xref: /NextBSD/sys/ofed/drivers/infiniband/core/umem.c (revision 287e3b14e9552995def1802ec9c5034f4adf28ec)
1 /*
2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <linux/mm.h>
36 #include <linux/dma-mapping.h>
37 #include <linux/sched.h>
38 #include <linux/dma-attrs.h>
39 #include <linux/slab.h>
40 #include <linux/module.h>
41 #include <sys/priv.h>
42 #include <sys/resourcevar.h>
43 #include <vm/vm_pageout.h>
44 #include <vm/vm_map.h>
45 #include "uverbs.h"
46 
47 #define IB_UMEM_MAX_PAGE_CHUNK		(PAGE_SIZE / sizeof (struct page *))
48 
49 static int allow_weak_ordering;
50 module_param_named(weak_ordering, allow_weak_ordering, int, 0444);
51 MODULE_PARM_DESC(weak_ordering,  "Allow weak ordering for data registered memory");
52 
peer_umem_get(struct ib_peer_memory_client * ib_peer_mem,struct ib_umem * umem,unsigned long addr,int dmasync,int invalidation_supported)53 static struct ib_umem *peer_umem_get(struct ib_peer_memory_client *ib_peer_mem,
54 				       struct ib_umem *umem, unsigned long addr,
55 				       int dmasync, int invalidation_supported)
56 {
57 	int ret;
58 	const struct peer_memory_client *peer_mem = ib_peer_mem->peer_mem;
59 	struct invalidation_ctx *invalidation_ctx = NULL;
60 
61 	umem->ib_peer_mem = ib_peer_mem;
62 	if (invalidation_supported) {
63 		invalidation_ctx = kzalloc(sizeof(*invalidation_ctx), GFP_KERNEL);
64 		if (!invalidation_ctx) {
65 			ret = -ENOMEM;
66 			goto out;
67 		}
68 		umem->invalidation_ctx = invalidation_ctx;
69 		invalidation_ctx->umem = umem;
70 		mutex_lock(&ib_peer_mem->lock);
71 		invalidation_ctx->context_ticket =
72 				ib_peer_insert_context(ib_peer_mem, invalidation_ctx);
73 		/* unlock before calling get pages to prevent a dead-lock from the callback */
74 		mutex_unlock(&ib_peer_mem->lock);
75 	}
76 
77 	ret = peer_mem->get_pages(addr, umem->length, umem->writable, 1,
78 				&umem->sg_head,
79 				umem->peer_mem_client_context,
80 				invalidation_ctx ?
81 				(void *)invalidation_ctx->context_ticket : NULL);
82 
83 	if (invalidation_ctx) {
84 		/* taking the lock back, checking that wasn't invalidated at that time */
85 		mutex_lock(&ib_peer_mem->lock);
86 		if (invalidation_ctx->peer_invalidated) {
87 			printk(KERN_ERR "peer_umem_get: pages were invalidated by peer\n");
88 			ret = -EINVAL;
89 		}
90 	}
91 
92 	if (ret)
93 		goto out;
94 
95 	umem->page_size = peer_mem->get_page_size
96 					(umem->peer_mem_client_context);
97 	if (umem->page_size <= 0)
98 		goto put_pages;
99 
100 	umem->offset = addr & ((unsigned long)umem->page_size - 1);
101 	ret = peer_mem->dma_map(&umem->sg_head,
102 					umem->peer_mem_client_context,
103 					umem->context->device->dma_device,
104 					dmasync,
105 					&umem->nmap);
106 	if (ret)
107 		goto put_pages;
108 
109 	ib_peer_mem->stats.num_reg_pages +=
110 			umem->nmap * (umem->page_size >> PAGE_SHIFT);
111 	ib_peer_mem->stats.num_alloc_mrs += 1;
112 	return umem;
113 
114 put_pages:
115 
116 	peer_mem->put_pages(umem->peer_mem_client_context,
117 					&umem->sg_head);
118 out:
119 	if (invalidation_ctx) {
120 		ib_peer_remove_context(ib_peer_mem, invalidation_ctx->context_ticket);
121 		mutex_unlock(&umem->ib_peer_mem->lock);
122 		kfree(invalidation_ctx);
123 	}
124 
125 	ib_put_peer_client(ib_peer_mem, umem->peer_mem_client_context,
126 				umem->peer_mem_srcu_key);
127 	kfree(umem);
128 	return ERR_PTR(ret);
129 }
130 
peer_umem_release(struct ib_umem * umem)131 static void peer_umem_release(struct ib_umem *umem)
132 {
133 	struct ib_peer_memory_client *ib_peer_mem = umem->ib_peer_mem;
134 	const struct peer_memory_client *peer_mem = ib_peer_mem->peer_mem;
135 	struct invalidation_ctx *invalidation_ctx = umem->invalidation_ctx;
136 
137 	if (invalidation_ctx) {
138 
139 		int peer_callback;
140 		int inflight_invalidation;
141 		/* If we are not under peer callback we must take the lock before removing
142 		  * core ticket from the tree and releasing its umem.
143 		  * It will let any inflight callbacks to be ended safely.
144 		  * If we are under peer callback or under error flow of reg_mr so that context
145 		  * wasn't activated yet lock was already taken.
146 		*/
147 		if (invalidation_ctx->func && !invalidation_ctx->peer_callback)
148 			mutex_lock(&ib_peer_mem->lock);
149 		ib_peer_remove_context(ib_peer_mem, invalidation_ctx->context_ticket);
150 		/* make sure to check inflight flag after took the lock and remove from tree.
151 		  * in addition, from that point using local variables for peer_callback and
152 		  * inflight_invalidation as after the complete invalidation_ctx can't be accessed
153 		  * any more as it may be freed by the callback.
154 		*/
155 		peer_callback = invalidation_ctx->peer_callback;
156 		inflight_invalidation = invalidation_ctx->inflight_invalidation;
157 		if (inflight_invalidation)
158 			complete(&invalidation_ctx->comp);
159 		/* On peer callback lock is handled externally */
160 		if (!peer_callback)
161 			/* unlocking before put_pages */
162 			mutex_unlock(&ib_peer_mem->lock);
163 		/* in case under callback context or callback is pending let it free the invalidation context */
164 		if (!peer_callback && !inflight_invalidation)
165 			kfree(invalidation_ctx);
166 	}
167 
168 	peer_mem->dma_unmap(&umem->sg_head,
169 					umem->peer_mem_client_context,
170 					umem->context->device->dma_device);
171 	peer_mem->put_pages(&umem->sg_head,
172 					  umem->peer_mem_client_context);
173 
174 	ib_peer_mem->stats.num_dereg_pages +=
175 			umem->nmap * (umem->page_size >> PAGE_SHIFT);
176 	ib_peer_mem->stats.num_dealloc_mrs += 1;
177 	ib_put_peer_client(ib_peer_mem, umem->peer_mem_client_context,
178 				umem->peer_mem_srcu_key);
179 	kfree(umem);
180 
181 	return;
182 
183 }
184 
__ib_umem_release(struct ib_device * dev,struct ib_umem * umem,int dirty)185 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
186 {
187 
188 	vm_object_t object;
189 	struct scatterlist *sg;
190 	struct page *page;
191 	int i;
192 
193 	object = NULL;
194 	if (umem->nmap > 0)
195 		ib_dma_unmap_sg(dev, umem->sg_head.sgl,
196 			umem->nmap,
197 			DMA_BIDIRECTIONAL);
198 	for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
199 		page = sg_page(sg);
200 			if (umem->writable && dirty) {
201 				if (object && object != page->object)
202 					VM_OBJECT_WUNLOCK(object);
203 				if (object != page->object) {
204 					object = page->object;
205 					VM_OBJECT_WLOCK(object);
206 				}
207 				vm_page_dirty(page);
208 			}
209 		}
210 	sg_free_table(&umem->sg_head);
211 	if (object)
212 		VM_OBJECT_WUNLOCK(object);
213 
214 }
215 
ib_umem_activate_invalidation_notifier(struct ib_umem * umem,umem_invalidate_func_t func,void * cookie)216 void ib_umem_activate_invalidation_notifier(struct ib_umem *umem,
217 					       umem_invalidate_func_t func,
218 					       void *cookie)
219 {
220 	struct invalidation_ctx *invalidation_ctx = umem->invalidation_ctx;
221 
222 	invalidation_ctx->func = func;
223 	invalidation_ctx->cookie = cookie;
224 
225 	/* from that point any pending invalidations can be called */
226 	mutex_unlock(&umem->ib_peer_mem->lock);
227 	return;
228 }
229 EXPORT_SYMBOL(ib_umem_activate_invalidation_notifier);
230 /**
231  * ib_umem_get - Pin and DMA map userspace memory.
232  * @context: userspace context to pin memory for
233  * @addr: userspace virtual address to start at
234  * @size: length of region to pin
235  * @access: IB_ACCESS_xxx flags for memory being pinned
236  * @dmasync: flush in-flight DMA when the memory region is written
237  */
ib_umem_get_ex(struct ib_ucontext * context,unsigned long addr,size_t size,int access,int dmasync,int invalidation_supported)238 struct ib_umem *ib_umem_get_ex(struct ib_ucontext *context, unsigned long addr,
239 			    size_t size, int access, int dmasync,
240 			    int invalidation_supported)
241 {
242 
243 	struct ib_umem *umem;
244         struct proc *proc;
245 	pmap_t pmap;
246         vm_offset_t end, last, start;
247         vm_size_t npages;
248         int error;
249 	int ret;
250 	int ents;
251 	int i;
252 	DEFINE_DMA_ATTRS(attrs);
253 	struct scatterlist *sg, *sg_list_start;
254 	int need_release = 0;
255 
256 	error = priv_check(curthread, PRIV_VM_MLOCK);
257 	if (error)
258 		return ERR_PTR(-error);
259 
260 	last = addr + size;
261 	start = addr & PAGE_MASK; /* Use the linux PAGE_MASK definition. */
262 	end = roundup2(last, PAGE_SIZE); /* Use PAGE_MASK safe operation. */
263 	if (last < addr || end < addr)
264 		return ERR_PTR(-EINVAL);
265 	npages = atop(end - start);
266 	if (npages > vm_page_max_wired)
267 		return ERR_PTR(-ENOMEM);
268 	umem = kzalloc(sizeof *umem, GFP_KERNEL);
269 	if (!umem)
270 		return ERR_PTR(-ENOMEM);
271 	proc = curthread->td_proc;
272 	PROC_LOCK(proc);
273 	if (ptoa(npages +
274 	    pmap_wired_count(vm_map_pmap(&proc->p_vmspace->vm_map))) >
275 	    lim_cur_proc(proc, RLIMIT_MEMLOCK)) {
276 		PROC_UNLOCK(proc);
277 		kfree(umem);
278 		return ERR_PTR(-ENOMEM);
279 	}
280         PROC_UNLOCK(proc);
281 	if (npages + vm_cnt.v_wire_count > vm_page_max_wired) {
282 		kfree(umem);
283 		return ERR_PTR(-EAGAIN);
284 	}
285 	error = vm_map_wire(&proc->p_vmspace->vm_map, start, end,
286 	    VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES |
287 	    (umem->writable ? VM_MAP_WIRE_WRITE : 0));
288 	if (error != KERN_SUCCESS) {
289 		kfree(umem);
290 		return ERR_PTR(-ENOMEM);
291 	}
292 
293 	umem->context   = context;
294 	umem->length    = size;
295 	umem->offset    = addr & ~PAGE_MASK;
296 	umem->page_size = PAGE_SIZE;
297 	umem->start	= addr;
298 	/*
299 	 * We ask for writable memory if any access flags other than
300 	 * "remote read" are set.  "Local write" and "remote write"
301 	 * obviously require write access.  "Remote atomic" can do
302 	 * things like fetch and add, which will modify memory, and
303 	 * "MW bind" can change permissions by binding a window.
304 	 */
305 	umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
306 
307 	if (invalidation_supported || context->peer_mem_private_data) {
308 
309 		struct ib_peer_memory_client *peer_mem_client;
310 
311 		peer_mem_client =  ib_get_peer_client(context, addr, size,
312 			&umem->peer_mem_client_context,
313 				&umem->peer_mem_srcu_key);
314 		if (peer_mem_client)
315 			return peer_umem_get(peer_mem_client, umem, addr,
316 				dmasync, invalidation_supported);
317 	}
318 
319 	umem->hugetlb = 0;
320 
321 	pmap = vm_map_pmap(&proc->p_vmspace->vm_map);
322 
323 	if (npages == 0) {
324 		ret = -EINVAL;
325 			goto out;
326 		}
327 
328 	ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
329 	if (ret)
330 		goto out;
331 
332 	need_release = 1;
333 	sg_list_start = umem->sg_head.sgl;
334 
335 	while (npages) {
336 
337 		ents = min_t(int, npages, IB_UMEM_MAX_PAGE_CHUNK);
338 		umem->npages += ents;
339 
340 		for_each_sg(sg_list_start, sg, ents, i) {
341 			vm_paddr_t pa;
342 
343 			pa = pmap_extract(pmap, start);
344 			if (pa == 0) {
345 				ret = -ENOMEM;
346 				goto out;
347 			}
348 			sg_set_page(sg, PHYS_TO_VM_PAGE(pa),
349 			    PAGE_SIZE, 0);
350 			npages--;
351 			start += PAGE_SIZE;
352 		}
353 
354 		/* preparing for next loop */
355 		sg_list_start = sg;
356 	}
357 
358 	umem->nmap = ib_dma_map_sg_attrs(context->device,
359 					umem->sg_head.sgl,
360 					umem->npages,
361 						  DMA_BIDIRECTIONAL,
362 						  &attrs);
363 	if (umem->nmap != umem->npages) {
364 			ret = -ENOMEM;
365 			goto out;
366 		}
367 
368 out:
369 	if (ret < 0) {
370 		if (need_release)
371 		__ib_umem_release(context->device, umem, 0);
372 		kfree(umem);
373 	}
374 
375 	return ret < 0 ? ERR_PTR(ret) : umem;
376 }
377 EXPORT_SYMBOL(ib_umem_get_ex);
378 
ib_umem_get(struct ib_ucontext * context,unsigned long addr,size_t size,int access,int dmasync)379 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
380 			    size_t size, int access, int dmasync)
381 {
382 	return ib_umem_get_ex(context, addr,
383 			    size, access, dmasync, 0);
384 }
385 EXPORT_SYMBOL(ib_umem_get);
386 
387 /**
388  * ib_umem_release - release memory pinned with ib_umem_get
389  * @umem: umem struct to release
390  */
ib_umem_release(struct ib_umem * umem)391 void ib_umem_release(struct ib_umem *umem)
392 {
393 
394 	vm_offset_t addr, end, last, start;
395 	vm_size_t size;
396 	int error;
397 
398 	if (umem->ib_peer_mem) {
399 		peer_umem_release(umem);
400 		return;
401 	}
402 
403 	__ib_umem_release(umem->context->device, umem, 1);
404 
405 	if (umem->context->closing) {
406 		kfree(umem);
407 		return;
408 	}
409 
410 	error = priv_check(curthread, PRIV_VM_MUNLOCK);
411 
412 	if (error)
413 		return;
414 
415 	addr = umem->start;
416 	size = umem->length;
417 	last = addr + size;
418         start = addr & PAGE_MASK; /* Use the linux PAGE_MASK definition. */
419 	end = roundup2(last, PAGE_SIZE); /* Use PAGE_MASK safe operation. */
420 	vm_map_unwire(&curthread->td_proc->p_vmspace->vm_map, start, end,
421 	    VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES);
422 	kfree(umem);
423 
424 }
425 EXPORT_SYMBOL(ib_umem_release);
426 
ib_umem_page_count(struct ib_umem * umem)427 int ib_umem_page_count(struct ib_umem *umem)
428 {
429 	int shift;
430 	int i;
431 	int n;
432 	struct scatterlist *sg;
433 
434 	shift = ilog2(umem->page_size);
435 
436 	n = 0;
437 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
438 		n += sg_dma_len(sg) >> shift;
439 
440 	return n;
441 }
442 EXPORT_SYMBOL(ib_umem_page_count);
443