forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2024-05-10 9999e48639b3cecb08ffb37358bcba3b48161b29 
 kernel/kernel/bpf/cpumap.c
....@@ -1,7 +1,7 @@
1
+// SPDX-License-Identifier: GPL-2.0-only
12 /* bpf/cpumap.c
23  *
34  * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
4
- * Released under terms in GPL version 2.  See COPYING.
55  */
66 
77 /* The 'cpumap' is primarily used as a backend map for XDP BPF helper
....@@ -32,14 +32,19 @@
3232 
3333 /* General idea: XDP packets getting XDP redirected to another CPU,
3434  * will maximum be stored/queued for one driver ->poll() call.  It is
35
- * guaranteed that setting flush bit and flush operation happen on
35
+ * guaranteed that queueing the frame and the flush operation happen on
3636  * same CPU.  Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
3737  * which queue in bpf_cpu_map_entry contains packets.
3838  */
3939 
4040 #define CPU_MAP_BULK_SIZE 8  /* 8 == one cacheline on 64-bit archs */
41
+struct bpf_cpu_map_entry;
42
+struct bpf_cpu_map;
43
+
4144 struct xdp_bulk_queue {
4245 	void *q[CPU_MAP_BULK_SIZE];
46
+	struct list_head flush_node;
47
+	struct bpf_cpu_map_entry *obj;
4348 	unsigned int count;
4449 };
4550 
....@@ -47,48 +52,49 @@
4752 struct bpf_cpu_map_entry {
4853 	u32 cpu;    /* kthread CPU and map index */
4954 	int map_id; /* Back reference to map */
50
-	u32 qsize;  /* Queue size placeholder for map lookup */
5155 
5256 	/* XDP can run multiple RX-ring queues, need __percpu enqueue store */
5357 	struct xdp_bulk_queue __percpu *bulkq;
5458 
59
+	struct bpf_cpu_map *cmap;
60
+
5561 	/* Queue with potential multi-producers, and single-consumer kthread */
5662 	struct ptr_ring *queue;
5763 	struct task_struct *kthread;
58
-	struct work_struct kthread_stop_wq;
64
+
65
+	struct bpf_cpumap_val value;
66
+	struct bpf_prog *prog;
5967 
6068 	atomic_t refcnt; /* Control when this struct can be free'ed */
6169 	struct rcu_head rcu;
70
+
71
+	struct work_struct kthread_stop_wq;
6272 };
6373 
6474 struct bpf_cpu_map {
6575 	struct bpf_map map;
6676 	/* Below members specific for map type */
6777 	struct bpf_cpu_map_entry **cpu_map;
68
-	unsigned long __percpu *flush_needed;
6978 };
7079 
71
-static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
72
-			     struct xdp_bulk_queue *bq, bool in_napi_ctx);
73
-
74
-static u64 cpu_map_bitmap_size(const union bpf_attr *attr)
75
-{
76
-	return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
77
-}
80
+static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
7881 
7982 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
8083 {
84
+	u32 value_size = attr->value_size;
8185 	struct bpf_cpu_map *cmap;
8286 	int err = -ENOMEM;
8387 	u64 cost;
8488 	int ret;
8589 
86
-	if (!capable(CAP_SYS_ADMIN))
90
+	if (!bpf_capable())
8791 		return ERR_PTR(-EPERM);
8892 
8993 	/* check sanity of attributes */
9094 	if (attr->max_entries == 0 || attr->key_size != 4 ||
91
-	    attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
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)
9298 		return ERR_PTR(-EINVAL);
9399 
94100 	cmap = kzalloc(sizeof(*cmap), GFP_USER);
....@@ -105,34 +111,24 @@
105111 
106112 	/* make sure page count doesn't overflow */
107113 	cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
108
-	cost += cpu_map_bitmap_size(attr) * num_possible_cpus();
109
-	if (cost >= U32_MAX - PAGE_SIZE)
110
-		goto free_cmap;
111
-	cmap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
112114 
113115 	/* Notice returns -EPERM on if map size is larger than memlock limit */
114
-	ret = bpf_map_precharge_memlock(cmap->map.pages);
116
+	ret = bpf_map_charge_init(&cmap->map.memory, cost);
115117 	if (ret) {
116118 		err = ret;
117119 		goto free_cmap;
118120 	}
119
-
120
-	/* A per cpu bitfield with a bit per possible CPU in map  */
121
-	cmap->flush_needed = __alloc_percpu(cpu_map_bitmap_size(attr),
122
-					    __alignof__(unsigned long));
123
-	if (!cmap->flush_needed)
124
-		goto free_cmap;
125121 
126122 	/* Alloc array for possible remote "destination" CPUs */
127123 	cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
128124 					   sizeof(struct bpf_cpu_map_entry *),
129125 					   cmap->map.numa_node);
130126 	if (!cmap->cpu_map)
131
-		goto free_percpu;
127
+		goto free_charge;
132128 
133129 	return &cmap->map;
134
-free_percpu:
135
-	free_percpu(cmap->flush_needed);
130
+free_charge:
131
+	bpf_map_charge_finish(&cmap->map.memory);
136132 free_cmap:
137133 	kfree(cmap);
138134 	return ERR_PTR(err);
....@@ -159,40 +155,24 @@
159155 	kthread_stop(rcpu->kthread);
160156 }
161157 
162
-static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
163
-					 struct xdp_frame *xdpf)
158
+static struct sk_buff *cpu_map_build_skb(struct xdp_frame *xdpf,
159
+					 struct sk_buff *skb)
164160 {
165161 	unsigned int hard_start_headroom;
166162 	unsigned int frame_size;
167163 	void *pkt_data_start;
168
-	struct sk_buff *skb;
169164 
170165 	/* Part of headroom was reserved to xdpf */
171166 	hard_start_headroom = sizeof(struct xdp_frame) +  xdpf->headroom;
172167 
173
-	/* build_skb need to place skb_shared_info after SKB end, and
174
-	 * also want to know the memory "truesize".  Thus, need to
175
-	 * know the memory frame size backing xdp_buff.
176
-	 *
177
-	 * XDP was designed to have PAGE_SIZE frames, but this
178
-	 * assumption is not longer true with ixgbe and i40e.  It
179
-	 * would be preferred to set frame_size to 2048 or 4096
180
-	 * depending on the driver.
181
-	 *   frame_size = 2048;
182
-	 *   frame_len  = frame_size - sizeof(*xdp_frame);
183
-	 *
184
-	 * Instead, with info avail, skb_shared_info in placed after
185
-	 * packet len.  This, unfortunately fakes the truesize.
186
-	 * Another disadvantage of this approach, the skb_shared_info
187
-	 * is not at a fixed memory location, with mixed length
188
-	 * packets, which is bad for cache-line hotness.
168
+	/* Memory size backing xdp_frame data already have reserved
169
+	 * room for build_skb to place skb_shared_info in tailroom.
189170 	 */
190
-	frame_size = SKB_DATA_ALIGN(xdpf->len + hard_start_headroom) +
191
-		SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
171
+	frame_size = xdpf->frame_sz;
192172 
193173 	pkt_data_start = xdpf->data - hard_start_headroom;
194
-	skb = build_skb(pkt_data_start, frame_size);
195
-	if (!skb)
174
+	skb = build_skb_around(skb, pkt_data_start, frame_size);
175
+	if (unlikely(!skb))
196176 		return NULL;
197177 
198178 	skb_reserve(skb, hard_start_headroom);
....@@ -208,6 +188,9 @@
208188 	 * - HW RX hash			(skb_set_hash)
209189 	 * - RX ring dev queue index	(skb_record_rx_queue)
210190 	 */
191
+
192
+	/* Until page_pool get SKB return path, release DMA here */
193
+	xdp_release_frame(xdpf);
211194 
212195 	/* Allow SKB to reuse area used by xdp_frame */
213196 	xdp_scrub_frame(xdpf);
....@@ -232,6 +215,8 @@
232215 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
233216 {
234217 	if (atomic_dec_and_test(&rcpu->refcnt)) {
218
+		if (rcpu->prog)
219
+			bpf_prog_put(rcpu->prog);
235220 		/* The queue should be empty at this point */
236221 		__cpu_map_ring_cleanup(rcpu->queue);
237222 		ptr_ring_cleanup(rcpu->queue, NULL);
....@@ -239,6 +224,77 @@
239224 		kfree(rcpu);
240225 	}
241226 }
227
+
228
+static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
229
+				    void **frames, int n,
230
+				    struct xdp_cpumap_stats *stats)
231
+{
232
+	struct xdp_rxq_info rxq;
233
+	struct xdp_buff xdp;
234
+	int i, nframes = 0;
235
+
236
+	if (!rcpu->prog)
237
+		return n;
238
+
239
+	rcu_read_lock_bh();
240
+
241
+	xdp_set_return_frame_no_direct();
242
+	xdp.rxq = &rxq;
243
+
244
+	for (i = 0; i < n; i++) {
245
+		struct xdp_frame *xdpf = frames[i];
246
+		u32 act;
247
+		int err;
248
+
249
+		rxq.dev = xdpf->dev_rx;
250
+		rxq.mem = xdpf->mem;
251
+		/* TODO: report queue_index to xdp_rxq_info */
252
+
253
+		xdp_convert_frame_to_buff(xdpf, &xdp);
254
+
255
+		act = bpf_prog_run_xdp(rcpu->prog, &xdp);
256
+		switch (act) {
257
+		case XDP_PASS:
258
+			err = xdp_update_frame_from_buff(&xdp, xdpf);
259
+			if (err < 0) {
260
+				xdp_return_frame(xdpf);
261
+				stats->drop++;
262
+			} else {
263
+				frames[nframes++] = xdpf;
264
+				stats->pass++;
265
+			}
266
+			break;
267
+		case XDP_REDIRECT:
268
+			err = xdp_do_redirect(xdpf->dev_rx, &xdp,
269
+					      rcpu->prog);
270
+			if (unlikely(err)) {
271
+				xdp_return_frame(xdpf);
272
+				stats->drop++;
273
+			} else {
274
+				stats->redirect++;
275
+			}
276
+			break;
277
+		default:
278
+			bpf_warn_invalid_xdp_action(act);
279
+			fallthrough;
280
+		case XDP_DROP:
281
+			xdp_return_frame(xdpf);
282
+			stats->drop++;
283
+			break;
284
+		}
285
+	}
286
+
287
+	if (stats->redirect)
288
+		xdp_do_flush_map();
289
+
290
+	xdp_clear_return_frame_no_direct();
291
+
292
+	rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
293
+
294
+	return nframes;
295
+}
296
+
297
+#define CPUMAP_BATCH 8
242298 
243299 static int cpu_map_kthread_run(void *data)
244300 {
....@@ -252,8 +308,12 @@
252308 	 * kthread_stop signal until queue is empty.
253309 	 */
254310 	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
255
-		unsigned int processed = 0, drops = 0, sched = 0;
256
-		struct xdp_frame *xdpf;
311
+		struct xdp_cpumap_stats stats = {}; /* zero stats */
312
+		gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
313
+		unsigned int drops = 0, sched = 0;
314
+		void *frames[CPUMAP_BATCH];
315
+		void *skbs[CPUMAP_BATCH];
316
+		int i, n, m, nframes;
257317 
258318 		/* Release CPU reschedule checks */
259319 		if (__ptr_ring_empty(rcpu->queue)) {
....@@ -269,18 +329,42 @@
269329 			sched = cond_resched();
270330 		}
271331 
272
-		/* Process packets in rcpu->queue */
273
-		local_bh_disable();
274332 		/*
275333 		 * The bpf_cpu_map_entry is single consumer, with this
276334 		 * kthread CPU pinned. Lockless access to ptr_ring
277335 		 * consume side valid as no-resize allowed of queue.
278336 		 */
279
-		while ((xdpf = __ptr_ring_consume(rcpu->queue))) {
280
-			struct sk_buff *skb;
337
+		n = __ptr_ring_consume_batched(rcpu->queue, frames,
338
+					       CPUMAP_BATCH);
339
+		for (i = 0; i < n; i++) {
340
+			void *f = frames[i];
341
+			struct page *page = virt_to_page(f);
342
+
343
+			/* Bring struct page memory area to curr CPU. Read by
344
+			 * build_skb_around via page_is_pfmemalloc(), and when
345
+			 * freed written by page_frag_free call.
346
+			 */
347
+			prefetchw(page);
348
+		}
349
+
350
+		/* Support running another XDP prog on this CPU */
351
+		nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, n, &stats);
352
+		if (nframes) {
353
+			m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
354
+			if (unlikely(m == 0)) {
355
+				for (i = 0; i < nframes; i++)
356
+					skbs[i] = NULL; /* effect: xdp_return_frame */
357
+				drops += nframes;
358
+			}
359
+		}
360
+
361
+		local_bh_disable();
362
+		for (i = 0; i < nframes; i++) {
363
+			struct xdp_frame *xdpf = frames[i];
364
+			struct sk_buff *skb = skbs[i];
281365 			int ret;
282366 
283
-			skb = cpu_map_build_skb(rcpu, xdpf);
367
+			skb = cpu_map_build_skb(xdpf, skb);
284368 			if (!skb) {
285369 				xdp_return_frame(xdpf);
286370 				continue;
....@@ -290,13 +374,9 @@
290374 			ret = netif_receive_skb_core(skb);
291375 			if (ret == NET_RX_DROP)
292376 				drops++;
293
-
294
-			/* Limit BH-disable period */
295
-			if (++processed == 8)
296
-				break;
297377 		}
298378 		/* Feedback loop via tracepoint */
299
-		trace_xdp_cpumap_kthread(rcpu->map_id, processed, drops, sched);
379
+		trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched, &stats);
300380 
301381 		local_bh_enable(); /* resched point, may call do_softirq() */
302382 	}
....@@ -306,12 +386,38 @@
306386 	return 0;
307387 }
308388 
309
-static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu,
310
-						       int map_id)
389
+bool cpu_map_prog_allowed(struct bpf_map *map)
311390 {
391
+	return map->map_type == BPF_MAP_TYPE_CPUMAP &&
392
+	       map->value_size != offsetofend(struct bpf_cpumap_val, qsize);
393
+}
394
+
395
+static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, int fd)
396
+{
397
+	struct bpf_prog *prog;
398
+
399
+	prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
400
+	if (IS_ERR(prog))
401
+		return PTR_ERR(prog);
402
+
403
+	if (prog->expected_attach_type != BPF_XDP_CPUMAP) {
404
+		bpf_prog_put(prog);
405
+		return -EINVAL;
406
+	}
407
+
408
+	rcpu->value.bpf_prog.id = prog->aux->id;
409
+	rcpu->prog = prog;
410
+
411
+	return 0;
412
+}
413
+
414
+static struct bpf_cpu_map_entry *
415
+__cpu_map_entry_alloc(struct bpf_cpumap_val *value, u32 cpu, int map_id)
416
+{
417
+	int numa, err, i, fd = value->bpf_prog.fd;
312418 	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
313419 	struct bpf_cpu_map_entry *rcpu;
314
-	int numa, err;
420
+	struct xdp_bulk_queue *bq;
315421 
316422 	/* Have map->numa_node, but choose node of redirect target CPU */
317423 	numa = cpu_to_node(cpu);
....@@ -326,24 +432,32 @@
326432 	if (!rcpu->bulkq)
327433 		goto free_rcu;
328434 
435
+	for_each_possible_cpu(i) {
436
+		bq = per_cpu_ptr(rcpu->bulkq, i);
437
+		bq->obj = rcpu;
438
+	}
439
+
329440 	/* Alloc queue */
330441 	rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
331442 	if (!rcpu->queue)
332443 		goto free_bulkq;
333444 
334
-	err = ptr_ring_init(rcpu->queue, qsize, gfp);
445
+	err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
335446 	if (err)
336447 		goto free_queue;
337448 
338449 	rcpu->cpu    = cpu;
339450 	rcpu->map_id = map_id;
340
-	rcpu->qsize  = qsize;
451
+	rcpu->value.qsize  = value->qsize;
452
+
453
+	if (fd > 0 && __cpu_map_load_bpf_program(rcpu, fd))
454
+		goto free_ptr_ring;
341455 
342456 	/* Setup kthread */
343457 	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
344458 					       "cpumap/%d/map:%d", cpu, map_id);
345459 	if (IS_ERR(rcpu->kthread))
346
-		goto free_ptr_ring;
460
+		goto free_prog;
347461 
348462 	get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
349463 	get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
....@@ -354,6 +468,9 @@
354468 
355469 	return rcpu;
356470 
471
+free_prog:
472
+	if (rcpu->prog)
473
+		bpf_prog_put(rcpu->prog);
357474 free_ptr_ring:
358475 	ptr_ring_cleanup(rcpu->queue, NULL);
359476 free_queue:
....@@ -368,22 +485,14 @@
368485 static void __cpu_map_entry_free(struct rcu_head *rcu)
369486 {
370487 	struct bpf_cpu_map_entry *rcpu;
371
-	int cpu;
372488 
373489 	/* This cpu_map_entry have been disconnected from map and one
374
-	 * RCU graze-period have elapsed.  Thus, XDP cannot queue any
490
+	 * RCU grace-period have elapsed.  Thus, XDP cannot queue any
375491 	 * new packets and cannot change/set flush_needed that can
376492 	 * find this entry.
377493 	 */
378494 	rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
379495 
380
-	/* Flush remaining packets in percpu bulkq */
381
-	for_each_online_cpu(cpu) {
382
-		struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu);
383
-
384
-		/* No concurrent bq_enqueue can run at this point */
385
-		bq_flush_to_queue(rcpu, bq, false);
386
-	}
387496 	free_percpu(rcpu->bulkq);
388497 	/* Cannot kthread_stop() here, last put free rcpu resources */
389498 	put_cpu_map_entry(rcpu);
....@@ -405,7 +514,7 @@
405514  * percpu bulkq to queue.  Due to caller map_delete_elem() disable
406515  * preemption, cannot call kthread_stop() to make sure queue is empty.
407516  * Instead a work_queue is started for stopping kthread,
408
- * cpu_map_kthread_stop, which waits for an RCU graze period before
517
+ * cpu_map_kthread_stop, which waits for an RCU grace period before
409518  * stopping kthread, emptying the queue.
410519  */
411520 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
....@@ -438,12 +547,12 @@
438547 			       u64 map_flags)
439548 {
440549 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
550
+	struct bpf_cpumap_val cpumap_value = {};
441551 	struct bpf_cpu_map_entry *rcpu;
442
-
443552 	/* Array index key correspond to CPU number */
444553 	u32 key_cpu = *(u32 *)key;
445
-	/* Value is the queue size */
446
-	u32 qsize = *(u32 *)value;
554
+
555
+	memcpy(&cpumap_value, value, map->value_size);
447556 
448557 	if (unlikely(map_flags > BPF_EXIST))
449558 		return -EINVAL;
....@@ -451,20 +560,21 @@
451560 		return -E2BIG;
452561 	if (unlikely(map_flags == BPF_NOEXIST))
453562 		return -EEXIST;
454
-	if (unlikely(qsize > 16384)) /* sanity limit on qsize */
563
+	if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
455564 		return -EOVERFLOW;
456565 
457566 	/* Make sure CPU is a valid possible cpu */
458567 	if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
459568 		return -ENODEV;
460569 
461
-	if (qsize == 0) {
570
+	if (cpumap_value.qsize == 0) {
462571 		rcpu = NULL; /* Same as deleting */
463572 	} else {
464573 		/* Updating qsize cause re-allocation of bpf_cpu_map_entry */
465
-		rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
574
+		rcpu = __cpu_map_entry_alloc(&cpumap_value, key_cpu, map->id);
466575 		if (!rcpu)
467576 			return -ENOMEM;
577
+		rcpu->cmap = cmap;
468578 	}
469579 	rcu_read_lock();
470580 	__cpu_map_entry_replace(cmap, key_cpu, rcpu);
....@@ -475,7 +585,6 @@
475585 static void cpu_map_free(struct bpf_map *map)
476586 {
477587 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
478
-	int cpu;
479588 	u32 i;
480589 
481590 	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
....@@ -490,18 +599,6 @@
490599 	bpf_clear_redirect_map(map);
491600 	synchronize_rcu();
492601 
493
-	/* To ensure all pending flush operations have completed wait for flush
494
-	 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
495
-	 * Because the above synchronize_rcu() ensures the map is disconnected
496
-	 * from the program we can assume no new bits will be set.
497
-	 */
498
-	for_each_online_cpu(cpu) {
499
-		unsigned long *bitmap = per_cpu_ptr(cmap->flush_needed, cpu);
500
-
501
-		while (!bitmap_empty(bitmap, cmap->map.max_entries))
502
-			cond_resched();
503
-	}
504
-
505602 	/* For cpu_map the remote CPUs can still be using the entries
506603 	 * (struct bpf_cpu_map_entry).
507604 	 */
....@@ -512,10 +609,9 @@
512609 		if (!rcpu)
513610 			continue;
514611 
515
-		/* bq flush and cleanup happens after RCU graze-period */
612
+		/* bq flush and cleanup happens after RCU grace-period */
516613 		__cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
517614 	}
518
-	free_percpu(cmap->flush_needed);
519615 	bpf_map_area_free(cmap->cpu_map);
520616 	kfree(cmap);
521617 }
....@@ -537,7 +633,7 @@
537633 	struct bpf_cpu_map_entry *rcpu =
538634 		__cpu_map_lookup_elem(map, *(u32 *)key);
539635 
540
-	return rcpu ? &rcpu->qsize : NULL;
636
+	return rcpu ? &rcpu->value : NULL;
541637 }
542638 
543639 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
....@@ -557,7 +653,9 @@
557653 	return 0;
558654 }
559655 
656
+static int cpu_map_btf_id;
560657 const struct bpf_map_ops cpu_map_ops = {
658
+	.map_meta_equal		= bpf_map_meta_equal,
561659 	.map_alloc		= cpu_map_alloc,
562660 	.map_free		= cpu_map_free,
563661 	.map_delete_elem	= cpu_map_delete_elem,
....@@ -565,18 +663,20 @@
565663 	.map_lookup_elem	= cpu_map_lookup_elem,
566664 	.map_get_next_key	= cpu_map_get_next_key,
567665 	.map_check_btf		= map_check_no_btf,
666
+	.map_btf_name		= "bpf_cpu_map",
667
+	.map_btf_id		= &cpu_map_btf_id,
568668 };
569669 
570
-static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
571
-			     struct xdp_bulk_queue *bq, bool in_napi_ctx)
670
+static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
572671 {
672
+	struct bpf_cpu_map_entry *rcpu = bq->obj;
573673 	unsigned int processed = 0, drops = 0;
574674 	const int to_cpu = rcpu->cpu;
575675 	struct ptr_ring *q;
576676 	int i;
577677 
578678 	if (unlikely(!bq->count))
579
-		return 0;
679
+		return;
580680 
581681 	q = rcpu->queue;
582682 	spin_lock(&q->producer_lock);
....@@ -588,30 +688,29 @@
588688 		err = __ptr_ring_produce(q, xdpf);
589689 		if (err) {
590690 			drops++;
591
-			if (likely(in_napi_ctx))
592
-				xdp_return_frame_rx_napi(xdpf);
593
-			else
594
-				xdp_return_frame(xdpf);
691
+			xdp_return_frame_rx_napi(xdpf);
595692 		}
596693 		processed++;
597694 	}
598695 	bq->count = 0;
599696 	spin_unlock(&q->producer_lock);
600697 
698
+	__list_del_clearprev(&bq->flush_node);
699
+
601700 	/* Feedback loop via tracepoints */
602701 	trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
603
-	return 0;
604702 }
605703 
606704 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
607705  * Thus, safe percpu variable access.
608706  */
609
-static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
707
+static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
610708 {
709
+	struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
611710 	struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
612711 
613712 	if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
614
-		bq_flush_to_queue(rcpu, bq, true);
713
+		bq_flush_to_queue(bq);
615714 
616715 	/* Notice, xdp_buff/page MUST be queued here, long enough for
617716 	 * driver to code invoking us to finished, due to driver
....@@ -623,7 +722,9 @@
623722 	 * operation, when completing napi->poll call.
624723 	 */
625724 	bq->q[bq->count++] = xdpf;
626
-	return 0;
725
+
726
+	if (!bq->flush_node.prev)
727
+		list_add(&bq->flush_node, flush_list);
627728 }
628729 
629730 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
....@@ -631,7 +732,7 @@
631732 {
632733 	struct xdp_frame *xdpf;
633734 
634
-	xdpf = convert_to_xdp_frame(xdp);
735
+	xdpf = xdp_convert_buff_to_frame(xdp);
635736 	if (unlikely(!xdpf))
636737 		return -EOVERFLOW;
637738 
....@@ -642,41 +743,26 @@
642743 	return 0;
643744 }
644745 
645
-void __cpu_map_insert_ctx(struct bpf_map *map, u32 bit)
746
+void __cpu_map_flush(void)
646747 {
647
-	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
648
-	unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
748
+	struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
749
+	struct xdp_bulk_queue *bq, *tmp;
649750 
650
-	__set_bit(bit, bitmap);
651
-}
652
-
653
-void __cpu_map_flush(struct bpf_map *map)
654
-{
655
-	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
656
-	unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
657
-	u32 bit;
658
-
659
-	/* The napi->poll softirq makes sure __cpu_map_insert_ctx()
660
-	 * and __cpu_map_flush() happen on same CPU. Thus, the percpu
661
-	 * bitmap indicate which percpu bulkq have packets.
662
-	 */
663
-	for_each_set_bit(bit, bitmap, map->max_entries) {
664
-		struct bpf_cpu_map_entry *rcpu = READ_ONCE(cmap->cpu_map[bit]);
665
-		struct xdp_bulk_queue *bq;
666
-
667
-		/* This is possible if entry is removed by user space
668
-		 * between xdp redirect and flush op.
669
-		 */
670
-		if (unlikely(!rcpu))
671
-			continue;
672
-
673
-		__clear_bit(bit, bitmap);
674
-
675
-		/* Flush all frames in bulkq to real queue */
676
-		bq = this_cpu_ptr(rcpu->bulkq);
677
-		bq_flush_to_queue(rcpu, bq, true);
751
+	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
752
+		bq_flush_to_queue(bq);
678753 
679754 		/* If already running, costs spin_lock_irqsave + smb_mb */
680
-		wake_up_process(rcpu->kthread);
755
+		wake_up_process(bq->obj->kthread);
681756 	}
682757 }
758
+
759
+static int __init cpu_map_init(void)
760
+{
761
+	int cpu;
762
+
763
+	for_each_possible_cpu(cpu)
764
+		INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
765
+	return 0;
766
+}
767
+
768
+subsys_initcall(cpu_map_init);