Thomas Gleixner | ddc64d0 | 2019-05-31 01:09:24 -0700 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0-only |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 2 | /* bpf/cpumap.c |
| 3 | * |
| 4 | * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc. |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 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/bpf.h> |
| 20 | #include <linux/filter.h> |
| 21 | #include <linux/ptr_ring.h> |
Jesper Dangaard Brouer | 5ab073f | 2018-04-17 16:45:26 +0200 | [diff] [blame] | 22 | #include <net/xdp.h> |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 23 | |
| 24 | #include <linux/sched.h> |
| 25 | #include <linux/workqueue.h> |
| 26 | #include <linux/kthread.h> |
| 27 | #include <linux/capability.h> |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 28 | #include <trace/events/xdp.h> |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 29 | |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 30 | #include <linux/netdevice.h> /* netif_receive_skb_core */ |
| 31 | #include <linux/etherdevice.h> /* eth_type_trans */ |
| 32 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 33 | /* General idea: XDP packets getting XDP redirected to another CPU, |
| 34 | * will maximum be stored/queued for one driver ->poll() call. It is |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 35 | * guaranteed that queueing the frame and the flush operation happen on |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 36 | * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr() |
| 37 | * which queue in bpf_cpu_map_entry contains packets. |
| 38 | */ |
| 39 | |
| 40 | #define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */ |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 41 | struct bpf_cpu_map_entry; |
| 42 | struct bpf_cpu_map; |
| 43 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 44 | struct xdp_bulk_queue { |
| 45 | void *q[CPU_MAP_BULK_SIZE]; |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 46 | struct list_head flush_node; |
| 47 | struct bpf_cpu_map_entry *obj; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 48 | unsigned int count; |
| 49 | }; |
| 50 | |
| 51 | /* Struct for every remote "destination" CPU in map */ |
| 52 | struct bpf_cpu_map_entry { |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 53 | u32 cpu; /* kthread CPU and map index */ |
| 54 | int map_id; /* Back reference to map */ |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 55 | |
| 56 | /* XDP can run multiple RX-ring queues, need __percpu enqueue store */ |
| 57 | struct xdp_bulk_queue __percpu *bulkq; |
| 58 | |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 59 | struct bpf_cpu_map *cmap; |
| 60 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 61 | /* Queue with potential multi-producers, and single-consumer kthread */ |
| 62 | struct ptr_ring *queue; |
| 63 | struct task_struct *kthread; |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 64 | |
| 65 | struct bpf_cpumap_val value; |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 66 | struct bpf_prog *prog; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 67 | |
| 68 | atomic_t refcnt; /* Control when this struct can be free'ed */ |
| 69 | struct rcu_head rcu; |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 70 | |
| 71 | struct work_struct kthread_stop_wq; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 72 | }; |
| 73 | |
| 74 | struct bpf_cpu_map { |
| 75 | struct bpf_map map; |
| 76 | /* Below members specific for map type */ |
| 77 | struct bpf_cpu_map_entry **cpu_map; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 78 | }; |
| 79 | |
Björn Töpel | cdfafe9 | 2019-12-19 07:10:04 +0100 | [diff] [blame] | 80 | static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list); |
| 81 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 82 | static struct bpf_map *cpu_map_alloc(union bpf_attr *attr) |
| 83 | { |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 84 | u32 value_size = attr->value_size; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 85 | struct bpf_cpu_map *cmap; |
| 86 | int err = -ENOMEM; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 87 | |
Alexei Starovoitov | 2c78ee8 | 2020-05-13 16:03:54 -0700 | [diff] [blame] | 88 | if (!bpf_capable()) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 89 | return ERR_PTR(-EPERM); |
| 90 | |
| 91 | /* check sanity of attributes */ |
| 92 | if (attr->max_entries == 0 || attr->key_size != 4 || |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 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) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 96 | return ERR_PTR(-EINVAL); |
| 97 | |
Roman Gushchin | e88cc05 | 2020-12-01 13:58:35 -0800 | [diff] [blame] | 98 | cmap = kzalloc(sizeof(*cmap), GFP_USER | __GFP_ACCOUNT); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 99 | if (!cmap) |
| 100 | return ERR_PTR(-ENOMEM); |
| 101 | |
Jakub Kicinski | bd47564 | 2018-01-11 20:29:06 -0800 | [diff] [blame] | 102 | bpf_map_init_from_attr(&cmap->map, attr); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 103 | |
| 104 | /* Pre-limit array size based on NR_CPUS, not final CPU check */ |
| 105 | if (cmap->map.max_entries > NR_CPUS) { |
| 106 | err = -E2BIG; |
| 107 | goto free_cmap; |
| 108 | } |
| 109 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 110 | /* Alloc array for possible remote "destination" CPUs */ |
| 111 | cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries * |
| 112 | sizeof(struct bpf_cpu_map_entry *), |
| 113 | cmap->map.numa_node); |
| 114 | if (!cmap->cpu_map) |
Roman Gushchin | 711caba | 2020-12-01 13:58:46 -0800 | [diff] [blame^] | 115 | goto free_cmap; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 116 | |
| 117 | return &cmap->map; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 118 | free_cmap: |
| 119 | kfree(cmap); |
| 120 | return ERR_PTR(err); |
| 121 | } |
| 122 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 123 | static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu) |
| 124 | { |
| 125 | atomic_inc(&rcpu->refcnt); |
| 126 | } |
| 127 | |
| 128 | /* called from workqueue, to workaround syscall using preempt_disable */ |
| 129 | static void cpu_map_kthread_stop(struct work_struct *work) |
| 130 | { |
| 131 | struct bpf_cpu_map_entry *rcpu; |
| 132 | |
| 133 | rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq); |
| 134 | |
| 135 | /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier, |
| 136 | * as it waits until all in-flight call_rcu() callbacks complete. |
| 137 | */ |
| 138 | rcu_barrier(); |
| 139 | |
| 140 | /* kthread_stop will wake_up_process and wait for it to complete */ |
| 141 | kthread_stop(rcpu->kthread); |
| 142 | } |
| 143 | |
Lorenzo Bianconi | efa90b5 | 2020-09-28 13:24:57 +0200 | [diff] [blame] | 144 | static struct sk_buff *cpu_map_build_skb(struct xdp_frame *xdpf, |
Jesper Dangaard Brouer | 8f0504a9 | 2019-04-12 17:07:43 +0200 | [diff] [blame] | 145 | struct sk_buff *skb) |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 146 | { |
Jesper Dangaard Brouer | 676e4a6 | 2019-03-29 10:18:00 +0100 | [diff] [blame] | 147 | unsigned int hard_start_headroom; |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 148 | unsigned int frame_size; |
| 149 | void *pkt_data_start; |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 150 | |
Jesper Dangaard Brouer | 676e4a6 | 2019-03-29 10:18:00 +0100 | [diff] [blame] | 151 | /* Part of headroom was reserved to xdpf */ |
| 152 | hard_start_headroom = sizeof(struct xdp_frame) + xdpf->headroom; |
| 153 | |
Jesper Dangaard Brouer | db612f7 | 2020-05-14 12:49:38 +0200 | [diff] [blame] | 154 | /* Memory size backing xdp_frame data already have reserved |
| 155 | * room for build_skb to place skb_shared_info in tailroom. |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 156 | */ |
Jesper Dangaard Brouer | db612f7 | 2020-05-14 12:49:38 +0200 | [diff] [blame] | 157 | frame_size = xdpf->frame_sz; |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 158 | |
Jesper Dangaard Brouer | 676e4a6 | 2019-03-29 10:18:00 +0100 | [diff] [blame] | 159 | pkt_data_start = xdpf->data - hard_start_headroom; |
Jesper Dangaard Brouer | 8f0504a9 | 2019-04-12 17:07:43 +0200 | [diff] [blame] | 160 | skb = build_skb_around(skb, pkt_data_start, frame_size); |
| 161 | if (unlikely(!skb)) |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 162 | return NULL; |
| 163 | |
Jesper Dangaard Brouer | 676e4a6 | 2019-03-29 10:18:00 +0100 | [diff] [blame] | 164 | skb_reserve(skb, hard_start_headroom); |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 165 | __skb_put(skb, xdpf->len); |
| 166 | if (xdpf->metasize) |
| 167 | skb_metadata_set(skb, xdpf->metasize); |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 168 | |
| 169 | /* Essential SKB info: protocol and skb->dev */ |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 170 | skb->protocol = eth_type_trans(skb, xdpf->dev_rx); |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 171 | |
| 172 | /* Optional SKB info, currently missing: |
| 173 | * - HW checksum info (skb->ip_summed) |
| 174 | * - HW RX hash (skb_set_hash) |
| 175 | * - RX ring dev queue index (skb_record_rx_queue) |
| 176 | */ |
| 177 | |
Jesper Dangaard Brouer | 6bf071b | 2019-06-18 15:05:27 +0200 | [diff] [blame] | 178 | /* Until page_pool get SKB return path, release DMA here */ |
| 179 | xdp_release_frame(xdpf); |
| 180 | |
Jesper Dangaard Brouer | 676e4a6 | 2019-03-29 10:18:00 +0100 | [diff] [blame] | 181 | /* Allow SKB to reuse area used by xdp_frame */ |
| 182 | xdp_scrub_frame(xdpf); |
| 183 | |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 184 | return skb; |
| 185 | } |
| 186 | |
Jesper Dangaard Brouer | 5ab073f | 2018-04-17 16:45:26 +0200 | [diff] [blame] | 187 | static void __cpu_map_ring_cleanup(struct ptr_ring *ring) |
| 188 | { |
| 189 | /* The tear-down procedure should have made sure that queue is |
| 190 | * empty. See __cpu_map_entry_replace() and work-queue |
| 191 | * invoked cpu_map_kthread_stop(). Catch any broken behaviour |
| 192 | * gracefully and warn once. |
| 193 | */ |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 194 | struct xdp_frame *xdpf; |
Jesper Dangaard Brouer | 5ab073f | 2018-04-17 16:45:26 +0200 | [diff] [blame] | 195 | |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 196 | while ((xdpf = ptr_ring_consume(ring))) |
| 197 | if (WARN_ON_ONCE(xdpf)) |
Jesper Dangaard Brouer | 0399309 | 2018-04-17 16:46:32 +0200 | [diff] [blame] | 198 | xdp_return_frame(xdpf); |
Jesper Dangaard Brouer | 5ab073f | 2018-04-17 16:45:26 +0200 | [diff] [blame] | 199 | } |
| 200 | |
| 201 | static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu) |
| 202 | { |
| 203 | if (atomic_dec_and_test(&rcpu->refcnt)) { |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 204 | if (rcpu->prog) |
| 205 | bpf_prog_put(rcpu->prog); |
Jesper Dangaard Brouer | 5ab073f | 2018-04-17 16:45:26 +0200 | [diff] [blame] | 206 | /* The queue should be empty at this point */ |
| 207 | __cpu_map_ring_cleanup(rcpu->queue); |
| 208 | ptr_ring_cleanup(rcpu->queue, NULL); |
| 209 | kfree(rcpu->queue); |
| 210 | kfree(rcpu); |
| 211 | } |
| 212 | } |
| 213 | |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 214 | static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu, |
| 215 | void **frames, int n, |
| 216 | struct xdp_cpumap_stats *stats) |
| 217 | { |
| 218 | struct xdp_rxq_info rxq; |
| 219 | struct xdp_buff xdp; |
| 220 | int i, nframes = 0; |
| 221 | |
| 222 | if (!rcpu->prog) |
| 223 | return n; |
| 224 | |
Lorenzo Bianconi | 28b1520 | 2020-07-14 15:56:39 +0200 | [diff] [blame] | 225 | rcu_read_lock_bh(); |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 226 | |
| 227 | xdp_set_return_frame_no_direct(); |
| 228 | xdp.rxq = &rxq; |
| 229 | |
| 230 | for (i = 0; i < n; i++) { |
| 231 | struct xdp_frame *xdpf = frames[i]; |
| 232 | u32 act; |
| 233 | int err; |
| 234 | |
| 235 | rxq.dev = xdpf->dev_rx; |
| 236 | rxq.mem = xdpf->mem; |
| 237 | /* TODO: report queue_index to xdp_rxq_info */ |
| 238 | |
| 239 | xdp_convert_frame_to_buff(xdpf, &xdp); |
| 240 | |
| 241 | act = bpf_prog_run_xdp(rcpu->prog, &xdp); |
| 242 | switch (act) { |
| 243 | case XDP_PASS: |
| 244 | err = xdp_update_frame_from_buff(&xdp, xdpf); |
| 245 | if (err < 0) { |
| 246 | xdp_return_frame(xdpf); |
| 247 | stats->drop++; |
| 248 | } else { |
| 249 | frames[nframes++] = xdpf; |
| 250 | stats->pass++; |
| 251 | } |
| 252 | break; |
Lorenzo Bianconi | 28b1520 | 2020-07-14 15:56:39 +0200 | [diff] [blame] | 253 | case XDP_REDIRECT: |
| 254 | err = xdp_do_redirect(xdpf->dev_rx, &xdp, |
| 255 | rcpu->prog); |
| 256 | if (unlikely(err)) { |
| 257 | xdp_return_frame(xdpf); |
| 258 | stats->drop++; |
| 259 | } else { |
| 260 | stats->redirect++; |
| 261 | } |
| 262 | break; |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 263 | default: |
| 264 | bpf_warn_invalid_xdp_action(act); |
Gustavo A. R. Silva | df561f66 | 2020-08-23 17:36:59 -0500 | [diff] [blame] | 265 | fallthrough; |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 266 | case XDP_DROP: |
| 267 | xdp_return_frame(xdpf); |
| 268 | stats->drop++; |
| 269 | break; |
| 270 | } |
| 271 | } |
| 272 | |
Lorenzo Bianconi | 28b1520 | 2020-07-14 15:56:39 +0200 | [diff] [blame] | 273 | if (stats->redirect) |
| 274 | xdp_do_flush_map(); |
| 275 | |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 276 | xdp_clear_return_frame_no_direct(); |
| 277 | |
Lorenzo Bianconi | 28b1520 | 2020-07-14 15:56:39 +0200 | [diff] [blame] | 278 | rcu_read_unlock_bh(); /* resched point, may call do_softirq() */ |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 279 | |
| 280 | return nframes; |
| 281 | } |
| 282 | |
Jesper Dangaard Brouer | 7736182 | 2019-04-12 17:07:32 +0200 | [diff] [blame] | 283 | #define CPUMAP_BATCH 8 |
| 284 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 285 | static int cpu_map_kthread_run(void *data) |
| 286 | { |
| 287 | struct bpf_cpu_map_entry *rcpu = data; |
| 288 | |
| 289 | set_current_state(TASK_INTERRUPTIBLE); |
| 290 | |
| 291 | /* When kthread gives stop order, then rcpu have been disconnected |
| 292 | * from map, thus no new packets can enter. Remaining in-flight |
| 293 | * per CPU stored packets are flushed to this queue. Wait honoring |
| 294 | * kthread_stop signal until queue is empty. |
| 295 | */ |
| 296 | while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) { |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 297 | struct xdp_cpumap_stats stats = {}; /* zero stats */ |
| 298 | gfp_t gfp = __GFP_ZERO | GFP_ATOMIC; |
Jesper Dangaard Brouer | 7736182 | 2019-04-12 17:07:32 +0200 | [diff] [blame] | 299 | unsigned int drops = 0, sched = 0; |
| 300 | void *frames[CPUMAP_BATCH]; |
Jesper Dangaard Brouer | 8f0504a9 | 2019-04-12 17:07:43 +0200 | [diff] [blame] | 301 | void *skbs[CPUMAP_BATCH]; |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 302 | int i, n, m, nframes; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 303 | |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 304 | /* Release CPU reschedule checks */ |
| 305 | if (__ptr_ring_empty(rcpu->queue)) { |
Jesper Dangaard Brouer | 3174946 | 2017-10-23 19:39:28 +0200 | [diff] [blame] | 306 | set_current_state(TASK_INTERRUPTIBLE); |
| 307 | /* Recheck to avoid lost wake-up */ |
| 308 | if (__ptr_ring_empty(rcpu->queue)) { |
| 309 | schedule(); |
| 310 | sched = 1; |
| 311 | } else { |
| 312 | __set_current_state(TASK_RUNNING); |
| 313 | } |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 314 | } else { |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 315 | sched = cond_resched(); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 316 | } |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 317 | |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 318 | /* |
| 319 | * The bpf_cpu_map_entry is single consumer, with this |
| 320 | * kthread CPU pinned. Lockless access to ptr_ring |
| 321 | * consume side valid as no-resize allowed of queue. |
| 322 | */ |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 323 | n = __ptr_ring_consume_batched(rcpu->queue, frames, |
| 324 | CPUMAP_BATCH); |
Jesper Dangaard Brouer | 86d2314 | 2019-04-12 17:07:48 +0200 | [diff] [blame] | 325 | for (i = 0; i < n; i++) { |
| 326 | void *f = frames[i]; |
| 327 | struct page *page = virt_to_page(f); |
| 328 | |
| 329 | /* Bring struct page memory area to curr CPU. Read by |
| 330 | * build_skb_around via page_is_pfmemalloc(), and when |
| 331 | * freed written by page_frag_free call. |
| 332 | */ |
| 333 | prefetchw(page); |
| 334 | } |
| 335 | |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 336 | /* Support running another XDP prog on this CPU */ |
| 337 | nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, n, &stats); |
| 338 | if (nframes) { |
| 339 | m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs); |
| 340 | if (unlikely(m == 0)) { |
| 341 | for (i = 0; i < nframes; i++) |
| 342 | skbs[i] = NULL; /* effect: xdp_return_frame */ |
| 343 | drops += nframes; |
| 344 | } |
Jesper Dangaard Brouer | 8f0504a9 | 2019-04-12 17:07:43 +0200 | [diff] [blame] | 345 | } |
Jesper Dangaard Brouer | 7736182 | 2019-04-12 17:07:32 +0200 | [diff] [blame] | 346 | |
| 347 | local_bh_disable(); |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 348 | for (i = 0; i < nframes; i++) { |
Jesper Dangaard Brouer | 7736182 | 2019-04-12 17:07:32 +0200 | [diff] [blame] | 349 | struct xdp_frame *xdpf = frames[i]; |
Jesper Dangaard Brouer | 8f0504a9 | 2019-04-12 17:07:43 +0200 | [diff] [blame] | 350 | struct sk_buff *skb = skbs[i]; |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 351 | int ret; |
| 352 | |
Lorenzo Bianconi | efa90b5 | 2020-09-28 13:24:57 +0200 | [diff] [blame] | 353 | skb = cpu_map_build_skb(xdpf, skb); |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 354 | if (!skb) { |
Jesper Dangaard Brouer | 0399309 | 2018-04-17 16:46:32 +0200 | [diff] [blame] | 355 | xdp_return_frame(xdpf); |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 356 | continue; |
| 357 | } |
| 358 | |
| 359 | /* Inject into network stack */ |
| 360 | ret = netif_receive_skb_core(skb); |
| 361 | if (ret == NET_RX_DROP) |
| 362 | drops++; |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 363 | } |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 364 | /* Feedback loop via tracepoint */ |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 365 | trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched, &stats); |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 366 | |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 367 | local_bh_enable(); /* resched point, may call do_softirq() */ |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 368 | } |
| 369 | __set_current_state(TASK_RUNNING); |
| 370 | |
| 371 | put_cpu_map_entry(rcpu); |
| 372 | return 0; |
| 373 | } |
| 374 | |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 375 | bool cpu_map_prog_allowed(struct bpf_map *map) |
| 376 | { |
| 377 | return map->map_type == BPF_MAP_TYPE_CPUMAP && |
| 378 | map->value_size != offsetofend(struct bpf_cpumap_val, qsize); |
| 379 | } |
| 380 | |
| 381 | static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, int fd) |
| 382 | { |
| 383 | struct bpf_prog *prog; |
| 384 | |
| 385 | prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP); |
| 386 | if (IS_ERR(prog)) |
| 387 | return PTR_ERR(prog); |
| 388 | |
| 389 | if (prog->expected_attach_type != BPF_XDP_CPUMAP) { |
| 390 | bpf_prog_put(prog); |
| 391 | return -EINVAL; |
| 392 | } |
| 393 | |
| 394 | rcpu->value.bpf_prog.id = prog->aux->id; |
| 395 | rcpu->prog = prog; |
| 396 | |
| 397 | return 0; |
| 398 | } |
| 399 | |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 400 | static struct bpf_cpu_map_entry * |
Roman Gushchin | e88cc05 | 2020-12-01 13:58:35 -0800 | [diff] [blame] | 401 | __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value, |
| 402 | u32 cpu) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 403 | { |
Lorenzo Bianconi | 9216477 | 2020-07-14 15:56:38 +0200 | [diff] [blame] | 404 | int numa, err, i, fd = value->bpf_prog.fd; |
Jason Wang | 7fc17e9 | 2018-02-14 22:17:34 +0800 | [diff] [blame] | 405 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 406 | struct bpf_cpu_map_entry *rcpu; |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 407 | struct xdp_bulk_queue *bq; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 408 | |
| 409 | /* Have map->numa_node, but choose node of redirect target CPU */ |
| 410 | numa = cpu_to_node(cpu); |
| 411 | |
Roman Gushchin | e88cc05 | 2020-12-01 13:58:35 -0800 | [diff] [blame] | 412 | rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 413 | if (!rcpu) |
| 414 | return NULL; |
| 415 | |
| 416 | /* Alloc percpu bulkq */ |
Roman Gushchin | e88cc05 | 2020-12-01 13:58:35 -0800 | [diff] [blame] | 417 | rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq), |
| 418 | sizeof(void *), gfp); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 419 | if (!rcpu->bulkq) |
| 420 | goto free_rcu; |
| 421 | |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 422 | for_each_possible_cpu(i) { |
| 423 | bq = per_cpu_ptr(rcpu->bulkq, i); |
| 424 | bq->obj = rcpu; |
| 425 | } |
| 426 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 427 | /* Alloc queue */ |
Roman Gushchin | e88cc05 | 2020-12-01 13:58:35 -0800 | [diff] [blame] | 428 | rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp, |
| 429 | numa); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 430 | if (!rcpu->queue) |
| 431 | goto free_bulkq; |
| 432 | |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 433 | err = ptr_ring_init(rcpu->queue, value->qsize, gfp); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 434 | if (err) |
| 435 | goto free_queue; |
| 436 | |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 437 | rcpu->cpu = cpu; |
Roman Gushchin | e88cc05 | 2020-12-01 13:58:35 -0800 | [diff] [blame] | 438 | rcpu->map_id = map->id; |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 439 | rcpu->value.qsize = value->qsize; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 440 | |
Lorenzo Bianconi | c576b9c | 2020-07-19 17:52:41 +0200 | [diff] [blame] | 441 | if (fd > 0 && __cpu_map_load_bpf_program(rcpu, fd)) |
| 442 | goto free_ptr_ring; |
| 443 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 444 | /* Setup kthread */ |
| 445 | rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa, |
Roman Gushchin | e88cc05 | 2020-12-01 13:58:35 -0800 | [diff] [blame] | 446 | "cpumap/%d/map:%d", cpu, |
| 447 | map->id); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 448 | if (IS_ERR(rcpu->kthread)) |
Lorenzo Bianconi | c576b9c | 2020-07-19 17:52:41 +0200 | [diff] [blame] | 449 | goto free_prog; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 450 | |
| 451 | get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */ |
| 452 | get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */ |
| 453 | |
| 454 | /* Make sure kthread runs on a single CPU */ |
| 455 | kthread_bind(rcpu->kthread, cpu); |
| 456 | wake_up_process(rcpu->kthread); |
| 457 | |
| 458 | return rcpu; |
| 459 | |
Lorenzo Bianconi | c576b9c | 2020-07-19 17:52:41 +0200 | [diff] [blame] | 460 | free_prog: |
| 461 | if (rcpu->prog) |
| 462 | bpf_prog_put(rcpu->prog); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 463 | free_ptr_ring: |
| 464 | ptr_ring_cleanup(rcpu->queue, NULL); |
| 465 | free_queue: |
| 466 | kfree(rcpu->queue); |
| 467 | free_bulkq: |
| 468 | free_percpu(rcpu->bulkq); |
| 469 | free_rcu: |
| 470 | kfree(rcpu); |
| 471 | return NULL; |
| 472 | } |
| 473 | |
Wei Yongjun | 0fe875c | 2018-01-16 11:27:05 +0000 | [diff] [blame] | 474 | static void __cpu_map_entry_free(struct rcu_head *rcu) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 475 | { |
| 476 | struct bpf_cpu_map_entry *rcpu; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 477 | |
| 478 | /* This cpu_map_entry have been disconnected from map and one |
Björn Töpel | fb5aacd | 2019-12-19 07:10:01 +0100 | [diff] [blame] | 479 | * RCU grace-period have elapsed. Thus, XDP cannot queue any |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 480 | * new packets and cannot change/set flush_needed that can |
| 481 | * find this entry. |
| 482 | */ |
| 483 | rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu); |
| 484 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 485 | free_percpu(rcpu->bulkq); |
| 486 | /* Cannot kthread_stop() here, last put free rcpu resources */ |
| 487 | put_cpu_map_entry(rcpu); |
| 488 | } |
| 489 | |
| 490 | /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to |
| 491 | * ensure any driver rcu critical sections have completed, but this |
| 492 | * does not guarantee a flush has happened yet. Because driver side |
| 493 | * rcu_read_lock/unlock only protects the running XDP program. The |
| 494 | * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a |
| 495 | * pending flush op doesn't fail. |
| 496 | * |
| 497 | * The bpf_cpu_map_entry is still used by the kthread, and there can |
| 498 | * still be pending packets (in queue and percpu bulkq). A refcnt |
| 499 | * makes sure to last user (kthread_stop vs. call_rcu) free memory |
| 500 | * resources. |
| 501 | * |
| 502 | * The rcu callback __cpu_map_entry_free flush remaining packets in |
| 503 | * percpu bulkq to queue. Due to caller map_delete_elem() disable |
| 504 | * preemption, cannot call kthread_stop() to make sure queue is empty. |
| 505 | * Instead a work_queue is started for stopping kthread, |
Björn Töpel | fb5aacd | 2019-12-19 07:10:01 +0100 | [diff] [blame] | 506 | * cpu_map_kthread_stop, which waits for an RCU grace period before |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 507 | * stopping kthread, emptying the queue. |
| 508 | */ |
Wei Yongjun | 0fe875c | 2018-01-16 11:27:05 +0000 | [diff] [blame] | 509 | static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap, |
| 510 | u32 key_cpu, struct bpf_cpu_map_entry *rcpu) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 511 | { |
| 512 | struct bpf_cpu_map_entry *old_rcpu; |
| 513 | |
| 514 | old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu); |
| 515 | if (old_rcpu) { |
| 516 | call_rcu(&old_rcpu->rcu, __cpu_map_entry_free); |
| 517 | INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop); |
| 518 | schedule_work(&old_rcpu->kthread_stop_wq); |
| 519 | } |
| 520 | } |
| 521 | |
Wei Yongjun | 0fe875c | 2018-01-16 11:27:05 +0000 | [diff] [blame] | 522 | static int cpu_map_delete_elem(struct bpf_map *map, void *key) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 523 | { |
| 524 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
| 525 | u32 key_cpu = *(u32 *)key; |
| 526 | |
| 527 | if (key_cpu >= map->max_entries) |
| 528 | return -EINVAL; |
| 529 | |
| 530 | /* notice caller map_delete_elem() use preempt_disable() */ |
| 531 | __cpu_map_entry_replace(cmap, key_cpu, NULL); |
| 532 | return 0; |
| 533 | } |
| 534 | |
Wei Yongjun | 0fe875c | 2018-01-16 11:27:05 +0000 | [diff] [blame] | 535 | static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value, |
| 536 | u64 map_flags) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 537 | { |
| 538 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 539 | struct bpf_cpumap_val cpumap_value = {}; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 540 | struct bpf_cpu_map_entry *rcpu; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 541 | /* Array index key correspond to CPU number */ |
| 542 | u32 key_cpu = *(u32 *)key; |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 543 | |
| 544 | memcpy(&cpumap_value, value, map->value_size); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 545 | |
| 546 | if (unlikely(map_flags > BPF_EXIST)) |
| 547 | return -EINVAL; |
| 548 | if (unlikely(key_cpu >= cmap->map.max_entries)) |
| 549 | return -E2BIG; |
| 550 | if (unlikely(map_flags == BPF_NOEXIST)) |
| 551 | return -EEXIST; |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 552 | if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */ |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 553 | return -EOVERFLOW; |
| 554 | |
| 555 | /* Make sure CPU is a valid possible cpu */ |
Toke Høiland-Jørgensen | bc23d0e | 2020-04-16 10:31:20 +0200 | [diff] [blame] | 556 | if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu)) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 557 | return -ENODEV; |
| 558 | |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 559 | if (cpumap_value.qsize == 0) { |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 560 | rcpu = NULL; /* Same as deleting */ |
| 561 | } else { |
| 562 | /* Updating qsize cause re-allocation of bpf_cpu_map_entry */ |
Roman Gushchin | e88cc05 | 2020-12-01 13:58:35 -0800 | [diff] [blame] | 563 | rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 564 | if (!rcpu) |
| 565 | return -ENOMEM; |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 566 | rcpu->cmap = cmap; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 567 | } |
| 568 | rcu_read_lock(); |
| 569 | __cpu_map_entry_replace(cmap, key_cpu, rcpu); |
| 570 | rcu_read_unlock(); |
| 571 | return 0; |
| 572 | } |
| 573 | |
Wei Yongjun | 0fe875c | 2018-01-16 11:27:05 +0000 | [diff] [blame] | 574 | static void cpu_map_free(struct bpf_map *map) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 575 | { |
| 576 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 577 | u32 i; |
| 578 | |
| 579 | /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0, |
| 580 | * so the bpf programs (can be more than one that used this map) were |
| 581 | * disconnected from events. Wait for outstanding critical sections in |
| 582 | * these programs to complete. The rcu critical section only guarantees |
| 583 | * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map. |
| 584 | * It does __not__ ensure pending flush operations (if any) are |
| 585 | * complete. |
| 586 | */ |
Daniel Borkmann | f6069b9 | 2018-08-17 23:26:14 +0200 | [diff] [blame] | 587 | |
| 588 | bpf_clear_redirect_map(map); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 589 | synchronize_rcu(); |
| 590 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 591 | /* For cpu_map the remote CPUs can still be using the entries |
| 592 | * (struct bpf_cpu_map_entry). |
| 593 | */ |
| 594 | for (i = 0; i < cmap->map.max_entries; i++) { |
| 595 | struct bpf_cpu_map_entry *rcpu; |
| 596 | |
| 597 | rcpu = READ_ONCE(cmap->cpu_map[i]); |
| 598 | if (!rcpu) |
| 599 | continue; |
| 600 | |
Björn Töpel | fb5aacd | 2019-12-19 07:10:01 +0100 | [diff] [blame] | 601 | /* bq flush and cleanup happens after RCU grace-period */ |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 602 | __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */ |
| 603 | } |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 604 | bpf_map_area_free(cmap->cpu_map); |
| 605 | kfree(cmap); |
| 606 | } |
| 607 | |
| 608 | struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key) |
| 609 | { |
| 610 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
| 611 | struct bpf_cpu_map_entry *rcpu; |
| 612 | |
| 613 | if (key >= map->max_entries) |
| 614 | return NULL; |
| 615 | |
| 616 | rcpu = READ_ONCE(cmap->cpu_map[key]); |
| 617 | return rcpu; |
| 618 | } |
| 619 | |
| 620 | static void *cpu_map_lookup_elem(struct bpf_map *map, void *key) |
| 621 | { |
| 622 | struct bpf_cpu_map_entry *rcpu = |
| 623 | __cpu_map_lookup_elem(map, *(u32 *)key); |
| 624 | |
Lorenzo Bianconi | 644bfe5 | 2020-07-14 15:56:37 +0200 | [diff] [blame] | 625 | return rcpu ? &rcpu->value : NULL; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 626 | } |
| 627 | |
| 628 | static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key) |
| 629 | { |
| 630 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
| 631 | u32 index = key ? *(u32 *)key : U32_MAX; |
| 632 | u32 *next = next_key; |
| 633 | |
| 634 | if (index >= cmap->map.max_entries) { |
| 635 | *next = 0; |
| 636 | return 0; |
| 637 | } |
| 638 | |
| 639 | if (index == cmap->map.max_entries - 1) |
| 640 | return -ENOENT; |
| 641 | *next = index + 1; |
| 642 | return 0; |
| 643 | } |
| 644 | |
Andrey Ignatov | 2872e9a | 2020-06-19 14:11:44 -0700 | [diff] [blame] | 645 | static int cpu_map_btf_id; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 646 | const struct bpf_map_ops cpu_map_ops = { |
Martin KaFai Lau | f4d0525 | 2020-08-27 18:18:06 -0700 | [diff] [blame] | 647 | .map_meta_equal = bpf_map_meta_equal, |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 648 | .map_alloc = cpu_map_alloc, |
| 649 | .map_free = cpu_map_free, |
| 650 | .map_delete_elem = cpu_map_delete_elem, |
| 651 | .map_update_elem = cpu_map_update_elem, |
| 652 | .map_lookup_elem = cpu_map_lookup_elem, |
| 653 | .map_get_next_key = cpu_map_get_next_key, |
Daniel Borkmann | e8d2bec | 2018-08-12 01:59:17 +0200 | [diff] [blame] | 654 | .map_check_btf = map_check_no_btf, |
Andrey Ignatov | 2872e9a | 2020-06-19 14:11:44 -0700 | [diff] [blame] | 655 | .map_btf_name = "bpf_cpu_map", |
| 656 | .map_btf_id = &cpu_map_btf_id, |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 657 | }; |
| 658 | |
Björn Töpel | ebc4ecd | 2020-09-01 10:39:28 +0200 | [diff] [blame] | 659 | static void bq_flush_to_queue(struct xdp_bulk_queue *bq) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 660 | { |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 661 | struct bpf_cpu_map_entry *rcpu = bq->obj; |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 662 | unsigned int processed = 0, drops = 0; |
| 663 | const int to_cpu = rcpu->cpu; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 664 | struct ptr_ring *q; |
| 665 | int i; |
| 666 | |
| 667 | if (unlikely(!bq->count)) |
Björn Töpel | ebc4ecd | 2020-09-01 10:39:28 +0200 | [diff] [blame] | 668 | return; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 669 | |
| 670 | q = rcpu->queue; |
| 671 | spin_lock(&q->producer_lock); |
| 672 | |
| 673 | for (i = 0; i < bq->count; i++) { |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 674 | struct xdp_frame *xdpf = bq->q[i]; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 675 | int err; |
| 676 | |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 677 | err = __ptr_ring_produce(q, xdpf); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 678 | if (err) { |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 679 | drops++; |
Björn Töpel | 4bc188c | 2019-12-19 07:10:00 +0100 | [diff] [blame] | 680 | xdp_return_frame_rx_napi(xdpf); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 681 | } |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 682 | processed++; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 683 | } |
| 684 | bq->count = 0; |
| 685 | spin_unlock(&q->producer_lock); |
| 686 | |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 687 | __list_del_clearprev(&bq->flush_node); |
| 688 | |
Jesper Dangaard Brouer | f9419f7 | 2017-10-16 12:19:44 +0200 | [diff] [blame] | 689 | /* Feedback loop via tracepoints */ |
| 690 | trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 691 | } |
| 692 | |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 693 | /* Runs under RCU-read-side, plus in softirq under NAPI protection. |
| 694 | * Thus, safe percpu variable access. |
| 695 | */ |
Björn Töpel | ebc4ecd | 2020-09-01 10:39:28 +0200 | [diff] [blame] | 696 | static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 697 | { |
Björn Töpel | cdfafe9 | 2019-12-19 07:10:04 +0100 | [diff] [blame] | 698 | struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 699 | struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq); |
| 700 | |
| 701 | if (unlikely(bq->count == CPU_MAP_BULK_SIZE)) |
Björn Töpel | 4bc188c | 2019-12-19 07:10:00 +0100 | [diff] [blame] | 702 | bq_flush_to_queue(bq); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 703 | |
| 704 | /* Notice, xdp_buff/page MUST be queued here, long enough for |
| 705 | * driver to code invoking us to finished, due to driver |
| 706 | * (e.g. ixgbe) recycle tricks based on page-refcnt. |
| 707 | * |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 708 | * Thus, incoming xdp_frame is always queued here (else we race |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 709 | * with another CPU on page-refcnt and remaining driver code). |
| 710 | * Queue time is very short, as driver will invoke flush |
| 711 | * operation, when completing napi->poll call. |
| 712 | */ |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 713 | bq->q[bq->count++] = xdpf; |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 714 | |
| 715 | if (!bq->flush_node.prev) |
| 716 | list_add(&bq->flush_node, flush_list); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 717 | } |
| 718 | |
Jesper Dangaard Brouer | 9c270af | 2017-10-16 12:19:34 +0200 | [diff] [blame] | 719 | int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp, |
| 720 | struct net_device *dev_rx) |
| 721 | { |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 722 | struct xdp_frame *xdpf; |
Jesper Dangaard Brouer | 9c270af | 2017-10-16 12:19:34 +0200 | [diff] [blame] | 723 | |
Lorenzo Bianconi | 1b698fa | 2020-05-28 22:47:29 +0200 | [diff] [blame] | 724 | xdpf = xdp_convert_buff_to_frame(xdp); |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 725 | if (unlikely(!xdpf)) |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 726 | return -EOVERFLOW; |
Jesper Dangaard Brouer | 9c270af | 2017-10-16 12:19:34 +0200 | [diff] [blame] | 727 | |
Jesper Dangaard Brouer | 1c601d8 | 2017-10-16 12:19:39 +0200 | [diff] [blame] | 728 | /* Info needed when constructing SKB on remote CPU */ |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 729 | xdpf->dev_rx = dev_rx; |
Jesper Dangaard Brouer | 9c270af | 2017-10-16 12:19:34 +0200 | [diff] [blame] | 730 | |
Jesper Dangaard Brouer | 70280ed | 2018-04-17 16:45:57 +0200 | [diff] [blame] | 731 | bq_enqueue(rcpu, xdpf); |
Jesper Dangaard Brouer | 9c270af | 2017-10-16 12:19:34 +0200 | [diff] [blame] | 732 | return 0; |
| 733 | } |
| 734 | |
Björn Töpel | cdfafe9 | 2019-12-19 07:10:04 +0100 | [diff] [blame] | 735 | void __cpu_map_flush(void) |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 736 | { |
Björn Töpel | cdfafe9 | 2019-12-19 07:10:04 +0100 | [diff] [blame] | 737 | struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list); |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 738 | struct xdp_bulk_queue *bq, *tmp; |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 739 | |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 740 | list_for_each_entry_safe(bq, tmp, flush_list, flush_node) { |
Björn Töpel | 4bc188c | 2019-12-19 07:10:00 +0100 | [diff] [blame] | 741 | bq_flush_to_queue(bq); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 742 | |
| 743 | /* If already running, costs spin_lock_irqsave + smb_mb */ |
Toke Høiland-Jørgensen | d5df283 | 2019-06-28 11:12:34 +0200 | [diff] [blame] | 744 | wake_up_process(bq->obj->kthread); |
Jesper Dangaard Brouer | 6710e11 | 2017-10-16 12:19:28 +0200 | [diff] [blame] | 745 | } |
| 746 | } |
Björn Töpel | cdfafe9 | 2019-12-19 07:10:04 +0100 | [diff] [blame] | 747 | |
| 748 | static int __init cpu_map_init(void) |
| 749 | { |
| 750 | int cpu; |
| 751 | |
| 752 | for_each_possible_cpu(cpu) |
| 753 | INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu)); |
| 754 | return 0; |
| 755 | } |
| 756 | |
| 757 | subsys_initcall(cpu_map_init); |