xref: /linux/net/smc/smc_loopback.c (revision b7e32ae6664285e156e9f0cd821e63e19798baf7)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Shared Memory Communications Direct over loopback-ism device.
4  *
5  *  Functions for loopback-ism device.
6  *
7  *  Copyright (c) 2024, Alibaba Inc.
8  *
9  *  Author: Wen Gu <guwen@linux.alibaba.com>
10  *          Tony Lu <tonylu@linux.alibaba.com>
11  *
12  */
13 
14 #include <linux/device.h>
15 #include <linux/types.h>
16 #include <net/smc.h>
17 
18 #include "smc_cdc.h"
19 #include "smc_ism.h"
20 #include "smc_loopback.h"
21 
22 #define SMC_LO_V2_CAPABLE	0x1 /* loopback-ism acts as ISMv2 */
23 #define SMC_LO_SUPPORT_NOCOPY	0x1
24 #define SMC_DMA_ADDR_INVALID	(~(dma_addr_t)0)
25 
26 static const char smc_lo_dev_name[] = "loopback-ism";
27 static struct smc_lo_dev *lo_dev;
28 
29 static void smc_lo_generate_ids(struct smc_lo_dev *ldev)
30 {
31 	struct smcd_gid *lgid = &ldev->local_gid;
32 	uuid_t uuid;
33 
34 	uuid_gen(&uuid);
35 	memcpy(&lgid->gid, &uuid, sizeof(lgid->gid));
36 	memcpy(&lgid->gid_ext, (u8 *)&uuid + sizeof(lgid->gid),
37 	       sizeof(lgid->gid_ext));
38 
39 	ldev->chid = SMC_LO_RESERVED_CHID;
40 }
41 
42 static int smc_lo_query_rgid(struct smcd_dev *smcd, struct smcd_gid *rgid,
43 			     u32 vid_valid, u32 vid)
44 {
45 	struct smc_lo_dev *ldev = smcd->priv;
46 
47 	/* rgid should be the same as lgid */
48 	if (!ldev || rgid->gid != ldev->local_gid.gid ||
49 	    rgid->gid_ext != ldev->local_gid.gid_ext)
50 		return -ENETUNREACH;
51 	return 0;
52 }
53 
54 static int smc_lo_register_dmb(struct smcd_dev *smcd, struct smcd_dmb *dmb,
55 			       void *client_priv)
56 {
57 	struct smc_lo_dmb_node *dmb_node, *tmp_node;
58 	struct smc_lo_dev *ldev = smcd->priv;
59 	struct folio *folio;
60 	int sba_idx, rc;
61 
62 	/* check space for new dmb */
63 	for_each_clear_bit(sba_idx, ldev->sba_idx_mask, SMC_LO_MAX_DMBS) {
64 		if (!test_and_set_bit(sba_idx, ldev->sba_idx_mask))
65 			break;
66 	}
67 	if (sba_idx == SMC_LO_MAX_DMBS)
68 		return -ENOSPC;
69 
70 	dmb_node = kzalloc(sizeof(*dmb_node), GFP_KERNEL);
71 	if (!dmb_node) {
72 		rc = -ENOMEM;
73 		goto err_bit;
74 	}
75 
76 	dmb_node->sba_idx = sba_idx;
77 	dmb_node->len = dmb->dmb_len;
78 
79 	/* not critical; fail under memory pressure and fallback to TCP */
80 	folio = folio_alloc(GFP_KERNEL | __GFP_NOWARN | __GFP_NOMEMALLOC |
81 			    __GFP_NORETRY | __GFP_ZERO,
82 			    get_order(dmb_node->len));
83 	if (!folio) {
84 		rc = -ENOMEM;
85 		goto err_node;
86 	}
87 	dmb_node->cpu_addr = folio_address(folio);
88 	dmb_node->dma_addr = SMC_DMA_ADDR_INVALID;
89 	refcount_set(&dmb_node->refcnt, 1);
90 
91 again:
92 	/* add new dmb into hash table */
93 	get_random_bytes(&dmb_node->token, sizeof(dmb_node->token));
94 	write_lock_bh(&ldev->dmb_ht_lock);
95 	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, dmb_node->token) {
96 		if (tmp_node->token == dmb_node->token) {
97 			write_unlock_bh(&ldev->dmb_ht_lock);
98 			goto again;
99 		}
100 	}
101 	hash_add(ldev->dmb_ht, &dmb_node->list, dmb_node->token);
102 	write_unlock_bh(&ldev->dmb_ht_lock);
103 	atomic_inc(&ldev->dmb_cnt);
104 
105 	dmb->sba_idx = dmb_node->sba_idx;
106 	dmb->dmb_tok = dmb_node->token;
107 	dmb->cpu_addr = dmb_node->cpu_addr;
108 	dmb->dma_addr = dmb_node->dma_addr;
109 	dmb->dmb_len = dmb_node->len;
110 
111 	return 0;
112 
113 err_node:
114 	kfree(dmb_node);
115 err_bit:
116 	clear_bit(sba_idx, ldev->sba_idx_mask);
117 	return rc;
118 }
119 
120 static void __smc_lo_unregister_dmb(struct smc_lo_dev *ldev,
121 				    struct smc_lo_dmb_node *dmb_node)
122 {
123 	/* remove dmb from hash table */
124 	write_lock_bh(&ldev->dmb_ht_lock);
125 	hash_del(&dmb_node->list);
126 	write_unlock_bh(&ldev->dmb_ht_lock);
127 
128 	clear_bit(dmb_node->sba_idx, ldev->sba_idx_mask);
129 	folio_put(virt_to_folio(dmb_node->cpu_addr));
130 	kfree(dmb_node);
131 
132 	if (atomic_dec_and_test(&ldev->dmb_cnt))
133 		wake_up(&ldev->ldev_release);
134 }
135 
136 static int smc_lo_unregister_dmb(struct smcd_dev *smcd, struct smcd_dmb *dmb)
137 {
138 	struct smc_lo_dmb_node *dmb_node = NULL, *tmp_node;
139 	struct smc_lo_dev *ldev = smcd->priv;
140 
141 	/* find dmb from hash table */
142 	read_lock_bh(&ldev->dmb_ht_lock);
143 	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, dmb->dmb_tok) {
144 		if (tmp_node->token == dmb->dmb_tok) {
145 			dmb_node = tmp_node;
146 			break;
147 		}
148 	}
149 	if (!dmb_node) {
150 		read_unlock_bh(&ldev->dmb_ht_lock);
151 		return -EINVAL;
152 	}
153 	read_unlock_bh(&ldev->dmb_ht_lock);
154 
155 	if (refcount_dec_and_test(&dmb_node->refcnt))
156 		__smc_lo_unregister_dmb(ldev, dmb_node);
157 	return 0;
158 }
159 
160 static int smc_lo_support_dmb_nocopy(struct smcd_dev *smcd)
161 {
162 	return SMC_LO_SUPPORT_NOCOPY;
163 }
164 
165 static int smc_lo_attach_dmb(struct smcd_dev *smcd, struct smcd_dmb *dmb)
166 {
167 	struct smc_lo_dmb_node *dmb_node = NULL, *tmp_node;
168 	struct smc_lo_dev *ldev = smcd->priv;
169 
170 	/* find dmb_node according to dmb->dmb_tok */
171 	read_lock_bh(&ldev->dmb_ht_lock);
172 	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, dmb->dmb_tok) {
173 		if (tmp_node->token == dmb->dmb_tok) {
174 			dmb_node = tmp_node;
175 			break;
176 		}
177 	}
178 	if (!dmb_node) {
179 		read_unlock_bh(&ldev->dmb_ht_lock);
180 		return -EINVAL;
181 	}
182 	read_unlock_bh(&ldev->dmb_ht_lock);
183 
184 	if (!refcount_inc_not_zero(&dmb_node->refcnt))
185 		/* the dmb is being unregistered, but has
186 		 * not been removed from the hash table.
187 		 */
188 		return -EINVAL;
189 
190 	/* provide dmb information */
191 	dmb->sba_idx = dmb_node->sba_idx;
192 	dmb->dmb_tok = dmb_node->token;
193 	dmb->cpu_addr = dmb_node->cpu_addr;
194 	dmb->dma_addr = dmb_node->dma_addr;
195 	dmb->dmb_len = dmb_node->len;
196 	return 0;
197 }
198 
199 static int smc_lo_detach_dmb(struct smcd_dev *smcd, u64 token)
200 {
201 	struct smc_lo_dmb_node *dmb_node = NULL, *tmp_node;
202 	struct smc_lo_dev *ldev = smcd->priv;
203 
204 	/* find dmb_node according to dmb->dmb_tok */
205 	read_lock_bh(&ldev->dmb_ht_lock);
206 	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, token) {
207 		if (tmp_node->token == token) {
208 			dmb_node = tmp_node;
209 			break;
210 		}
211 	}
212 	if (!dmb_node) {
213 		read_unlock_bh(&ldev->dmb_ht_lock);
214 		return -EINVAL;
215 	}
216 	read_unlock_bh(&ldev->dmb_ht_lock);
217 
218 	if (refcount_dec_and_test(&dmb_node->refcnt))
219 		__smc_lo_unregister_dmb(ldev, dmb_node);
220 	return 0;
221 }
222 
223 static int smc_lo_move_data(struct smcd_dev *smcd, u64 dmb_tok,
224 			    unsigned int idx, bool sf, unsigned int offset,
225 			    void *data, unsigned int size)
226 {
227 	struct smc_lo_dmb_node *rmb_node = NULL, *tmp_node;
228 	struct smc_lo_dev *ldev = smcd->priv;
229 	struct smc_connection *conn;
230 
231 	if (!sf)
232 		/* since sndbuf is merged with peer DMB, there is
233 		 * no need to copy data from sndbuf to peer DMB.
234 		 */
235 		return 0;
236 
237 	read_lock_bh(&ldev->dmb_ht_lock);
238 	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, dmb_tok) {
239 		if (tmp_node->token == dmb_tok) {
240 			rmb_node = tmp_node;
241 			break;
242 		}
243 	}
244 	if (!rmb_node) {
245 		read_unlock_bh(&ldev->dmb_ht_lock);
246 		return -EINVAL;
247 	}
248 	memcpy((char *)rmb_node->cpu_addr + offset, data, size);
249 	read_unlock_bh(&ldev->dmb_ht_lock);
250 
251 	conn = smcd->conn[rmb_node->sba_idx];
252 	if (!conn || conn->killed)
253 		return -EPIPE;
254 	tasklet_schedule(&conn->rx_tsklet);
255 	return 0;
256 }
257 
258 static void smc_lo_get_local_gid(struct smcd_dev *smcd,
259 				 struct smcd_gid *smcd_gid)
260 {
261 	struct smc_lo_dev *ldev = smcd->priv;
262 
263 	smcd_gid->gid = ldev->local_gid.gid;
264 	smcd_gid->gid_ext = ldev->local_gid.gid_ext;
265 }
266 
267 static u16 smc_lo_get_chid(struct smcd_dev *smcd)
268 {
269 	return ((struct smc_lo_dev *)smcd->priv)->chid;
270 }
271 
272 static struct device *smc_lo_get_dev(struct smcd_dev *smcd)
273 {
274 	return &((struct smc_lo_dev *)smcd->priv)->dev;
275 }
276 
277 static const struct smcd_ops lo_ops = {
278 	.query_remote_gid = smc_lo_query_rgid,
279 	.register_dmb = smc_lo_register_dmb,
280 	.unregister_dmb = smc_lo_unregister_dmb,
281 	.support_dmb_nocopy = smc_lo_support_dmb_nocopy,
282 	.attach_dmb = smc_lo_attach_dmb,
283 	.detach_dmb = smc_lo_detach_dmb,
284 	.add_vlan_id		= NULL,
285 	.del_vlan_id		= NULL,
286 	.set_vlan_required	= NULL,
287 	.reset_vlan_required	= NULL,
288 	.signal_event		= NULL,
289 	.move_data = smc_lo_move_data,
290 	.get_local_gid = smc_lo_get_local_gid,
291 	.get_chid = smc_lo_get_chid,
292 	.get_dev = smc_lo_get_dev,
293 };
294 
295 static struct smcd_dev *smcd_lo_alloc_dev(const struct smcd_ops *ops,
296 					  int max_dmbs)
297 {
298 	struct smcd_dev *smcd;
299 
300 	smcd = kzalloc(sizeof(*smcd), GFP_KERNEL);
301 	if (!smcd)
302 		return NULL;
303 
304 	smcd->conn = kcalloc(max_dmbs, sizeof(struct smc_connection *),
305 			     GFP_KERNEL);
306 	if (!smcd->conn)
307 		goto out_smcd;
308 
309 	smcd->ops = ops;
310 
311 	spin_lock_init(&smcd->lock);
312 	spin_lock_init(&smcd->lgr_lock);
313 	INIT_LIST_HEAD(&smcd->vlan);
314 	INIT_LIST_HEAD(&smcd->lgr_list);
315 	init_waitqueue_head(&smcd->lgrs_deleted);
316 	return smcd;
317 
318 out_smcd:
319 	kfree(smcd);
320 	return NULL;
321 }
322 
323 static int smcd_lo_register_dev(struct smc_lo_dev *ldev)
324 {
325 	struct smcd_dev *smcd;
326 
327 	smcd = smcd_lo_alloc_dev(&lo_ops, SMC_LO_MAX_DMBS);
328 	if (!smcd)
329 		return -ENOMEM;
330 	ldev->smcd = smcd;
331 	smcd->priv = ldev;
332 	smc_ism_set_v2_capable();
333 	mutex_lock(&smcd_dev_list.mutex);
334 	list_add(&smcd->list, &smcd_dev_list.list);
335 	mutex_unlock(&smcd_dev_list.mutex);
336 	pr_warn_ratelimited("smc: adding smcd device %s\n",
337 			    dev_name(&ldev->dev));
338 	return 0;
339 }
340 
341 static void smcd_lo_unregister_dev(struct smc_lo_dev *ldev)
342 {
343 	struct smcd_dev *smcd = ldev->smcd;
344 
345 	pr_warn_ratelimited("smc: removing smcd device %s\n",
346 			    dev_name(&ldev->dev));
347 	smcd->going_away = 1;
348 	smc_smcd_terminate_all(smcd);
349 	mutex_lock(&smcd_dev_list.mutex);
350 	list_del_init(&smcd->list);
351 	mutex_unlock(&smcd_dev_list.mutex);
352 	kfree(smcd->conn);
353 	kfree(smcd);
354 }
355 
356 static int smc_lo_dev_init(struct smc_lo_dev *ldev)
357 {
358 	smc_lo_generate_ids(ldev);
359 	rwlock_init(&ldev->dmb_ht_lock);
360 	hash_init(ldev->dmb_ht);
361 	atomic_set(&ldev->dmb_cnt, 0);
362 	init_waitqueue_head(&ldev->ldev_release);
363 
364 	return smcd_lo_register_dev(ldev);
365 }
366 
367 static void smc_lo_dev_exit(struct smc_lo_dev *ldev)
368 {
369 	smcd_lo_unregister_dev(ldev);
370 	if (atomic_read(&ldev->dmb_cnt))
371 		wait_event(ldev->ldev_release, !atomic_read(&ldev->dmb_cnt));
372 }
373 
374 static void smc_lo_dev_release(struct device *dev)
375 {
376 	struct smc_lo_dev *ldev =
377 		container_of(dev, struct smc_lo_dev, dev);
378 
379 	kfree(ldev);
380 }
381 
382 static int smc_lo_dev_probe(void)
383 {
384 	struct smc_lo_dev *ldev;
385 	int ret;
386 
387 	ldev = kzalloc(sizeof(*ldev), GFP_KERNEL);
388 	if (!ldev)
389 		return -ENOMEM;
390 
391 	ldev->dev.parent = NULL;
392 	ldev->dev.release = smc_lo_dev_release;
393 	device_initialize(&ldev->dev);
394 	dev_set_name(&ldev->dev, smc_lo_dev_name);
395 
396 	ret = smc_lo_dev_init(ldev);
397 	if (ret)
398 		goto free_dev;
399 
400 	lo_dev = ldev; /* global loopback device */
401 	return 0;
402 
403 free_dev:
404 	put_device(&ldev->dev);
405 	return ret;
406 }
407 
408 static void smc_lo_dev_remove(void)
409 {
410 	if (!lo_dev)
411 		return;
412 
413 	smc_lo_dev_exit(lo_dev);
414 	put_device(&lo_dev->dev); /* device_initialize in smc_lo_dev_probe */
415 }
416 
417 int smc_loopback_init(void)
418 {
419 	return smc_lo_dev_probe();
420 }
421 
422 void smc_loopback_exit(void)
423 {
424 	smc_lo_dev_remove();
425 }
426