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, 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 "../../include/linux/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)) {
50 /* first unlink attempt... */
51 lnet_me_t *me = md->md_me;
53 md->md_flags |= LNET_MD_FLAG_ZOMBIE;
56 * Disassociate from ME (if any),
57 * and unlink it if it was created
61 /* detach MD from portal */
62 lnet_ptl_detach_md(me, md);
63 if (me->me_unlink == LNET_UNLINK)
67 /* ensure all future handle lookups fail */
68 lnet_res_lh_invalidate(&md->md_lh);
71 if (md->md_refcount) {
72 CDEBUG(D_NET, "Queueing unlink of md %p\n", md);
76 CDEBUG(D_NET, "Unlinking md %p\n", md);
79 int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
81 LASSERT(*md->md_eq->eq_refs[cpt] > 0);
82 (*md->md_eq->eq_refs[cpt])--;
85 LASSERT(!list_empty(&md->md_list));
86 list_del_init(&md->md_list);
91 lnet_md_build(lnet_libmd_t *lmd, lnet_md_t *umd, int unlink)
98 lmd->md_start = umd->start;
100 lmd->md_max_size = umd->max_size;
101 lmd->md_options = umd->options;
102 lmd->md_user_ptr = umd->user_ptr;
104 lmd->md_threshold = umd->threshold;
105 lmd->md_refcount = 0;
106 lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
108 if (umd->options & LNET_MD_IOVEC) {
109 if (umd->options & LNET_MD_KIOV) /* Can't specify both */
113 lmd->md_niov = umd->length;
114 memcpy(lmd->md_iov.iov, umd->start,
115 niov * sizeof(lmd->md_iov.iov[0]));
117 for (i = 0; i < (int)niov; i++) {
118 /* We take the base address on trust */
120 if (lmd->md_iov.iov[i].iov_len <= 0)
123 total_length += lmd->md_iov.iov[i].iov_len;
126 lmd->md_length = total_length;
128 if ((umd->options & LNET_MD_MAX_SIZE) && /* use max size */
129 (umd->max_size < 0 ||
130 umd->max_size > total_length)) /* illegal max_size */
133 } else if (umd->options & LNET_MD_KIOV) {
135 lmd->md_niov = umd->length;
136 memcpy(lmd->md_iov.kiov, umd->start,
137 niov * sizeof(lmd->md_iov.kiov[0]));
139 for (i = 0; i < (int)niov; i++) {
140 /* We take the page pointer on trust */
141 if (lmd->md_iov.kiov[i].kiov_offset +
142 lmd->md_iov.kiov[i].kiov_len > PAGE_SIZE)
143 return -EINVAL; /* invalid length */
145 total_length += lmd->md_iov.kiov[i].kiov_len;
148 lmd->md_length = total_length;
150 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
151 (umd->max_size < 0 ||
152 umd->max_size > total_length)) /* illegal max_size */
154 } else { /* contiguous */
155 lmd->md_length = umd->length;
158 lmd->md_iov.iov[0].iov_base = umd->start;
159 lmd->md_iov.iov[0].iov_len = umd->length;
161 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
162 (umd->max_size < 0 ||
163 umd->max_size > (int)umd->length)) /* illegal max_size */
170 /* must be called with resource lock held */
172 lnet_md_link(lnet_libmd_t *md, lnet_handle_eq_t eq_handle, int cpt)
174 struct lnet_res_container *container = the_lnet.ln_md_containers[cpt];
177 * NB we are passed an allocated, but inactive md.
178 * if we return success, caller may lnet_md_unlink() it.
179 * otherwise caller may only lnet_md_free() it.
182 * This implementation doesn't know how to create START events or
183 * disable END events. Best to LASSERT our caller is compliant so
184 * we find out quickly...
187 * TODO - reevaluate what should be here in light of
188 * the removal of the start and end events
189 * maybe there we shouldn't even allow LNET_EQ_NONE!)
192 if (!LNetHandleIsInvalid(eq_handle)) {
193 md->md_eq = lnet_handle2eq(&eq_handle);
198 (*md->md_eq->eq_refs[cpt])++;
201 lnet_res_lh_initialize(container, &md->md_lh);
203 LASSERT(list_empty(&md->md_list));
204 list_add(&md->md_list, &container->rec_active);
209 /* must be called with lnet_res_lock held */
211 lnet_md_deconstruct(lnet_libmd_t *lmd, lnet_md_t *umd)
213 /* NB this doesn't copy out all the iov entries so when a
214 * discontiguous MD is copied out, the target gets to know the
215 * original iov pointer (in start) and the number of entries it had
218 umd->start = lmd->md_start;
219 umd->length = !(lmd->md_options &
220 (LNET_MD_IOVEC | LNET_MD_KIOV)) ?
221 lmd->md_length : lmd->md_niov;
222 umd->threshold = lmd->md_threshold;
223 umd->max_size = lmd->md_max_size;
224 umd->options = lmd->md_options;
225 umd->user_ptr = lmd->md_user_ptr;
226 lnet_eq2handle(&umd->eq_handle, lmd->md_eq);
230 lnet_md_validate(lnet_md_t *umd)
232 if (!umd->start && umd->length) {
233 CERROR("MD start pointer can not be NULL with length %u\n",
238 if ((umd->options & (LNET_MD_KIOV | LNET_MD_IOVEC)) &&
239 umd->length > LNET_MAX_IOV) {
240 CERROR("Invalid option: too many fragments %u, %d max\n",
241 umd->length, LNET_MAX_IOV);
249 * Create a memory descriptor and attach it to a ME
251 * \param meh A handle for a ME to associate the new MD with.
252 * \param umd Provides initial values for the user-visible parts of a MD.
253 * Other than its use for initialization, there is no linkage between this
254 * structure and the MD maintained by the LNet.
255 * \param unlink A flag to indicate whether the MD is automatically unlinked
256 * when it becomes inactive, either because the operation threshold drops to
257 * zero or because the available memory becomes less than \a umd.max_size.
258 * (Note that the check for unlinking a MD only occurs after the completion
259 * of a successful operation on the MD.) The value LNET_UNLINK enables auto
260 * unlinking; the value LNET_RETAIN disables it.
261 * \param handle On successful returns, a handle to the newly created MD is
262 * saved here. This handle can be used later in LNetMDUnlink().
264 * \retval 0 On success.
265 * \retval -EINVAL If \a umd is not valid.
266 * \retval -ENOMEM If new MD cannot be allocated.
267 * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a
268 * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by
269 * calling LNetInvalidateHandle() on it.
270 * \retval -EBUSY If the ME pointed to by \a meh is already associated with
274 LNetMDAttach(lnet_handle_me_t meh, lnet_md_t umd,
275 lnet_unlink_t unlink, lnet_handle_md_t *handle)
280 struct lnet_libmd *md;
284 LASSERT(the_lnet.ln_refcount > 0);
286 if (lnet_md_validate(&umd))
289 if (!(umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT))) {
290 CERROR("Invalid option: no MD_OP set\n");
294 md = lnet_md_alloc(&umd);
298 rc = lnet_md_build(md, &umd, unlink);
299 cpt = lnet_cpt_of_cookie(meh.cookie);
305 me = lnet_handle2me(&meh);
311 rc = lnet_md_link(md, umd.eq_handle, cpt);
317 * attach this MD to portal of ME and check if it matches any
318 * blocked msgs on this portal
320 lnet_ptl_attach_md(me, md, &matches, &drops);
322 lnet_md2handle(handle, md);
324 lnet_res_unlock(cpt);
326 lnet_drop_delayed_msg_list(&drops, "Bad match");
327 lnet_recv_delayed_msg_list(&matches);
334 lnet_res_unlock(cpt);
337 EXPORT_SYMBOL(LNetMDAttach);
340 * Create a "free floating" memory descriptor - a MD that is not associated
341 * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
343 * \param umd,unlink See the discussion for LNetMDAttach().
344 * \param handle On successful returns, a handle to the newly created MD is
345 * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(),
346 * and LNetGet() operations.
348 * \retval 0 On success.
349 * \retval -EINVAL If \a umd is not valid.
350 * \retval -ENOMEM If new MD cannot be allocated.
351 * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that
352 * it's OK to supply a NULL \a umd.eq_handle by calling
353 * LNetInvalidateHandle() on it.
356 LNetMDBind(lnet_md_t umd, lnet_unlink_t unlink, lnet_handle_md_t *handle)
362 LASSERT(the_lnet.ln_refcount > 0);
364 if (lnet_md_validate(&umd))
367 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT))) {
368 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
372 md = lnet_md_alloc(&umd);
376 rc = lnet_md_build(md, &umd, unlink);
378 cpt = lnet_res_lock_current();
382 rc = lnet_md_link(md, umd.eq_handle, cpt);
386 lnet_md2handle(handle, md);
388 lnet_res_unlock(cpt);
394 lnet_res_unlock(cpt);
397 EXPORT_SYMBOL(LNetMDBind);
400 * Unlink the memory descriptor from any ME it may be linked to and release
401 * the internal resources associated with it. As a result, active messages
402 * associated with the MD may get aborted.
404 * This function does not free the memory region associated with the MD;
405 * i.e., the memory the user allocated for this MD. If the ME associated with
406 * this MD is not NULL and was created with auto unlink enabled, the ME is
407 * unlinked as well (see LNetMEAttach()).
409 * Explicitly unlinking a MD via this function call has the same behavior as
410 * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
411 * is generated in the latter case.
413 * An unlinked event can be reported in two ways:
414 * - If there's no pending operations on the MD, it's unlinked immediately
415 * and an LNET_EVENT_UNLINK event is logged before this function returns.
416 * - Otherwise, the MD is only marked for deletion when this function
417 * returns, and the unlinked event will be piggybacked on the event of
418 * the completion of the last operation by setting the unlinked field of
419 * the event. No dedicated LNET_EVENT_UNLINK event is generated.
421 * Note that in both cases the unlinked field of the event is always set; no
422 * more event will happen on the MD after such an event is logged.
424 * \param mdh A handle for the MD to be unlinked.
426 * \retval 0 On success.
427 * \retval -ENOENT If \a mdh does not point to a valid MD object.
430 LNetMDUnlink(lnet_handle_md_t mdh)
436 LASSERT(the_lnet.ln_refcount > 0);
438 cpt = lnet_cpt_of_cookie(mdh.cookie);
441 md = lnet_handle2md(&mdh);
443 lnet_res_unlock(cpt);
447 md->md_flags |= LNET_MD_FLAG_ABORTED;
449 * If the MD is busy, lnet_md_unlink just marks it for deletion, and
450 * when the LND is done, the completion event flags that the MD was
451 * unlinked. Otherwise, we enqueue an event now...
453 if (md->md_eq && !md->md_refcount) {
454 lnet_build_unlink_event(md, &ev);
455 lnet_eq_enqueue_event(md->md_eq, &ev);
460 lnet_res_unlock(cpt);
463 EXPORT_SYMBOL(LNetMDUnlink);