4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * Memory Descriptor management routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <lnet/lib-lnet.h>
41 /* must be called with lnet_res_lock held */
43 lnet_md_unlink(struct lnet_libmd *md)
45 if ((md->md_flags & LNET_MD_FLAG_ZOMBIE) == 0) {
46 /* first unlink attempt... */
47 struct lnet_me *me = md->md_me;
49 md->md_flags |= LNET_MD_FLAG_ZOMBIE;
51 /* Disassociate from ME (if any), and unlink it if it was created
54 /* detach MD from portal */
55 lnet_ptl_detach_md(me, md);
56 if (me->me_unlink == LNET_UNLINK)
60 /* ensure all future handle lookups fail */
61 lnet_res_lh_invalidate(&md->md_lh);
64 if (md->md_refcount != 0) {
65 CDEBUG(D_NET, "Queueing unlink of md %p\n", md);
69 CDEBUG(D_NET, "Unlinking md %p\n", md);
71 LASSERT(!list_empty(&md->md_list));
72 list_del_init(&md->md_list);
73 LASSERT(!(md->md_flags & LNET_MD_FLAG_HANDLING));
78 lnet_kvaddr_to_page(unsigned long vaddr)
80 if (is_vmalloc_addr((void *)vaddr))
81 return vmalloc_to_page((void *)vaddr);
85 #ifdef HAVE_KMAP_TO_PAGE
87 * This ifdef is added to handle the kernel versions
88 * which have kmap_to_page() function exported. If so,
89 * we should use it. Otherwise, remain with the legacy check.
91 return kmap_to_page((void *)vaddr);
94 if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) {
95 /* No highmem pages only used for bulk (kiov) I/O */
96 CERROR("find page for address in highmem\n");
99 return virt_to_page(vaddr);
100 #endif /* HAVE_KMAP_TO_PAGE */
103 return virt_to_page(vaddr);
104 #endif /* CONFIG_HIGHMEM */
106 EXPORT_SYMBOL(lnet_kvaddr_to_page);
109 lnet_cpt_of_md(struct lnet_libmd *md, unsigned int offset)
111 int cpt = CFS_CPT_ANY;
113 struct bio_vec *kiov;
116 * if the md_options has a bulk handle then we want to look at the
117 * bulk md because that's the data which we will be DMAing
119 if (md && (md->md_options & LNET_MD_BULK_HANDLE) != 0 &&
120 !LNetMDHandleIsInvalid(md->md_bulk_handle))
121 md = lnet_handle2md(&md->md_bulk_handle);
123 if (!md || md->md_niov == 0)
129 while (offset >= kiov->bv_len) {
130 offset -= kiov->bv_len;
134 CERROR("offset %d goes beyond kiov\n", offset);
139 cpt = cfs_cpt_of_node(lnet_cpt_table(),
140 page_to_nid(kiov->bv_page));
145 static int lnet_md_validate(const struct lnet_md *umd);
147 static struct lnet_libmd *
148 lnet_md_build(const struct lnet_md *umd, int unlink)
152 int total_length = 0;
153 struct lnet_libmd *lmd;
156 if (lnet_md_validate(umd) != 0)
157 return ERR_PTR(-EINVAL);
159 if (umd->options & LNET_MD_KIOV)
162 niov = DIV_ROUND_UP(offset_in_page(umd->start) + umd->length,
164 size = offsetof(struct lnet_libmd, md_kiov[niov]);
166 if (size <= LNET_SMALL_MD_SIZE) {
167 lmd = kmem_cache_zalloc(lnet_small_mds_cachep, GFP_NOFS);
170 "slab-alloced 'md' of size %u at %p.\n",
173 CDEBUG(D_MALLOC, "failed to allocate 'md' of size %u\n",
177 LIBCFS_ALLOC(lmd, size);
181 return ERR_PTR(-ENOMEM);
184 INIT_LIST_HEAD(&lmd->md_list);
186 lmd->md_start = umd->start;
188 lmd->md_max_size = umd->max_size;
189 lmd->md_options = umd->options;
190 lmd->md_user_ptr = umd->user_ptr;
191 lmd->md_handler = NULL;
192 lmd->md_threshold = umd->threshold;
193 lmd->md_refcount = 0;
194 lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
195 lmd->md_bulk_handle = umd->bulk_handle;
197 if (umd->options & LNET_MD_KIOV) {
198 memcpy(lmd->md_kiov, umd->start,
199 niov * sizeof(lmd->md_kiov[0]));
201 for (i = 0; i < (int)niov; i++) {
202 /* We take the page pointer on trust */
203 if (lmd->md_kiov[i].bv_offset +
204 lmd->md_kiov[i].bv_len > PAGE_SIZE) {
206 return ERR_PTR(-EINVAL); /* invalid length */
209 total_length += lmd->md_kiov[i].bv_len;
212 lmd->md_length = total_length;
214 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
215 (umd->max_size < 0 ||
216 umd->max_size > total_length)) { /* illegal max_size */
218 return ERR_PTR(-EINVAL);
220 } else { /* contiguous - split into pages */
221 void *pa = umd->start;
222 int len = umd->length;
224 lmd->md_length = len;
229 plen = min_t(int, len, PAGE_SIZE - offset_in_page(pa));
231 lmd->md_kiov[i].bv_page =
232 lnet_kvaddr_to_page((unsigned long) pa);
233 lmd->md_kiov[i].bv_offset = offset_in_page(pa);
234 lmd->md_kiov[i].bv_len = plen;
240 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
241 (umd->max_size < 0 ||
242 umd->max_size > (int)umd->length)) { /* illegal max_size */
244 return ERR_PTR(-EINVAL);
246 lmd->md_options |= LNET_MD_KIOV;
252 /* must be called with resource lock held */
254 lnet_md_link(struct lnet_libmd *md, lnet_handler_t handler, int cpt)
256 struct lnet_res_container *container = the_lnet.ln_md_containers[cpt];
258 /* NB we are passed an allocated, but inactive md.
259 * Caller may lnet_md_unlink() it, or may lnet_md_free() it.
261 /* This implementation doesn't know how to create START events or
262 * disable END events. Best to LASSERT our caller is compliant so
263 * we find out quickly... */
264 /* TODO - reevaluate what should be here in light of
265 * the removal of the start and end events
266 * maybe there we shouldn't even allow LNET_EQ_NONE!)
267 * LASSERT (handler != NULL);
269 md->md_handler = handler;
271 lnet_res_lh_initialize(container, &md->md_lh);
273 LASSERT(list_empty(&md->md_list));
274 list_add(&md->md_list, &container->rec_active);
277 void lnet_assert_handler_unused(lnet_handler_t handler)
279 struct lnet_res_container *container;
284 cfs_percpt_for_each(container, cpt, the_lnet.ln_md_containers) {
285 struct lnet_libmd *md;
288 list_for_each_entry(md, &container->rec_active, md_list)
289 LASSERT(md->md_handler != handler);
290 lnet_res_unlock(cpt);
293 EXPORT_SYMBOL(lnet_assert_handler_unused);
295 /* must be called with lnet_res_lock held */
297 lnet_md_deconstruct(struct lnet_libmd *lmd, struct lnet_event *ev)
299 ev->md_start = lmd->md_start;
300 ev->md_options = lmd->md_options;
301 ev->md_user_ptr = lmd->md_user_ptr;
305 lnet_md_validate(const struct lnet_md *umd)
307 if (umd->start == NULL && umd->length != 0) {
308 CERROR("MD start pointer can not be NULL with length %u\n",
313 if ((umd->options & LNET_MD_KIOV) &&
314 umd->length > LNET_MAX_IOV) {
315 CERROR("Invalid option: too many fragments %u, %d max\n",
316 umd->length, LNET_MAX_IOV);
324 * Create a memory descriptor and attach it to a ME
326 * \param me An ME to associate the new MD with.
327 * \param umd Provides initial values for the user-visible parts of a MD.
328 * Other than its use for initialization, there is no linkage between this
329 * structure and the MD maintained by the LNet.
330 * \param unlink A flag to indicate whether the MD is automatically unlinked
331 * when it becomes inactive, either because the operation threshold drops to
332 * zero or because the available memory becomes less than \a umd.max_size.
333 * (Note that the check for unlinking a MD only occurs after the completion
334 * of a successful operation on the MD.) The value LNET_UNLINK enables auto
335 * unlinking; the value LNET_RETAIN disables it.
336 * \param handle On successful returns, a handle to the newly created MD is
337 * saved here. This handle can be used later in LNetMDUnlink().
339 * The ME will either be linked to the new MD, or it will be freed.
341 * \retval 0 On success.
342 * \retval -EINVAL If \a umd is not valid.
343 * \retval -ENOMEM If new MD cannot be allocated.
346 LNetMDAttach(struct lnet_me *me, const struct lnet_md *umd,
347 enum lnet_unlink unlink, struct lnet_handle_md *handle)
351 struct lnet_libmd *md;
354 LASSERT(the_lnet.ln_refcount > 0);
357 if ((umd->options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) == 0) {
358 CERROR("Invalid option: no MD_OP set\n");
359 md = ERR_PTR(-EINVAL);
361 md = lnet_md_build(umd, unlink);
368 lnet_res_unlock(cpt);
372 lnet_md_link(md, umd->handler, cpt);
374 /* attach this MD to portal of ME and check if it matches any
375 * blocked msgs on this portal */
376 lnet_ptl_attach_md(me, md, &matches, &drops);
378 lnet_md2handle(handle, md);
380 lnet_res_unlock(cpt);
382 lnet_drop_delayed_msg_list(&drops, "Bad match");
383 lnet_recv_delayed_msg_list(&matches);
387 EXPORT_SYMBOL(LNetMDAttach);
390 * Create a "free floating" memory descriptor - a MD that is not associated
391 * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
393 * \param umd,unlink See the discussion for LNetMDAttach().
394 * \param handle On successful returns, a handle to the newly created MD is
395 * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(),
396 * and LNetGet() operations.
398 * \retval 0 On success.
399 * \retval -EINVAL If \a umd is not valid.
400 * \retval -ENOMEM If new MD cannot be allocated.
403 LNetMDBind(const struct lnet_md *umd, enum lnet_unlink unlink,
404 struct lnet_handle_md *handle)
406 struct lnet_libmd *md;
410 LASSERT(the_lnet.ln_refcount > 0);
412 if ((umd->options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) != 0) {
413 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
417 md = lnet_md_build(umd, unlink);
421 if (md->md_length > LNET_MTU) {
422 CERROR("Invalid length: too big transfer size %u, %d max\n",
423 md->md_length, LNET_MTU);
428 cpt = lnet_res_lock_current();
430 lnet_md_link(md, umd->handler, cpt);
432 lnet_md2handle(handle, md);
434 lnet_res_unlock(cpt);
441 EXPORT_SYMBOL(LNetMDBind);
444 * Unlink the memory descriptor from any ME it may be linked to and release
445 * the internal resources associated with it. As a result, active messages
446 * associated with the MD may get aborted.
448 * This function does not free the memory region associated with the MD;
449 * i.e., the memory the user allocated for this MD. If the ME associated with
450 * this MD is not NULL and was created with auto unlink enabled, the ME is
451 * unlinked as well (see LNetMEAttach()).
453 * Explicitly unlinking a MD via this function call has the same behavior as
454 * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
455 * is generated in the latter case.
457 * An unlinked event can be reported in two ways:
458 * - If there's no pending operations on the MD, it's unlinked immediately
459 * and an LNET_EVENT_UNLINK event is logged before this function returns.
460 * - Otherwise, the MD is only marked for deletion when this function
461 * returns, and the unlinked event will be piggybacked on the event of
462 * the completion of the last operation by setting the unlinked field of
463 * the event. No dedicated LNET_EVENT_UNLINK event is generated.
465 * Note that in both cases the unlinked field of the event is always set; no
466 * more event will happen on the MD after such an event is logged.
468 * \param mdh A handle for the MD to be unlinked.
470 * \retval 0 On success.
471 * \retval -ENOENT If \a mdh does not point to a valid MD object.
474 __LNetMDUnlink(struct lnet_handle_md mdh, bool discard)
476 struct lnet_event ev;
477 struct lnet_libmd *md = NULL;
478 lnet_handler_t handler = NULL;
481 LASSERT(the_lnet.ln_refcount > 0);
483 cpt = lnet_cpt_of_cookie(mdh.cookie);
486 md = lnet_handle2md(&mdh);
488 lnet_res_unlock(cpt);
491 if (md->md_refcount == 0 &&
492 md->md_flags & LNET_MD_FLAG_HANDLING) {
493 /* Race with unlocked call to ->md_handler. */
494 lnet_md_wait_handling(md, cpt);
499 md->md_flags |= LNET_MD_FLAG_ABORTED;
500 /* If the MD is busy, lnet_md_unlink just marks it for deletion, and
501 * when the LND is done, the completion event flags that the MD was
502 * unlinked. Otherwise, we enqueue an event now... */
503 if (md->md_handler && md->md_refcount == 0) {
504 lnet_build_unlink_event(md, &ev);
505 handler = md->md_handler;
509 md->md_flags |= LNET_MD_FLAG_DISCARD;
511 if (md->md_rspt_ptr != NULL)
512 lnet_detach_rsp_tracker(md, cpt);
516 lnet_res_unlock(cpt);
523 EXPORT_SYMBOL(__LNetMDUnlink);
526 lnet_md_discarded(struct lnet_libmd *md)
534 cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
536 rc = md->md_flags & LNET_MD_FLAG_DISCARD;
537 lnet_res_unlock(cpt);
541 EXPORT_SYMBOL(lnet_md_discarded);