mlx5_pagealloc.c revision 331807 
1/*-
2 * Copyright (c) 2013-2017, Mellanox Technologies, Ltd.  All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 *    notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 *    notice, this list of conditions and the following disclaimer in the
11 *    documentation and/or other materials provided with the distribution.
12 *
13 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23 * SUCH DAMAGE.
24 *
25 * $FreeBSD: stable/11/sys/dev/mlx5/mlx5_core/mlx5_pagealloc.c 331807 2018-03-30 19:13:17Z hselasky $
26 */
27
28#include <linux/kernel.h>
29#include <linux/module.h>
30#include <linux/delay.h>
31#include <dev/mlx5/driver.h>
32#include "mlx5_core.h"
33
34CTASSERT((uintptr_t)PAGE_MASK > (uintptr_t)PAGE_SIZE);
35
36struct mlx5_pages_req {
37	struct mlx5_core_dev *dev;
38	u16	func_id;
39	s32	npages;
40	struct work_struct work;
41};
42
43
44enum {
45	MAX_RECLAIM_TIME_MSECS	= 5000,
46};
47
48static void
49mlx5_fwp_load_mem_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
50{
51	struct mlx5_fw_page *fwp;
52	uint8_t owned;
53
54	fwp = (struct mlx5_fw_page *)arg;
55	owned = MLX5_DMA_OWNED(fwp->dev);
56
57	if (!owned)
58		MLX5_DMA_LOCK(fwp->dev);
59
60	if (error == 0) {
61		KASSERT(nseg == 1, ("Number of segments is different from 1"));
62		fwp->dma_addr = segs->ds_addr;
63		fwp->load_done = MLX5_LOAD_ST_SUCCESS;
64	} else {
65		fwp->load_done = MLX5_LOAD_ST_FAILURE;
66	}
67	MLX5_DMA_DONE(fwp->dev);
68
69	if (!owned)
70		MLX5_DMA_UNLOCK(fwp->dev);
71}
72
73void
74mlx5_fwp_flush(struct mlx5_fw_page *fwp)
75{
76	unsigned num = fwp->numpages;
77
78	while (num--)
79		bus_dmamap_sync(fwp[num].dev->cmd.dma_tag, fwp[num].dma_map, BUS_DMASYNC_PREWRITE);
80}
81
82void
83mlx5_fwp_invalidate(struct mlx5_fw_page *fwp)
84{
85	unsigned num = fwp->numpages;
86
87	while (num--) {
88		bus_dmamap_sync(fwp[num].dev->cmd.dma_tag, fwp[num].dma_map, BUS_DMASYNC_POSTREAD);
89		bus_dmamap_sync(fwp[num].dev->cmd.dma_tag, fwp[num].dma_map, BUS_DMASYNC_PREREAD);
90	}
91}
92
93struct mlx5_fw_page *
94mlx5_fwp_alloc(struct mlx5_core_dev *dev, gfp_t flags, unsigned num)
95{
96	struct mlx5_fw_page *fwp;
97	unsigned x;
98	int err;
99
100	/* check for special case */
101	if (num == 0) {
102		fwp = kzalloc(sizeof(*fwp), flags);
103		if (fwp != NULL)
104			fwp->dev = dev;
105		return (fwp);
106	}
107
108	/* we need sleeping context for this function */
109	if (flags & M_NOWAIT)
110		return (NULL);
111
112	fwp = kzalloc(sizeof(*fwp) * num, flags);
113
114	/* serialize loading the DMA map(s) */
115	sx_xlock(&dev->cmd.dma_sx);
116
117	for (x = 0; x != num; x++) {
118		/* store pointer to MLX5 core device */
119		fwp[x].dev = dev;
120		/* store number of pages left from the array */
121		fwp[x].numpages = num - x;
122
123		/* allocate memory */
124		err = bus_dmamem_alloc(dev->cmd.dma_tag, &fwp[x].virt_addr,
125		    BUS_DMA_WAITOK | BUS_DMA_COHERENT, &fwp[x].dma_map);
126		if (err != 0)
127			goto failure;
128
129		/* load memory into DMA */
130		MLX5_DMA_LOCK(dev);
131		err = bus_dmamap_load(
132		    dev->cmd.dma_tag, fwp[x].dma_map, fwp[x].virt_addr,
133		    MLX5_ADAPTER_PAGE_SIZE, &mlx5_fwp_load_mem_cb,
134		    fwp + x, BUS_DMA_WAITOK | BUS_DMA_COHERENT);
135
136		while (fwp[x].load_done == MLX5_LOAD_ST_NONE)
137			MLX5_DMA_WAIT(dev);
138		MLX5_DMA_UNLOCK(dev);
139
140		/* check for error */
141		if (fwp[x].load_done != MLX5_LOAD_ST_SUCCESS) {
142			bus_dmamem_free(dev->cmd.dma_tag, fwp[x].virt_addr,
143			    fwp[x].dma_map);
144			goto failure;
145		}
146	}
147	sx_xunlock(&dev->cmd.dma_sx);
148	return (fwp);
149
150failure:
151	while (x--) {
152		bus_dmamap_unload(dev->cmd.dma_tag, fwp[x].dma_map);
153		bus_dmamem_free(dev->cmd.dma_tag, fwp[x].virt_addr, fwp[x].dma_map);
154	}
155	sx_xunlock(&dev->cmd.dma_sx);
156	return (NULL);
157}
158
159void
160mlx5_fwp_free(struct mlx5_fw_page *fwp)
161{
162	struct mlx5_core_dev *dev;
163	unsigned num;
164
165	/* be NULL safe */
166	if (fwp == NULL)
167		return;
168
169	/* check for special case */
170	if (fwp->numpages == 0) {
171		kfree(fwp);
172		return;
173	}
174
175	num = fwp->numpages;
176	dev = fwp->dev;
177
178	while (num--) {
179		bus_dmamap_unload(dev->cmd.dma_tag, fwp[num].dma_map);
180		bus_dmamem_free(dev->cmd.dma_tag, fwp[num].virt_addr, fwp[num].dma_map);
181	}
182
183	kfree(fwp);
184}
185
186u64
187mlx5_fwp_get_dma(struct mlx5_fw_page *fwp, size_t offset)
188{
189	size_t index = (offset / MLX5_ADAPTER_PAGE_SIZE);
190	KASSERT(index < fwp->numpages, ("Invalid offset: %lld", (long long)offset));
191
192	return ((fwp + index)->dma_addr + (offset % MLX5_ADAPTER_PAGE_SIZE));
193}
194
195void *
196mlx5_fwp_get_virt(struct mlx5_fw_page *fwp, size_t offset)
197{
198	size_t index = (offset / MLX5_ADAPTER_PAGE_SIZE);
199	KASSERT(index < fwp->numpages, ("Invalid offset: %lld", (long long)offset));
200
201	return ((char *)(fwp + index)->virt_addr + (offset % MLX5_ADAPTER_PAGE_SIZE));
202}
203
204static int
205mlx5_insert_fw_page_locked(struct mlx5_core_dev *dev, struct mlx5_fw_page *nfp)
206{
207	struct rb_root *root = &dev->priv.page_root;
208	struct rb_node **new = &root->rb_node;
209	struct rb_node *parent = NULL;
210	struct mlx5_fw_page *tfp;
211
212	while (*new) {
213		parent = *new;
214		tfp = rb_entry(parent, struct mlx5_fw_page, rb_node);
215		if (tfp->dma_addr < nfp->dma_addr)
216			new = &parent->rb_left;
217		else if (tfp->dma_addr > nfp->dma_addr)
218			new = &parent->rb_right;
219		else
220			return (-EEXIST);
221	}
222
223	rb_link_node(&nfp->rb_node, parent, new);
224	rb_insert_color(&nfp->rb_node, root);
225	return (0);
226}
227
228static struct mlx5_fw_page *
229mlx5_remove_fw_page_locked(struct mlx5_core_dev *dev, bus_addr_t addr)
230{
231	struct rb_root *root = &dev->priv.page_root;
232	struct rb_node *tmp = root->rb_node;
233	struct mlx5_fw_page *result = NULL;
234	struct mlx5_fw_page *tfp;
235
236	while (tmp) {
237		tfp = rb_entry(tmp, struct mlx5_fw_page, rb_node);
238		if (tfp->dma_addr < addr) {
239			tmp = tmp->rb_left;
240		} else if (tfp->dma_addr > addr) {
241			tmp = tmp->rb_right;
242		} else {
243			rb_erase(&tfp->rb_node, &dev->priv.page_root);
244			result = tfp;
245			break;
246		}
247	}
248	return (result);
249}
250
251static int
252alloc_4k(struct mlx5_core_dev *dev, u64 *addr, u16 func_id)
253{
254	struct mlx5_fw_page *fwp;
255	int err;
256
257	fwp = mlx5_fwp_alloc(dev, GFP_KERNEL, 1);
258	if (fwp == NULL)
259		return (-ENOMEM);
260
261	fwp->func_id = func_id;
262
263	MLX5_DMA_LOCK(dev);
264	err = mlx5_insert_fw_page_locked(dev, fwp);
265	MLX5_DMA_UNLOCK(dev);
266
267	if (err != 0) {
268		mlx5_fwp_free(fwp);
269	} else {
270		/* make sure cached data is cleaned */
271		mlx5_fwp_invalidate(fwp);
272
273		/* store DMA address */
274		*addr = fwp->dma_addr;
275	}
276	return (err);
277}
278
279static void
280free_4k(struct mlx5_core_dev *dev, u64 addr)
281{
282	struct mlx5_fw_page *fwp;
283
284	MLX5_DMA_LOCK(dev);
285	fwp = mlx5_remove_fw_page_locked(dev, addr);
286	MLX5_DMA_UNLOCK(dev);
287
288	if (fwp == NULL) {
289		mlx5_core_warn(dev, "Cannot free 4K page at 0x%llx\n", (long long)addr);
290		return;
291	}
292	mlx5_fwp_free(fwp);
293}
294
295static int mlx5_cmd_query_pages(struct mlx5_core_dev *dev, u16 *func_id,
296				s32 *npages, int boot)
297{
298	u32 in[MLX5_ST_SZ_DW(query_pages_in)] = {0};
299	u32 out[MLX5_ST_SZ_DW(query_pages_out)] = {0};
300	int err;
301
302	MLX5_SET(query_pages_in, in, opcode, MLX5_CMD_OP_QUERY_PAGES);
303	MLX5_SET(query_pages_in, in, op_mod, boot ?
304		 MLX5_QUERY_PAGES_IN_OP_MOD_BOOT_PAGES :
305		 MLX5_QUERY_PAGES_IN_OP_MOD_INIT_PAGES);
306
307	err = mlx5_cmd_exec(dev, in, sizeof(in), out, sizeof(out));
308	if (err)
309		return err;
310
311	*npages = MLX5_GET(query_pages_out, out, num_pages);
312	*func_id = MLX5_GET(query_pages_out, out, function_id);
313
314	return 0;
315}
316
317static int give_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
318		      int notify_fail)
319{
320	u32 out[MLX5_ST_SZ_DW(manage_pages_out)] = {0};
321	int inlen = MLX5_ST_SZ_BYTES(manage_pages_in);
322	u64 addr;
323	int err;
324	u32 *in, *nin;
325	int i = 0;
326
327	inlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_in, pas[0]);
328	in = mlx5_vzalloc(inlen);
329	if (!in) {
330		mlx5_core_warn(dev, "vzalloc failed %d\n", inlen);
331		err = -ENOMEM;
332		goto out_alloc;
333	}
334
335	for (i = 0; i < npages; i++) {
336		err = alloc_4k(dev, &addr, func_id);
337		if (err)
338			goto out_alloc;
339		MLX5_ARRAY_SET64(manage_pages_in, in, pas, i, addr);
340	}
341
342	MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
343	MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_GIVE);
344	MLX5_SET(manage_pages_in, in, function_id, func_id);
345	MLX5_SET(manage_pages_in, in, input_num_entries, npages);
346
347	err = mlx5_cmd_exec(dev, in, inlen, out, sizeof(out));
348	if (err) {
349		mlx5_core_warn(dev, "func_id 0x%x, npages %d, err %d\n",
350			       func_id, npages, err);
351		goto out_alloc;
352	}
353	dev->priv.fw_pages += npages;
354	dev->priv.pages_per_func[func_id] += npages;
355
356	mlx5_core_dbg(dev, "err %d\n", err);
357
358	goto out_free;
359
360out_alloc:
361	if (notify_fail) {
362		nin = mlx5_vzalloc(inlen);
363		if (!nin)
364			goto out_4k;
365
366		memset(&out, 0, sizeof(out));
367		MLX5_SET(manage_pages_in, nin, opcode, MLX5_CMD_OP_MANAGE_PAGES);
368		MLX5_SET(manage_pages_in, nin, op_mod, MLX5_PAGES_CANT_GIVE);
369		MLX5_SET(manage_pages_in, nin, function_id, func_id);
370		if (mlx5_cmd_exec(dev, nin, inlen, out, sizeof(out)))
371			mlx5_core_warn(dev, "page notify failed\n");
372		kvfree(nin);
373	}
374
375out_4k:
376	for (i--; i >= 0; i--)
377		free_4k(dev, MLX5_GET64(manage_pages_in, in, pas[i]));
378out_free:
379	kvfree(in);
380	return err;
381}
382
383static int reclaim_pages(struct mlx5_core_dev *dev, u32 func_id, int npages,
384			 int *nclaimed)
385{
386	int outlen = MLX5_ST_SZ_BYTES(manage_pages_out);
387	u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {0};
388	int num_claimed;
389	u32 *out;
390	int err;
391	int i;
392
393	if (nclaimed)
394		*nclaimed = 0;
395
396	outlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
397	out = mlx5_vzalloc(outlen);
398	if (!out)
399		return -ENOMEM;
400
401	MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
402	MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_TAKE);
403	MLX5_SET(manage_pages_in, in, function_id, func_id);
404	MLX5_SET(manage_pages_in, in, input_num_entries, npages);
405
406	mlx5_core_dbg(dev, "npages %d, outlen %d\n", npages, outlen);
407	err = mlx5_cmd_exec(dev, in, sizeof(in), out, outlen);
408	if (err) {
409		mlx5_core_err(dev, "failed reclaiming pages\n");
410		goto out_free;
411	}
412
413	num_claimed = MLX5_GET(manage_pages_out, out, output_num_entries);
414	if (nclaimed)
415		*nclaimed = num_claimed;
416
417	dev->priv.fw_pages -= num_claimed;
418	dev->priv.pages_per_func[func_id] -= num_claimed;
419	for (i = 0; i < num_claimed; i++)
420		free_4k(dev, MLX5_GET64(manage_pages_out, out, pas[i]));
421
422out_free:
423	kvfree(out);
424	return err;
425}
426
427static void pages_work_handler(struct work_struct *work)
428{
429	struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
430	struct mlx5_core_dev *dev = req->dev;
431	int err = 0;
432
433	if (req->npages < 0)
434		err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL);
435	else if (req->npages > 0)
436		err = give_pages(dev, req->func_id, req->npages, 1);
437
438	if (err)
439		mlx5_core_warn(dev, "%s fail %d\n",
440			       req->npages < 0 ? "reclaim" : "give", err);
441
442	kfree(req);
443}
444
445void mlx5_core_req_pages_handler(struct mlx5_core_dev *dev, u16 func_id,
446				 s32 npages)
447{
448	struct mlx5_pages_req *req;
449
450	req = kzalloc(sizeof(*req), GFP_ATOMIC);
451	if (!req) {
452		mlx5_core_warn(dev, "failed to allocate pages request\n");
453		return;
454	}
455
456	req->dev = dev;
457	req->func_id = func_id;
458	req->npages = npages;
459	INIT_WORK(&req->work, pages_work_handler);
460	if (!queue_work(dev->priv.pg_wq, &req->work))
461		mlx5_core_warn(dev, "failed to queue pages handler work\n");
462}
463
464int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
465{
466	u16 uninitialized_var(func_id);
467	s32 uninitialized_var(npages);
468	int err;
469
470	err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
471	if (err)
472		return err;
473
474	mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
475		      npages, boot ? "boot" : "init", func_id);
476
477	return give_pages(dev, func_id, npages, 0);
478}
479
480enum {
481	MLX5_BLKS_FOR_RECLAIM_PAGES = 12
482};
483
484s64 mlx5_wait_for_reclaim_vfs_pages(struct mlx5_core_dev *dev)
485{
486	int end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
487	s64 prevpages = 0;
488	s64 npages = 0;
489
490	while (!time_after(jiffies, end)) {
491		/* exclude own function, VFs only */
492		npages = dev->priv.fw_pages - dev->priv.pages_per_func[0];
493		if (!npages)
494			break;
495
496		if (npages != prevpages)
497			end = end + msecs_to_jiffies(100);
498
499		prevpages = npages;
500		msleep(1);
501	}
502
503	if (npages)
504		mlx5_core_warn(dev, "FW did not return all VFs pages, will cause to memory leak\n");
505
506	return -npages;
507}
508
509static int optimal_reclaimed_pages(void)
510{
511	struct mlx5_cmd_prot_block *block;
512	struct mlx5_cmd_layout *lay;
513	int ret;
514
515	ret = (sizeof(lay->out) + MLX5_BLKS_FOR_RECLAIM_PAGES * sizeof(block->data) -
516	       MLX5_ST_SZ_BYTES(manage_pages_out)) /
517	       MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
518
519	return ret;
520}
521
522int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
523{
524	int end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
525	struct mlx5_fw_page *fwp;
526	struct rb_node *p;
527	int nclaimed = 0;
528	int err;
529
530	do {
531		p = rb_first(&dev->priv.page_root);
532		if (p) {
533			fwp = rb_entry(p, struct mlx5_fw_page, rb_node);
534			if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
535				--dev->priv.fw_pages;
536				free_4k(dev, fwp->dma_addr);
537				nclaimed = 1;
538			} else {
539				err = reclaim_pages(dev, fwp->func_id,
540						    optimal_reclaimed_pages(),
541						    &nclaimed);
542				if (err) {
543					mlx5_core_warn(dev, "failed reclaiming pages (%d)\n",
544						       err);
545					return err;
546				}
547			}
548
549			if (nclaimed)
550				end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
551		}
552		if (time_after(jiffies, end)) {
553			mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
554			break;
555		}
556	} while (p);
557
558	return 0;
559}
560
561void mlx5_pagealloc_init(struct mlx5_core_dev *dev)
562{
563
564	dev->priv.page_root = RB_ROOT;
565}
566
567void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
568{
569	/* nothing */
570}
571
572int mlx5_pagealloc_start(struct mlx5_core_dev *dev)
573{
574	dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
575	if (!dev->priv.pg_wq)
576		return -ENOMEM;
577
578	return 0;
579}
580
581void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
582{
583	destroy_workqueue(dev->priv.pg_wq);
584}
585