure queue is empty.
 524  * Instead a work_queue is started for stopping kthread,
 525  * cpu_map_kthread_stop, which waits for an RCU grace period before
 526  * stopping kthread, emptying the queue.
 527  */
 528 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
 529                                     u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
 530 {
 531         struct bpf_cpu_map_entry *old_rcpu;
 532
 533         old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
 534         if (old_rcpu) {
 535                 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
 536                 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
 537                 schedule_work(&old_rcpu->kthread_stop_wq);
 538         }
 539 }
 540
 541 static int cpu_map_delete_elem(struct bpf_map *map, void *key)
 542 {
 543         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 544         u32 key_cpu = *(u32 *)key;
 545
 546         if (key_cpu >= map->max_entries)
 547                 return -EINVAL;
 548
 549         /* notice caller map_delete_elem() use preempt_disable() */
 550         __cpu_map_entry_replace(cmap, key_cpu, NULL);
 551         return 0;
 552 }
 553
 554 static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
 555                                u64 map_flags)
 556 {
 557         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 558         struct bpf_cpumap_val cpumap_value = {};
 559         struct bpf_cpu_map_entry *rcpu;
 560         /* Array index key correspond to CPU number */
 561         u32 key_cpu = *(u32 *)key;
 562
 563         memcpy(&cpumap_value, value, map->value_size);
 564
 565         if (unlikely(map_flags > BPF_EXIST))
 566                 return -EINVAL;
 567         if (unlikely(key_cpu >= cmap->map.max_entries))
 568                 return -E2BIG;
 569         if (unlikely(map_flags == BPF_NOEXIST))
 570                 return -EEXIST;
 571         if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
 572                 return -EOVERFLOW;
 573
 574         /* Make sure CPU is a valid possible cpu */
 575         if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
 576                 return -ENODEV;
 577
 578         if (cpumap_value.qsize == 0) {
 579                 rcpu = NULL; /* Same as deleting */
 580         } else {
 581                 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
 582                 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
 583                 if (!rcpu)
 584                         return -ENOMEM;
 585                 rcpu->cmap = cmap;
 586         }
 587         rcu_read_lock();
 588         __cpu_map_entry_replace(cmap, key_cpu, rcpu);
 589         rcu_read_unlock();
 590         return 0;
 591 }
 592
 593 static void cpu_map_free(struct bpf_map *map)
 594 {
 595         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 596         u32 i;
 597
 598         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 599          * so the bpf programs (can be more than one that used this map) were
 600          * disconnected from events. Wait for outstanding critical sections in
 601          * these programs to complete. The rcu critical section only guarantees
 602          * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
 603          * It does __not__ ensure pending flush operations (if any) are
 604          * complete.
 605          */
 606
 607         synchronize_rcu();
 608
 609         /* For cpu_map the remote CPUs can still be using the entries
 610          * (struct bpf_cpu_map_entry).
 611          */
 612         for (i = 0; i < cmap->map.max_entries; i++) {
 613                 struct bpf_cpu_map_entry *rcpu;
 614
 615                 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
 616                 if (!rcpu)
 617                         continue;
 618
 619                 /* bq flush and cleanup happens after RCU grace-period */
 620                 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
 621         }
 622         bpf_map_area_free(cmap->cpu_map);
 623         kfree(cmap);
 624 }
 625
 626 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 627  * by local_bh_disable() (from XDP calls inside NAPI). The
 628  * rcu_read_lock_bh_held() below makes lockdep accept both.
 629  */
 630 static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
 631 {
 632         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 633         struct bpf_cpu_map_entry *rcpu;
 634
 635         if (key >= map->max_entries)
 636                 return NULL;
 637
 638         rcpu = rcu_dereference_check(cmap->cpu_map[key],
 639                                      rcu_read_lock_bh_held());
 640         return rcpu;
 641 }
 642
 643 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
 644 {
 645         struct bpf_cpu_map_entry *rcpu =
 646                 __cpu_map_lookup_elem(map, *(u32 *)key);
 647
 648         return rcpu ? &rcpu->value : NULL;
 649 }
 650
 651 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 652 {
 653         struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 654         u32 index = key ? *(u32 *)key : U32_MAX;
 655         u32 *next = next_key;
 656
 657         if (index >= cmap->map.max_entries) {
 658                 *next = 0;
 659                 return 0;
 660         }
 661
 662         if (index == cmap->map.max_entries - 1)
 663                 return -ENOENT;
 664         *next = index + 1;
 665         return 0;
 666 }
 667
 668 static int cpu_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
 669 {
 670         return __bpf_xdp_redirect_map(map, ifindex, flags, 0,
 671                                       __cpu_map_lookup_elem);
 672 }
 673
 674 static int cpu_map_btf_id;
 675 const struct bpf_map_ops cpu_map_ops = {
 676         .map_meta_equal         = bpf_map_meta_equal,
 677         .map_alloc              = cpu_map_alloc,
 678         .map_free               = cpu_map_free,
 679         .map_delete_elem        = cpu_map_delete_elem,
 680         .map_update_elem        = cpu_map_update_elem,
 681         .map_lookup_elem        = cpu_map_lookup_elem,
 682         .map_get_next_key       = cpu_map_get_next_key,
 683         .map_check_btf          = map_check_no_btf,
 684         .map_btf_name           = "bpf_cpu_map",
 685         .map_btf_id             = &cpu_map_btf_id,
 686         .map_redirect           = cpu_map_redirect,
 687 };
 688
 689 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
 690 {
 691         struct bpf_cpu_map_entry *rcpu = bq->obj;
 692         unsigned int processed = 0, drops = 0;
 693         const int to_cpu = rcpu->cpu;
 694         struct ptr_ring *q;
 695         int i;
 696
 697         if (unlikely(!bq->count))
 698                 return;
 699
 700         q = rcpu->queue;
 701         spin_lock(&q->producer_lock);
 702
 703         for (i = 0; i < bq->count; i++) {
 704                 struct xdp_frame *xdpf = bq->q[i];
 705                 int err;
 706
 707                 err = __ptr_ring_produce(q, xdpf);
 708                 if (err) {
 709                         drops++;
 710                         xdp_return_frame_rx_napi(xdpf);
 711                 }
 712                 processed++;
 713         }
 714         bq->count = 0;
 715         spin_unlock(&q->producer_lock);
 716
 717         __list_del_clearprev(&bq->flush_node);
 718
 719         /* Feedback loop via tracepoints */
 720         trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
 721 }
 722
 723 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
 724  * Thus, safe percpu variable access.
 725  */
 726 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
 727 {
 728         struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
 729         struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
 730
 731         if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
 732                 bq_flush_to_queue(bq);
 733
 734         /* Notice, xdp_buff/page MUST be queued here, long enough for
 735          * driver to code invoking us to finished, due to driver
 736          * (e.g. ixgbe) recycle tricks based on page-refcnt.
 737          *
 738          * Thus, incoming xdp_frame is always queued here (else we race
 739          * with another CPU on page-refcnt and remaining driver code).
 740          * Queue time is very short, as driver will invoke flush
 741          * operation, when completing napi->poll call.
 742          */
 743         bq->q[bq->count++] = xdpf;
 744
 745         if (!bq->flush_node.prev)
 746                 list_add(&bq->flush_node, flush_list);
 747 }
 748
 749 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
 750                     struct net_device *dev_rx)
 751 {
 752         struct xdp_frame *xdpf;
 753
 754         xdpf = xdp_convert_buff_to_frame(xdp);
 755         if (unlikely(!xdpf))
 756                 return -EOVERFLOW;
 757
 758         /* Info needed when constructing SKB on remote CPU */
 759         xdpf->dev_rx = dev_rx;
 760
 761         bq_enqueue(rcpu, xdpf);
 762         return 0;
 763 }
 764
 765 int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
 766                              struct sk_buff *skb)
 767 {
 768         int ret;
 769
 770         __skb_pull(skb, skb->mac_len);
 771         skb_set_redirected(skb, false);
 772         __ptr_set_bit(0, &skb);
 773
 774         ret = ptr_ring_produce(rcpu->queue, skb);
 775         if (ret < 0)
 776                 goto trace;
 777
 778         wake_up_process(rcpu->kthread);
 779 trace:
 780         trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
 781         return ret;
 782 }
 783
 784 void __cpu_map_flush(void)
 785 {
 786         struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
 787         struct xdp_bulk_queue *bq, *tmp;
 788
 789         list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 790                 bq_flush_to_queue(bq);
 791
 792                 /* If already running, costs spin_lock_irqsave + smb_mb */
 793                 wake_up_process(bq->obj->kthread);
 794         }
 795 }
 796
 797 static int __init cpu_map_init(void)
 798 {
 799         int cpu;
 800
 801         for_each_possible_cpu(cpu)
 802                 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
 803         return 0;
 804 }
 805
 806 subsys_initcall(cpu_map_init);