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_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 md->md_eq->eq_refcount--;
75 LASSERT (md->md_eq->eq_refcount >= 0);
78 LASSERT (!cfs_list_empty(&md->md_list));
79 cfs_list_del_init (&md->md_list);
80 lnet_md_free_locked(md);
84 lnet_md_build(lnet_libmd_t *lmd, lnet_md_t *umd, int unlink)
91 lmd->md_start = umd->start;
93 lmd->md_max_size = umd->max_size;
94 lmd->md_options = umd->options;
95 lmd->md_user_ptr = umd->user_ptr;
97 lmd->md_threshold = umd->threshold;
99 lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
101 if ((umd->options & LNET_MD_IOVEC) != 0) {
103 if ((umd->options & LNET_MD_KIOV) != 0) /* Can't specify both */
106 lmd->md_niov = niov = umd->length;
107 memcpy(lmd->md_iov.iov, umd->start,
108 niov * sizeof (lmd->md_iov.iov[0]));
110 for (i = 0; i < (int)niov; i++) {
111 /* We take the base address on trust */
112 if (lmd->md_iov.iov[i].iov_len <= 0) /* invalid length */
115 total_length += lmd->md_iov.iov[i].iov_len;
118 lmd->md_length = total_length;
120 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
121 (umd->max_size < 0 ||
122 umd->max_size > total_length)) // illegal max_size
125 } else if ((umd->options & LNET_MD_KIOV) != 0) {
129 lmd->md_niov = niov = umd->length;
130 memcpy(lmd->md_iov.kiov, umd->start,
131 niov * sizeof (lmd->md_iov.kiov[0]));
133 for (i = 0; i < (int)niov; i++) {
134 /* We take the page pointer on trust */
135 if (lmd->md_iov.kiov[i].kiov_offset +
136 lmd->md_iov.kiov[i].kiov_len > CFS_PAGE_SIZE )
137 return -EINVAL; /* invalid length */
139 total_length += lmd->md_iov.kiov[i].kiov_len;
142 lmd->md_length = total_length;
144 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
145 (umd->max_size < 0 ||
146 umd->max_size > total_length)) // illegal max_size
149 } else { /* contiguous */
150 lmd->md_length = umd->length;
151 lmd->md_niov = niov = 1;
152 lmd->md_iov.iov[0].iov_base = umd->start;
153 lmd->md_iov.iov[0].iov_len = umd->length;
155 if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */
156 (umd->max_size < 0 ||
157 umd->max_size > (int)umd->length)) // illegal max_size
164 /* must be called with resource lock held */
166 lnet_md_link(lnet_libmd_t *md, lnet_handle_eq_t eq_handle)
168 struct lnet_res_container *container = &the_lnet.ln_md_container;
170 /* NB we are passed an allocated, but inactive md.
171 * if we return success, caller may lnet_md_unlink() it.
172 * otherwise caller may only lnet_md_free() it.
174 /* This implementation doesn't know how to create START events or
175 * disable END events. Best to LASSERT our caller is compliant so
176 * we find out quickly... */
177 /* TODO - reevaluate what should be here in light of
178 * the removal of the start and end events
179 * maybe there we shouldn't even allow LNET_EQ_NONE!)
180 * LASSERT (eq == NULL);
182 if (!LNetHandleIsInvalid(eq_handle)) {
183 md->md_eq = lnet_handle2eq(&eq_handle);
185 if (md->md_eq == NULL)
188 md->md_eq->eq_refcount++;
191 lnet_res_lh_initialize(container, &md->md_lh);
193 LASSERT(cfs_list_empty(&md->md_list));
194 cfs_list_add(&md->md_list, &container->rec_active);
199 /* must be called with LNET_LOCK held */
201 lnet_md_deconstruct(lnet_libmd_t *lmd, lnet_md_t *umd)
203 /* NB this doesn't copy out all the iov entries so when a
204 * discontiguous MD is copied out, the target gets to know the
205 * original iov pointer (in start) and the number of entries it had
208 umd->start = lmd->md_start;
209 umd->length = ((lmd->md_options & (LNET_MD_IOVEC | LNET_MD_KIOV)) == 0) ?
210 lmd->md_length : lmd->md_niov;
211 umd->threshold = lmd->md_threshold;
212 umd->max_size = lmd->md_max_size;
213 umd->options = lmd->md_options;
214 umd->user_ptr = lmd->md_user_ptr;
215 lnet_eq2handle(&umd->eq_handle, lmd->md_eq);
219 lnet_md_validate(lnet_md_t *umd)
221 if (umd->start == NULL && umd->length != 0) {
222 CERROR("MD start pointer can not be NULL with length %u\n",
227 if ((umd->options & (LNET_MD_KIOV | LNET_MD_IOVEC)) != 0 &&
228 umd->length > LNET_MAX_IOV) {
229 CERROR("Invalid option: too many fragments %u, %d max\n",
230 umd->length, LNET_MAX_IOV);
238 * Create a memory descriptor and attach it to a ME
240 * \param meh A handle for a ME to associate the new MD with.
241 * \param umd Provides initial values for the user-visible parts of a MD.
242 * Other than its use for initialization, there is no linkage between this
243 * structure and the MD maintained by the LNet.
244 * \param unlink A flag to indicate whether the MD is automatically unlinked
245 * when it becomes inactive, either because the operation threshold drops to
246 * zero or because the available memory becomes less than \a umd.max_size.
247 * (Note that the check for unlinking a MD only occurs after the completion
248 * of a successful operation on the MD.) The value LNET_UNLINK enables auto
249 * unlinking; the value LNET_RETAIN disables it.
250 * \param handle On successful returns, a handle to the newly created MD is
251 * saved here. This handle can be used later in LNetMDUnlink().
253 * \retval 0 On success.
254 * \retval -EINVAL If \a umd is not valid.
255 * \retval -ENOMEM If new MD cannot be allocated.
256 * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a
257 * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by
258 * calling LNetInvalidateHandle() on it.
259 * \retval -EBUSY If the ME pointed to by \a meh is already associated with
263 LNetMDAttach(lnet_handle_me_t meh, lnet_md_t umd,
264 lnet_unlink_t unlink, lnet_handle_md_t *handle)
266 CFS_LIST_HEAD (matches);
267 CFS_LIST_HEAD (drops);
272 LASSERT (the_lnet.ln_init);
273 LASSERT (the_lnet.ln_refcount > 0);
275 if (lnet_md_validate(&umd) != 0)
278 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) == 0) {
279 CERROR("Invalid option: no MD_OP set\n");
283 md = lnet_md_alloc(&umd);
287 rc = lnet_md_build(md, &umd, unlink);
293 me = lnet_handle2me(&meh);
296 else if (me->me_md != NULL)
299 rc = lnet_md_link(md, umd.eq_handle);
304 /* attach this MD to portal of ME and check if it matches any
305 * blocked msgs on this portal */
306 lnet_ptl_attach_md(me, md, &matches, &drops);
308 lnet_md2handle(handle, md);
312 lnet_drop_delayed_msg_list(&drops, "Bad match");
313 lnet_recv_delayed_msg_list(&matches);
318 lnet_md_free_locked(md);
325 * Create a "free floating" memory descriptor - a MD that is not associated
326 * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
328 * \param umd,unlink See the discussion for LNetMDAttach().
329 * \param handle On successful returns, a handle to the newly created MD is
330 * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(),
331 * and LNetGet() operations.
333 * \retval 0 On success.
334 * \retval -EINVAL If \a umd is not valid.
335 * \retval -ENOMEM If new MD cannot be allocated.
336 * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that
337 * it's OK to supply a NULL \a umd.eq_handle by calling
338 * LNetInvalidateHandle() on it.
341 LNetMDBind(lnet_md_t umd, lnet_unlink_t unlink, lnet_handle_md_t *handle)
346 LASSERT (the_lnet.ln_init);
347 LASSERT (the_lnet.ln_refcount > 0);
349 if (lnet_md_validate(&umd) != 0)
352 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) != 0) {
353 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
357 md = lnet_md_alloc(&umd);
361 rc = lnet_md_build(md, &umd, unlink);
367 rc = lnet_md_link(md, umd.eq_handle);
371 lnet_md2handle(handle, md);
377 lnet_md_free_locked(md);
384 * Unlink the memory descriptor from any ME it may be linked to and release
385 * the internal resources associated with it.
387 * This function does not free the memory region associated with the MD;
388 * i.e., the memory the user allocated for this MD. If the ME associated with
389 * this MD is not NULL and was created with auto unlink enabled, the ME is
390 * unlinked as well (see LNetMEAttach()).
392 * Explicitly unlinking a MD via this function call has the same behavior as
393 * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
394 * is generated in the latter case.
396 * An unlinked event can be reported in two ways:
397 * - If there's no pending operations on the MD, it's unlinked immediately
398 * and an LNET_EVENT_UNLINK event is logged before this function returns.
399 * - Otherwise, the MD is only marked for deletion when this function
400 * returns, and the unlinked event will be piggybacked on the event of
401 * the completion of the last operation by setting the unlinked field of
402 * the event. No dedicated LNET_EVENT_UNLINK event is generated.
404 * Note that in both cases the unlinked field of the event is always set; no
405 * more event will happen on the MD after such an event is logged.
407 * \param mdh A handle for the MD to be unlinked.
409 * \retval 0 On success.
410 * \retval -ENOENT If \a mdh does not point to a valid MD object.
413 LNetMDUnlink (lnet_handle_md_t mdh)
418 LASSERT (the_lnet.ln_init);
419 LASSERT (the_lnet.ln_refcount > 0);
423 md = lnet_handle2md(&mdh);
429 /* If the MD is busy, lnet_md_unlink just marks it for deletion, and
430 * when the NAL is done, the completion event flags that the MD was
431 * unlinked. Otherwise, we enqueue an event now... */
433 if (md->md_eq != NULL &&
434 md->md_refcount == 0) {
435 lnet_build_unlink_event(md, &ev);
436 lnet_eq_enqueue_event(md->md_eq, &ev);