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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Memory Descriptor management routines
41 #define DEBUG_SUBSYSTEM S_LNET
43 #include <lnet/lib-lnet.h>
45 /* must be called with lnet_res_lock held */
47 lnet_md_unlink(lnet_libmd_t *md)
49 if ((md->md_flags & LNET_MD_FLAG_ZOMBIE) == 0) {
50 /* first unlink attempt... */
51 lnet_me_t *me = md->md_me;
53 md->md_flags |= LNET_MD_FLAG_ZOMBIE;
55 /* Disassociate from ME (if any), and unlink it if it was created
58 /* detach MD from portal */
59 lnet_ptl_detach_md(me, md);
60 if (me->me_unlink == LNET_UNLINK)
64 /* ensure all future handle lookups fail */
65 lnet_res_lh_invalidate(&md->md_lh);
68 if (md->md_refcount != 0) {
69 CDEBUG(D_NET, "Queueing unlink of md %p\n", md);
73 CDEBUG(D_NET, "Unlinking md %p\n", md);
75 if (md->md_eq != NULL) {
76 int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
78 LASSERT(*md->md_eq->eq_refs[cpt] > 0);
79 (*md->md_eq->eq_refs[cpt])--;
82 LASSERT(!list_empty(&md->md_list));
83 list_del_init(&md->md_list);
84 lnet_md_free_locked(md);
88 lnet_md_build(lnet_libmd_t *lmd, lnet_md_t *umd, int unlink)
95 lmd->md_start = umd->start;
97 lmd->md_max_size = umd->max_size;
98 lmd->md_options = umd->options;
99 lmd->md_user_ptr = umd->user_ptr;
101 lmd->md_threshold = umd->threshold;
102 lmd->md_refcount = 0;
103 lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
105 if ((umd->options & LNET_MD_IOVEC) != 0) {
107 if ((umd->options & LNET_MD_KIOV) != 0) /* Can't specify both */
110 lmd->md_niov = niov = umd->length;
111 memcpy(lmd->md_iov.iov, umd->start,
112 niov * sizeof (lmd->md_iov.iov[0]));
114 for (i = 0; i < (int)niov; i++) {
115 /* We take the base address on trust */
116 if (lmd->md_iov.iov[i].iov_len <= 0) /* invalid length */
119 total_length += lmd->md_iov.iov[i].iov_len;
122 lmd->md_length = total_length;
124 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
125 (umd->max_size < 0 ||
126 umd->max_size > total_length)) // illegal max_size
129 } else if ((umd->options & LNET_MD_KIOV) != 0) {
133 lmd->md_niov = niov = umd->length;
134 memcpy(lmd->md_iov.kiov, umd->start,
135 niov * sizeof (lmd->md_iov.kiov[0]));
137 for (i = 0; i < (int)niov; i++) {
138 /* We take the page pointer on trust */
139 if (lmd->md_iov.kiov[i].kiov_offset +
140 lmd->md_iov.kiov[i].kiov_len > PAGE_CACHE_SIZE)
141 return -EINVAL; /* invalid length */
143 total_length += lmd->md_iov.kiov[i].kiov_len;
146 lmd->md_length = total_length;
148 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
149 (umd->max_size < 0 ||
150 umd->max_size > total_length)) // illegal max_size
153 } else { /* contiguous */
154 lmd->md_length = umd->length;
155 lmd->md_niov = niov = 1;
156 lmd->md_iov.iov[0].iov_base = umd->start;
157 lmd->md_iov.iov[0].iov_len = umd->length;
159 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
160 (umd->max_size < 0 ||
161 umd->max_size > (int)umd->length)) // illegal max_size
168 /* must be called with resource lock held */
170 lnet_md_link(lnet_libmd_t *md, lnet_handle_eq_t eq_handle, int cpt)
172 struct lnet_res_container *container = the_lnet.ln_md_containers[cpt];
174 /* NB we are passed an allocated, but inactive md.
175 * if we return success, caller may lnet_md_unlink() it.
176 * otherwise caller may only lnet_md_free() it.
178 /* This implementation doesn't know how to create START events or
179 * disable END events. Best to LASSERT our caller is compliant so
180 * we find out quickly... */
181 /* TODO - reevaluate what should be here in light of
182 * the removal of the start and end events
183 * maybe there we shouldn't even allow LNET_EQ_NONE!)
184 * LASSERT (eq == NULL);
186 if (!LNetHandleIsInvalid(eq_handle)) {
187 md->md_eq = lnet_handle2eq(&eq_handle);
189 if (md->md_eq == NULL)
192 (*md->md_eq->eq_refs[cpt])++;
195 lnet_res_lh_initialize(container, &md->md_lh);
197 LASSERT(list_empty(&md->md_list));
198 list_add(&md->md_list, &container->rec_active);
203 /* must be called with lnet_res_lock held */
205 lnet_md_deconstruct(lnet_libmd_t *lmd, lnet_md_t *umd)
207 /* NB this doesn't copy out all the iov entries so when a
208 * discontiguous MD is copied out, the target gets to know the
209 * original iov pointer (in start) and the number of entries it had
212 umd->start = lmd->md_start;
213 umd->length = ((lmd->md_options & (LNET_MD_IOVEC | LNET_MD_KIOV)) == 0) ?
214 lmd->md_length : lmd->md_niov;
215 umd->threshold = lmd->md_threshold;
216 umd->max_size = lmd->md_max_size;
217 umd->options = lmd->md_options;
218 umd->user_ptr = lmd->md_user_ptr;
219 lnet_eq2handle(&umd->eq_handle, lmd->md_eq);
223 lnet_md_validate(lnet_md_t *umd)
225 if (umd->start == NULL && umd->length != 0) {
226 CERROR("MD start pointer can not be NULL with length %u\n",
231 if ((umd->options & (LNET_MD_KIOV | LNET_MD_IOVEC)) != 0 &&
232 umd->length > LNET_MAX_IOV) {
233 CERROR("Invalid option: too many fragments %u, %d max\n",
234 umd->length, LNET_MAX_IOV);
242 * Create a memory descriptor and attach it to a ME
244 * \param meh A handle for a ME to associate the new MD with.
245 * \param umd Provides initial values for the user-visible parts of a MD.
246 * Other than its use for initialization, there is no linkage between this
247 * structure and the MD maintained by the LNet.
248 * \param unlink A flag to indicate whether the MD is automatically unlinked
249 * when it becomes inactive, either because the operation threshold drops to
250 * zero or because the available memory becomes less than \a umd.max_size.
251 * (Note that the check for unlinking a MD only occurs after the completion
252 * of a successful operation on the MD.) The value LNET_UNLINK enables auto
253 * unlinking; the value LNET_RETAIN disables it.
254 * \param handle On successful returns, a handle to the newly created MD is
255 * saved here. This handle can be used later in LNetMDUnlink().
257 * \retval 0 On success.
258 * \retval -EINVAL If \a umd is not valid.
259 * \retval -ENOMEM If new MD cannot be allocated.
260 * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a
261 * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by
262 * calling LNetInvalidateHandle() on it.
263 * \retval -EBUSY If the ME pointed to by \a meh is already associated with
267 LNetMDAttach(lnet_handle_me_t meh, lnet_md_t umd,
268 lnet_unlink_t unlink, lnet_handle_md_t *handle)
270 struct list_head matches = LIST_HEAD_INIT(matches);
271 struct list_head drops = LIST_HEAD_INIT(drops);
273 struct lnet_libmd *md;
277 LASSERT (the_lnet.ln_init);
278 LASSERT (the_lnet.ln_refcount > 0);
280 if (lnet_md_validate(&umd) != 0)
283 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) == 0) {
284 CERROR("Invalid option: no MD_OP set\n");
288 md = lnet_md_alloc(&umd);
292 rc = lnet_md_build(md, &umd, unlink);
293 cpt = lnet_cpt_of_cookie(meh.cookie);
299 me = lnet_handle2me(&meh);
302 else if (me->me_md != NULL)
305 rc = lnet_md_link(md, umd.eq_handle, cpt);
310 /* attach this MD to portal of ME and check if it matches any
311 * blocked msgs on this portal */
312 lnet_ptl_attach_md(me, md, &matches, &drops);
314 lnet_md2handle(handle, md);
316 lnet_res_unlock(cpt);
318 lnet_drop_delayed_msg_list(&drops, "Bad match");
319 lnet_recv_delayed_msg_list(&matches);
324 lnet_md_free_locked(md);
326 lnet_res_unlock(cpt);
329 EXPORT_SYMBOL(LNetMDAttach);
332 * Create a "free floating" memory descriptor - a MD that is not associated
333 * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
335 * \param umd,unlink See the discussion for LNetMDAttach().
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(), LNetPut(),
338 * and LNetGet() operations.
340 * \retval 0 On success.
341 * \retval -EINVAL If \a umd is not valid.
342 * \retval -ENOMEM If new MD cannot be allocated.
343 * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that
344 * it's OK to supply a NULL \a umd.eq_handle by calling
345 * LNetInvalidateHandle() on it.
348 LNetMDBind(lnet_md_t umd, lnet_unlink_t unlink, lnet_handle_md_t *handle)
354 LASSERT (the_lnet.ln_init);
355 LASSERT (the_lnet.ln_refcount > 0);
357 if (lnet_md_validate(&umd) != 0)
360 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) != 0) {
361 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
365 md = lnet_md_alloc(&umd);
369 rc = lnet_md_build(md, &umd, unlink);
371 cpt = lnet_res_lock_current();
375 rc = lnet_md_link(md, umd.eq_handle, cpt);
379 lnet_md2handle(handle, md);
381 lnet_res_unlock(cpt);
385 lnet_md_free_locked(md);
387 lnet_res_unlock(cpt);
390 EXPORT_SYMBOL(LNetMDBind);
393 * Unlink the memory descriptor from any ME it may be linked to and release
394 * the internal resources associated with it. As a result, active messages
395 * associated with the MD may get aborted.
397 * This function does not free the memory region associated with the MD;
398 * i.e., the memory the user allocated for this MD. If the ME associated with
399 * this MD is not NULL and was created with auto unlink enabled, the ME is
400 * unlinked as well (see LNetMEAttach()).
402 * Explicitly unlinking a MD via this function call has the same behavior as
403 * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
404 * is generated in the latter case.
406 * An unlinked event can be reported in two ways:
407 * - If there's no pending operations on the MD, it's unlinked immediately
408 * and an LNET_EVENT_UNLINK event is logged before this function returns.
409 * - Otherwise, the MD is only marked for deletion when this function
410 * returns, and the unlinked event will be piggybacked on the event of
411 * the completion of the last operation by setting the unlinked field of
412 * the event. No dedicated LNET_EVENT_UNLINK event is generated.
414 * Note that in both cases the unlinked field of the event is always set; no
415 * more event will happen on the MD after such an event is logged.
417 * \param mdh A handle for the MD to be unlinked.
419 * \retval 0 On success.
420 * \retval -ENOENT If \a mdh does not point to a valid MD object.
423 LNetMDUnlink (lnet_handle_md_t mdh)
429 LASSERT(the_lnet.ln_init);
430 LASSERT(the_lnet.ln_refcount > 0);
432 cpt = lnet_cpt_of_cookie(mdh.cookie);
435 md = lnet_handle2md(&mdh);
437 lnet_res_unlock(cpt);
441 md->md_flags |= LNET_MD_FLAG_ABORTED;
442 /* If the MD is busy, lnet_md_unlink just marks it for deletion, and
443 * when the LND is done, the completion event flags that the MD was
444 * unlinked. Otherwise, we enqueue an event now... */
445 if (md->md_eq != NULL && md->md_refcount == 0) {
446 lnet_build_unlink_event(md, &ev);
447 lnet_eq_enqueue_event(md->md_eq, &ev);
452 lnet_res_unlock(cpt);
455 EXPORT_SYMBOL(LNetMDUnlink);