dibs_loopback.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 *  Functions for dibs loopback/loopback-ism device.
  4 *
  5 *  Copyright (c) 2024, Alibaba Inc.
  6 *
  7 *  Author: Wen Gu <guwen@linux.alibaba.com>
  8 *          Tony Lu <tonylu@linux.alibaba.com>
  9 *
 10 */
 11
 12#include <linux/bitops.h>
 13#include <linux/device.h>
 14#include <linux/dibs.h>
 15#include <linux/mm.h>
 16#include <linux/slab.h>
 17#include <linux/spinlock.h>
 18#include <linux/types.h>
 19
 20#include "dibs_loopback.h"
 21
 22#define DIBS_LO_SUPPORT_NOCOPY	0x1
 23#define DIBS_DMA_ADDR_INVALID	(~(dma_addr_t)0)
 24
 25static const char dibs_lo_dev_name[] = "lo";
 26/* global loopback device */
 27static struct dibs_lo_dev *lo_dev;
 28
 29static u16 dibs_lo_get_fabric_id(struct dibs_dev *dibs)
 30{
 31	return DIBS_LOOPBACK_FABRIC;
 32}
 33
 34static int dibs_lo_query_rgid(struct dibs_dev *dibs, const uuid_t *rgid,
 35			      u32 vid_valid, u32 vid)
 36{
 37	/* rgid should be the same as lgid */
 38	if (!uuid_equal(rgid, &dibs->gid))
 39		return -ENETUNREACH;
 40	return 0;
 41}
 42
 43static int dibs_lo_max_dmbs(void)
 44{
 45	return DIBS_LO_MAX_DMBS;
 46}
 47
 48static int dibs_lo_register_dmb(struct dibs_dev *dibs, struct dibs_dmb *dmb,
 49				struct dibs_client *client)
 50{
 51	struct dibs_lo_dmb_node *dmb_node, *tmp_node;
 52	struct dibs_lo_dev *ldev;
 53	struct folio *folio;
 54	unsigned long flags;
 55	int sba_idx, rc;
 56
 57	ldev = dibs->drv_priv;
 58	sba_idx = dmb->idx;
 59	/* check space for new dmb */
 60	for_each_clear_bit(sba_idx, ldev->sba_idx_mask, DIBS_LO_MAX_DMBS) {
 61		if (!test_and_set_bit(sba_idx, ldev->sba_idx_mask))
 62			break;
 63	}
 64	if (sba_idx == DIBS_LO_MAX_DMBS)
 65		return -ENOSPC;
 66
 67	dmb_node = kzalloc(sizeof(*dmb_node), GFP_KERNEL);
 68	if (!dmb_node) {
 69		rc = -ENOMEM;
 70		goto err_bit;
 71	}
 72
 73	dmb_node->sba_idx = sba_idx;
 74	dmb_node->len = dmb->dmb_len;
 75
 76	/* not critical; fail under memory pressure and fallback to TCP */
 77	folio = folio_alloc(GFP_KERNEL | __GFP_NOWARN | __GFP_NOMEMALLOC |
 78			    __GFP_NORETRY | __GFP_ZERO,
 79			    get_order(dmb_node->len));
 80	if (!folio) {
 81		rc = -ENOMEM;
 82		goto err_node;
 83	}
 84	dmb_node->cpu_addr = folio_address(folio);
 85	dmb_node->dma_addr = DIBS_DMA_ADDR_INVALID;
 86	refcount_set(&dmb_node->refcnt, 1);
 87
 88again:
 89	/* add new dmb into hash table */
 90	get_random_bytes(&dmb_node->token, sizeof(dmb_node->token));
 91	write_lock_bh(&ldev->dmb_ht_lock);
 92	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, dmb_node->token) {
 93		if (tmp_node->token == dmb_node->token) {
 94			write_unlock_bh(&ldev->dmb_ht_lock);
 95			goto again;
 96		}
 97	}
 98	hash_add(ldev->dmb_ht, &dmb_node->list, dmb_node->token);
 99	write_unlock_bh(&ldev->dmb_ht_lock);
100	atomic_inc(&ldev->dmb_cnt);
101
102	dmb->idx = dmb_node->sba_idx;
103	dmb->dmb_tok = dmb_node->token;
104	dmb->cpu_addr = dmb_node->cpu_addr;
105	dmb->dma_addr = dmb_node->dma_addr;
106	dmb->dmb_len = dmb_node->len;
107
108	spin_lock_irqsave(&dibs->lock, flags);
109	dibs->dmb_clientid_arr[sba_idx] = client->id;
110	spin_unlock_irqrestore(&dibs->lock, flags);
111
112	return 0;
113
114err_node:
115	kfree(dmb_node);
116err_bit:
117	clear_bit(sba_idx, ldev->sba_idx_mask);
118	return rc;
119}
120
121static void __dibs_lo_unregister_dmb(struct dibs_lo_dev *ldev,
122				     struct dibs_lo_dmb_node *dmb_node)
123{
124	/* remove dmb from hash table */
125	write_lock_bh(&ldev->dmb_ht_lock);
126	hash_del(&dmb_node->list);
127	write_unlock_bh(&ldev->dmb_ht_lock);
128
129	clear_bit(dmb_node->sba_idx, ldev->sba_idx_mask);
130	folio_put(virt_to_folio(dmb_node->cpu_addr));
131	kfree(dmb_node);
132
133	if (atomic_dec_and_test(&ldev->dmb_cnt))
134		wake_up(&ldev->ldev_release);
135}
136
137static int dibs_lo_unregister_dmb(struct dibs_dev *dibs, struct dibs_dmb *dmb)
138{
139	struct dibs_lo_dmb_node *dmb_node = NULL, *tmp_node;
140	struct dibs_lo_dev *ldev;
141	unsigned long flags;
142
143	ldev = dibs->drv_priv;
144
145	/* find dmb from hash table */
146	read_lock_bh(&ldev->dmb_ht_lock);
147	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, dmb->dmb_tok) {
148		if (tmp_node->token == dmb->dmb_tok) {
149			dmb_node = tmp_node;
150			break;
151		}
152	}
153	read_unlock_bh(&ldev->dmb_ht_lock);
154	if (!dmb_node)
155		return -EINVAL;
156
157	if (refcount_dec_and_test(&dmb_node->refcnt)) {
158		spin_lock_irqsave(&dibs->lock, flags);
159		dibs->dmb_clientid_arr[dmb_node->sba_idx] = NO_DIBS_CLIENT;
160		spin_unlock_irqrestore(&dibs->lock, flags);
161
162		__dibs_lo_unregister_dmb(ldev, dmb_node);
163	}
164	return 0;
165}
166
167static int dibs_lo_support_dmb_nocopy(struct dibs_dev *dibs)
168{
169	return DIBS_LO_SUPPORT_NOCOPY;
170}
171
172static int dibs_lo_attach_dmb(struct dibs_dev *dibs, struct dibs_dmb *dmb)
173{
174	struct dibs_lo_dmb_node *dmb_node = NULL, *tmp_node;
175	struct dibs_lo_dev *ldev;
176
177	ldev = dibs->drv_priv;
178
179	/* find dmb_node according to dmb->dmb_tok */
180	read_lock_bh(&ldev->dmb_ht_lock);
181	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, dmb->dmb_tok) {
182		if (tmp_node->token == dmb->dmb_tok) {
183			dmb_node = tmp_node;
184			break;
185		}
186	}
187	if (!dmb_node) {
188		read_unlock_bh(&ldev->dmb_ht_lock);
189		return -EINVAL;
190	}
191	read_unlock_bh(&ldev->dmb_ht_lock);
192
193	if (!refcount_inc_not_zero(&dmb_node->refcnt))
194		/* the dmb is being unregistered, but has
195		 * not been removed from the hash table.
196		 */
197		return -EINVAL;
198
199	/* provide dmb information */
200	dmb->idx = dmb_node->sba_idx;
201	dmb->dmb_tok = dmb_node->token;
202	dmb->cpu_addr = dmb_node->cpu_addr;
203	dmb->dma_addr = dmb_node->dma_addr;
204	dmb->dmb_len = dmb_node->len;
205	return 0;
206}
207
208static int dibs_lo_detach_dmb(struct dibs_dev *dibs, u64 token)
209{
210	struct dibs_lo_dmb_node *dmb_node = NULL, *tmp_node;
211	struct dibs_lo_dev *ldev;
212
213	ldev = dibs->drv_priv;
214
215	/* find dmb_node according to dmb->dmb_tok */
216	read_lock_bh(&ldev->dmb_ht_lock);
217	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, token) {
218		if (tmp_node->token == token) {
219			dmb_node = tmp_node;
220			break;
221		}
222	}
223	if (!dmb_node) {
224		read_unlock_bh(&ldev->dmb_ht_lock);
225		return -EINVAL;
226	}
227	read_unlock_bh(&ldev->dmb_ht_lock);
228
229	if (refcount_dec_and_test(&dmb_node->refcnt))
230		__dibs_lo_unregister_dmb(ldev, dmb_node);
231	return 0;
232}
233
234static int dibs_lo_move_data(struct dibs_dev *dibs, u64 dmb_tok,
235			     unsigned int idx, bool sf, unsigned int offset,
236			     void *data, unsigned int size)
237{
238	struct dibs_lo_dmb_node *rmb_node = NULL, *tmp_node;
239	struct dibs_lo_dev *ldev;
240	u16 s_mask;
241	u8 client_id;
242	u32 sba_idx;
243
244	ldev = dibs->drv_priv;
245
246	read_lock_bh(&ldev->dmb_ht_lock);
247	hash_for_each_possible(ldev->dmb_ht, tmp_node, list, dmb_tok) {
248		if (tmp_node->token == dmb_tok) {
249			rmb_node = tmp_node;
250			break;
251		}
252	}
253	if (!rmb_node) {
254		read_unlock_bh(&ldev->dmb_ht_lock);
255		return -EINVAL;
256	}
257	memcpy((char *)rmb_node->cpu_addr + offset, data, size);
258	sba_idx = rmb_node->sba_idx;
259	read_unlock_bh(&ldev->dmb_ht_lock);
260
261	if (!sf)
262		return 0;
263
264	spin_lock(&dibs->lock);
265	client_id = dibs->dmb_clientid_arr[sba_idx];
266	s_mask = ror16(0x1000, idx);
267	if (likely(client_id != NO_DIBS_CLIENT && dibs->subs[client_id]))
268		dibs->subs[client_id]->ops->handle_irq(dibs, sba_idx, s_mask);
269	spin_unlock(&dibs->lock);
270
271	return 0;
272}
273
274static const struct dibs_dev_ops dibs_lo_ops = {
275	.get_fabric_id = dibs_lo_get_fabric_id,
276	.query_remote_gid = dibs_lo_query_rgid,
277	.max_dmbs = dibs_lo_max_dmbs,
278	.register_dmb = dibs_lo_register_dmb,
279	.unregister_dmb = dibs_lo_unregister_dmb,
280	.move_data = dibs_lo_move_data,
281	.support_mmapped_rdmb = dibs_lo_support_dmb_nocopy,
282	.attach_dmb = dibs_lo_attach_dmb,
283	.detach_dmb = dibs_lo_detach_dmb,
284};
285
286static void dibs_lo_dev_init(struct dibs_lo_dev *ldev)
287{
288	rwlock_init(&ldev->dmb_ht_lock);
289	hash_init(ldev->dmb_ht);
290	atomic_set(&ldev->dmb_cnt, 0);
291	init_waitqueue_head(&ldev->ldev_release);
292}
293
294static void dibs_lo_dev_exit(struct dibs_lo_dev *ldev)
295{
296	if (atomic_read(&ldev->dmb_cnt))
297		wait_event(ldev->ldev_release, !atomic_read(&ldev->dmb_cnt));
298}
299
300static int dibs_lo_dev_probe(void)
301{
302	struct dibs_lo_dev *ldev;
303	struct dibs_dev *dibs;
304	int ret;
305
306	ldev = kzalloc(sizeof(*ldev), GFP_KERNEL);
307	if (!ldev)
308		return -ENOMEM;
309
310	dibs = dibs_dev_alloc();
311	if (!dibs) {
312		kfree(ldev);
313		return -ENOMEM;
314	}
315
316	ldev->dibs = dibs;
317	dibs->drv_priv = ldev;
318	dibs_lo_dev_init(ldev);
319	uuid_gen(&dibs->gid);
320	dibs->ops = &dibs_lo_ops;
321
322	dibs->dev.parent = NULL;
323	dev_set_name(&dibs->dev, "%s", dibs_lo_dev_name);
324
325	ret = dibs_dev_add(dibs);
326	if (ret)
327		goto err_reg;
328	lo_dev = ldev;
329	return 0;
330
331err_reg:
332	kfree(dibs->dmb_clientid_arr);
333	/* pairs with dibs_dev_alloc() */
334	put_device(&dibs->dev);
335	kfree(ldev);
336
337	return ret;
338}
339
340static void dibs_lo_dev_remove(void)
341{
342	if (!lo_dev)
343		return;
344
345	dibs_dev_del(lo_dev->dibs);
346	dibs_lo_dev_exit(lo_dev);
347	/* pairs with dibs_dev_alloc() */
348	put_device(&lo_dev->dibs->dev);
349	kfree(lo_dev);
350	lo_dev = NULL;
351}
352
353int dibs_loopback_init(void)
354{
355	return dibs_lo_dev_probe();
356}
357
358void dibs_loopback_exit(void)
359{
360	dibs_lo_dev_remove();
361}