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.
31 * This file is part of Lustre, http://www.lustre.org/
32 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Memory Descriptor management routines
39 #define DEBUG_SUBSYSTEM S_LNET
41 #include <lnet/lib-lnet.h>
43 /* must be called with lnet_res_lock held */
45 lnet_md_unlink(lnet_libmd_t *md)
47 if ((md->md_flags & LNET_MD_FLAG_ZOMBIE) == 0) {
48 /* first unlink attempt... */
49 lnet_me_t *me = md->md_me;
51 md->md_flags |= LNET_MD_FLAG_ZOMBIE;
53 /* Disassociate from ME (if any), and unlink it if it was created
56 /* detach MD from portal */
57 lnet_ptl_detach_md(me, md);
58 if (me->me_unlink == LNET_UNLINK)
62 /* ensure all future handle lookups fail */
63 lnet_res_lh_invalidate(&md->md_lh);
66 if (md->md_refcount != 0) {
67 CDEBUG(D_NET, "Queueing unlink of md %p\n", md);
71 CDEBUG(D_NET, "Unlinking md %p\n", md);
73 if (md->md_eq != NULL) {
74 int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
76 LASSERT(*md->md_eq->eq_refs[cpt] > 0);
77 (*md->md_eq->eq_refs[cpt])--;
80 LASSERT(!cfs_list_empty(&md->md_list));
81 cfs_list_del_init(&md->md_list);
82 lnet_md_free_locked(md);
86 lnet_md_build(lnet_libmd_t *lmd, lnet_md_t *umd, int unlink)
93 lmd->md_start = umd->start;
95 lmd->md_max_size = umd->max_size;
96 lmd->md_options = umd->options;
97 lmd->md_user_ptr = umd->user_ptr;
99 lmd->md_threshold = umd->threshold;
100 lmd->md_refcount = 0;
101 lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
103 if ((umd->options & LNET_MD_IOVEC) != 0) {
105 if ((umd->options & LNET_MD_KIOV) != 0) /* Can't specify both */
108 lmd->md_niov = niov = umd->length;
109 memcpy(lmd->md_iov.iov, umd->start,
110 niov * sizeof (lmd->md_iov.iov[0]));
112 for (i = 0; i < (int)niov; i++) {
113 /* We take the base address on trust */
114 if (lmd->md_iov.iov[i].iov_len <= 0) /* invalid length */
117 total_length += lmd->md_iov.iov[i].iov_len;
120 lmd->md_length = total_length;
122 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
123 (umd->max_size < 0 ||
124 umd->max_size > total_length)) // illegal max_size
127 } else if ((umd->options & LNET_MD_KIOV) != 0) {
131 lmd->md_niov = niov = umd->length;
132 memcpy(lmd->md_iov.kiov, umd->start,
133 niov * sizeof (lmd->md_iov.kiov[0]));
135 for (i = 0; i < (int)niov; i++) {
136 /* We take the page pointer on trust */
137 if (lmd->md_iov.kiov[i].kiov_offset +
138 lmd->md_iov.kiov[i].kiov_len > CFS_PAGE_SIZE )
139 return -EINVAL; /* invalid length */
141 total_length += lmd->md_iov.kiov[i].kiov_len;
144 lmd->md_length = total_length;
146 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
147 (umd->max_size < 0 ||
148 umd->max_size > total_length)) // illegal max_size
151 } else { /* contiguous */
152 lmd->md_length = umd->length;
153 lmd->md_niov = niov = 1;
154 lmd->md_iov.iov[0].iov_base = umd->start;
155 lmd->md_iov.iov[0].iov_len = umd->length;
157 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
158 (umd->max_size < 0 ||
159 umd->max_size > (int)umd->length)) // illegal max_size
166 /* must be called with resource lock held */
168 lnet_md_link(lnet_libmd_t *md, lnet_handle_eq_t eq_handle, int cpt)
170 struct lnet_res_container *container = the_lnet.ln_md_containers[cpt];
172 /* NB we are passed an allocated, but inactive md.
173 * if we return success, caller may lnet_md_unlink() it.
174 * otherwise caller may only lnet_md_free() it.
176 /* This implementation doesn't know how to create START events or
177 * disable END events. Best to LASSERT our caller is compliant so
178 * we find out quickly... */
179 /* TODO - reevaluate what should be here in light of
180 * the removal of the start and end events
181 * maybe there we shouldn't even allow LNET_EQ_NONE!)
182 * LASSERT (eq == NULL);
184 if (!LNetHandleIsInvalid(eq_handle)) {
185 md->md_eq = lnet_handle2eq(&eq_handle);
187 if (md->md_eq == NULL)
190 (*md->md_eq->eq_refs[cpt])++;
193 lnet_res_lh_initialize(container, &md->md_lh);
195 LASSERT(cfs_list_empty(&md->md_list));
196 cfs_list_add(&md->md_list, &container->rec_active);
201 /* must be called with lnet_res_lock held */
203 lnet_md_deconstruct(lnet_libmd_t *lmd, lnet_md_t *umd)
205 /* NB this doesn't copy out all the iov entries so when a
206 * discontiguous MD is copied out, the target gets to know the
207 * original iov pointer (in start) and the number of entries it had
210 umd->start = lmd->md_start;
211 umd->length = ((lmd->md_options & (LNET_MD_IOVEC | LNET_MD_KIOV)) == 0) ?
212 lmd->md_length : lmd->md_niov;
213 umd->threshold = lmd->md_threshold;
214 umd->max_size = lmd->md_max_size;
215 umd->options = lmd->md_options;
216 umd->user_ptr = lmd->md_user_ptr;
217 lnet_eq2handle(&umd->eq_handle, lmd->md_eq);
221 lnet_md_validate(lnet_md_t *umd)
223 if (umd->start == NULL && umd->length != 0) {
224 CERROR("MD start pointer can not be NULL with length %u\n",
229 if ((umd->options & (LNET_MD_KIOV | LNET_MD_IOVEC)) != 0 &&
230 umd->length > LNET_MAX_IOV) {
231 CERROR("Invalid option: too many fragments %u, %d max\n",
232 umd->length, LNET_MAX_IOV);
240 * Create a memory descriptor and attach it to a ME
242 * \param meh A handle for a ME to associate the new MD with.
243 * \param umd Provides initial values for the user-visible parts of a MD.
244 * Other than its use for initialization, there is no linkage between this
245 * structure and the MD maintained by the LNet.
246 * \param unlink A flag to indicate whether the MD is automatically unlinked
247 * when it becomes inactive, either because the operation threshold drops to
248 * zero or because the available memory becomes less than \a umd.max_size.
249 * (Note that the check for unlinking a MD only occurs after the completion
250 * of a successful operation on the MD.) The value LNET_UNLINK enables auto
251 * unlinking; the value LNET_RETAIN disables it.
252 * \param handle On successful returns, a handle to the newly created MD is
253 * saved here. This handle can be used later in LNetMDUnlink().
255 * \retval 0 On success.
256 * \retval -EINVAL If \a umd is not valid.
257 * \retval -ENOMEM If new MD cannot be allocated.
258 * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a
259 * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by
260 * calling LNetInvalidateHandle() on it.
261 * \retval -EBUSY If the ME pointed to by \a meh is already associated with
265 LNetMDAttach(lnet_handle_me_t meh, lnet_md_t umd,
266 lnet_unlink_t unlink, lnet_handle_md_t *handle)
268 CFS_LIST_HEAD (matches);
269 CFS_LIST_HEAD (drops);
271 struct lnet_libmd *md;
275 LASSERT (the_lnet.ln_init);
276 LASSERT (the_lnet.ln_refcount > 0);
278 if (lnet_md_validate(&umd) != 0)
281 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) == 0) {
282 CERROR("Invalid option: no MD_OP set\n");
286 md = lnet_md_alloc(&umd);
290 rc = lnet_md_build(md, &umd, unlink);
291 cpt = lnet_cpt_of_cookie(meh.cookie);
297 me = lnet_handle2me(&meh);
300 else if (me->me_md != NULL)
303 rc = lnet_md_link(md, umd.eq_handle, cpt);
308 /* attach this MD to portal of ME and check if it matches any
309 * blocked msgs on this portal */
310 lnet_ptl_attach_md(me, md, &matches, &drops);
312 lnet_md2handle(handle, md);
314 lnet_res_unlock(cpt);
316 lnet_drop_delayed_msg_list(&drops, "Bad match");
317 lnet_recv_delayed_msg_list(&matches);
322 lnet_md_free_locked(md);
324 lnet_res_unlock(cpt);
329 * Create a "free floating" memory descriptor - a MD that is not associated
330 * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
332 * \param umd,unlink See the discussion for LNetMDAttach().
333 * \param handle On successful returns, a handle to the newly created MD is
334 * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(),
335 * and LNetGet() operations.
337 * \retval 0 On success.
338 * \retval -EINVAL If \a umd is not valid.
339 * \retval -ENOMEM If new MD cannot be allocated.
340 * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that
341 * it's OK to supply a NULL \a umd.eq_handle by calling
342 * LNetInvalidateHandle() on it.
345 LNetMDBind(lnet_md_t umd, lnet_unlink_t unlink, lnet_handle_md_t *handle)
351 LASSERT (the_lnet.ln_init);
352 LASSERT (the_lnet.ln_refcount > 0);
354 if (lnet_md_validate(&umd) != 0)
357 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) != 0) {
358 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
362 md = lnet_md_alloc(&umd);
366 rc = lnet_md_build(md, &umd, unlink);
368 cpt = lnet_res_lock_current();
372 rc = lnet_md_link(md, umd.eq_handle, cpt);
376 lnet_md2handle(handle, md);
378 lnet_res_unlock(cpt);
382 lnet_md_free_locked(md);
384 lnet_res_unlock(cpt);
389 * Unlink the memory descriptor from any ME it may be linked to and release
390 * the internal resources associated with it.
392 * This function does not free the memory region associated with the MD;
393 * i.e., the memory the user allocated for this MD. If the ME associated with
394 * this MD is not NULL and was created with auto unlink enabled, the ME is
395 * unlinked as well (see LNetMEAttach()).
397 * Explicitly unlinking a MD via this function call has the same behavior as
398 * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
399 * is generated in the latter case.
401 * An unlinked event can be reported in two ways:
402 * - If there's no pending operations on the MD, it's unlinked immediately
403 * and an LNET_EVENT_UNLINK event is logged before this function returns.
404 * - Otherwise, the MD is only marked for deletion when this function
405 * returns, and the unlinked event will be piggybacked on the event of
406 * the completion of the last operation by setting the unlinked field of
407 * the event. No dedicated LNET_EVENT_UNLINK event is generated.
409 * Note that in both cases the unlinked field of the event is always set; no
410 * more event will happen on the MD after such an event is logged.
412 * \param mdh A handle for the MD to be unlinked.
414 * \retval 0 On success.
415 * \retval -ENOENT If \a mdh does not point to a valid MD object.
418 LNetMDUnlink (lnet_handle_md_t mdh)
424 LASSERT(the_lnet.ln_init);
425 LASSERT(the_lnet.ln_refcount > 0);
427 cpt = lnet_cpt_of_cookie(mdh.cookie);
430 md = lnet_handle2md(&mdh);
432 lnet_res_unlock(cpt);
436 /* If the MD is busy, lnet_md_unlink just marks it for deletion, and
437 * when the NAL is done, the completion event flags that the MD was
438 * unlinked. Otherwise, we enqueue an event now... */
440 if (md->md_eq != NULL &&
441 md->md_refcount == 0) {
442 lnet_build_unlink_event(md, &ev);
443 lnet_eq_enqueue_event(md->md_eq, &ev);
448 lnet_res_unlock(cpt);