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, 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 "../../include/linux/lnet/lib-lnet.h"
41 /* must be called with lnet_res_lock held */
43 lnet_md_unlink(lnet_libmd_t *md)
45 if (!(md->md_flags & LNET_MD_FLAG_ZOMBIE)) {
46 /* first unlink attempt... */
47 lnet_me_t *me = md->md_me;
49 md->md_flags |= LNET_MD_FLAG_ZOMBIE;
52 * Disassociate from ME (if any),
53 * and unlink it if it was created
57 /* detach MD from portal */
58 lnet_ptl_detach_md(me, md);
59 if (me->me_unlink == LNET_UNLINK)
63 /* ensure all future handle lookups fail */
64 lnet_res_lh_invalidate(&md->md_lh);
67 if (md->md_refcount) {
68 CDEBUG(D_NET, "Queueing unlink of md %p\n", md);
72 CDEBUG(D_NET, "Unlinking md %p\n", md);
75 int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
77 LASSERT(*md->md_eq->eq_refs[cpt] > 0);
78 (*md->md_eq->eq_refs[cpt])--;
81 LASSERT(!list_empty(&md->md_list));
82 list_del_init(&md->md_list);
87 lnet_md_build(lnet_libmd_t *lmd, lnet_md_t *umd, int unlink)
94 lmd->md_start = umd->start;
96 lmd->md_max_size = umd->max_size;
97 lmd->md_options = umd->options;
98 lmd->md_user_ptr = umd->user_ptr;
100 lmd->md_threshold = umd->threshold;
101 lmd->md_refcount = 0;
102 lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
104 if (umd->options & LNET_MD_IOVEC) {
105 if (umd->options & LNET_MD_KIOV) /* Can't specify both */
109 lmd->md_niov = umd->length;
110 memcpy(lmd->md_iov.iov, umd->start,
111 niov * sizeof(lmd->md_iov.iov[0]));
113 for (i = 0; i < (int)niov; i++) {
114 /* We take the base address on trust */
116 if (lmd->md_iov.iov[i].iov_len <= 0)
119 total_length += lmd->md_iov.iov[i].iov_len;
122 lmd->md_length = total_length;
124 if ((umd->options & LNET_MD_MAX_SIZE) && /* use max size */
125 (umd->max_size < 0 ||
126 umd->max_size > total_length)) /* illegal max_size */
129 } else if (umd->options & LNET_MD_KIOV) {
131 lmd->md_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].bv_offset +
138 lmd->md_iov.kiov[i].bv_len > PAGE_SIZE)
139 return -EINVAL; /* invalid length */
141 total_length += lmd->md_iov.kiov[i].bv_len;
144 lmd->md_length = total_length;
146 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
147 (umd->max_size < 0 ||
148 umd->max_size > total_length)) /* illegal max_size */
150 } else { /* contiguous */
151 lmd->md_length = umd->length;
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) && /* 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];
173 * NB we are passed an allocated, but inactive md.
174 * if we return success, caller may lnet_md_unlink() it.
175 * 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...
183 * TODO - reevaluate what should be here in light of
184 * the removal of the start and end events
185 * maybe there we shouldn't even allow LNET_EQ_NONE!)
188 if (!LNetHandleIsInvalid(eq_handle)) {
189 md->md_eq = lnet_handle2eq(&eq_handle);
194 (*md->md_eq->eq_refs[cpt])++;
197 lnet_res_lh_initialize(container, &md->md_lh);
199 LASSERT(list_empty(&md->md_list));
200 list_add(&md->md_list, &container->rec_active);
205 /* must be called with lnet_res_lock held */
207 lnet_md_deconstruct(lnet_libmd_t *lmd, lnet_md_t *umd)
209 /* NB this doesn't copy out all the iov entries so when a
210 * discontiguous MD is copied out, the target gets to know the
211 * original iov pointer (in start) and the number of entries it had
214 umd->start = lmd->md_start;
215 umd->length = !(lmd->md_options &
216 (LNET_MD_IOVEC | LNET_MD_KIOV)) ?
217 lmd->md_length : lmd->md_niov;
218 umd->threshold = lmd->md_threshold;
219 umd->max_size = lmd->md_max_size;
220 umd->options = lmd->md_options;
221 umd->user_ptr = lmd->md_user_ptr;
222 lnet_eq2handle(&umd->eq_handle, lmd->md_eq);
226 lnet_md_validate(lnet_md_t *umd)
228 if (!umd->start && umd->length) {
229 CERROR("MD start pointer can not be NULL with length %u\n",
234 if ((umd->options & (LNET_MD_KIOV | LNET_MD_IOVEC)) &&
235 umd->length > LNET_MAX_IOV) {
236 CERROR("Invalid option: too many fragments %u, %d max\n",
237 umd->length, LNET_MAX_IOV);
245 * Create a memory descriptor and attach it to a ME
247 * \param meh A handle for a ME to associate the new MD with.
248 * \param umd Provides initial values for the user-visible parts of a MD.
249 * Other than its use for initialization, there is no linkage between this
250 * structure and the MD maintained by the LNet.
251 * \param unlink A flag to indicate whether the MD is automatically unlinked
252 * when it becomes inactive, either because the operation threshold drops to
253 * zero or because the available memory becomes less than \a umd.max_size.
254 * (Note that the check for unlinking a MD only occurs after the completion
255 * of a successful operation on the MD.) The value LNET_UNLINK enables auto
256 * unlinking; the value LNET_RETAIN disables it.
257 * \param handle On successful returns, a handle to the newly created MD is
258 * saved here. This handle can be used later in LNetMDUnlink().
260 * \retval 0 On success.
261 * \retval -EINVAL If \a umd is not valid.
262 * \retval -ENOMEM If new MD cannot be allocated.
263 * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a
264 * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by
265 * calling LNetInvalidateHandle() on it.
266 * \retval -EBUSY If the ME pointed to by \a meh is already associated with
270 LNetMDAttach(lnet_handle_me_t meh, lnet_md_t umd,
271 lnet_unlink_t unlink, lnet_handle_md_t *handle)
276 struct lnet_libmd *md;
280 LASSERT(the_lnet.ln_refcount > 0);
282 if (lnet_md_validate(&umd))
285 if (!(umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT))) {
286 CERROR("Invalid option: no MD_OP set\n");
290 md = lnet_md_alloc(&umd);
294 rc = lnet_md_build(md, &umd, unlink);
298 cpt = lnet_cpt_of_cookie(meh.cookie);
302 me = lnet_handle2me(&meh);
308 rc = lnet_md_link(md, umd.eq_handle, cpt);
314 * attach this MD to portal of ME and check if it matches any
315 * blocked msgs on this portal
317 lnet_ptl_attach_md(me, md, &matches, &drops);
319 lnet_md2handle(handle, md);
321 lnet_res_unlock(cpt);
323 lnet_drop_delayed_msg_list(&drops, "Bad match");
324 lnet_recv_delayed_msg_list(&matches);
329 lnet_res_unlock(cpt);
334 EXPORT_SYMBOL(LNetMDAttach);
337 * Create a "free floating" memory descriptor - a MD that is not associated
338 * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
340 * \param umd,unlink See the discussion for LNetMDAttach().
341 * \param handle On successful returns, a handle to the newly created MD is
342 * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(),
343 * and LNetGet() operations.
345 * \retval 0 On success.
346 * \retval -EINVAL If \a umd is not valid.
347 * \retval -ENOMEM If new MD cannot be allocated.
348 * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that
349 * it's OK to supply a NULL \a umd.eq_handle by calling
350 * LNetInvalidateHandle() on it.
353 LNetMDBind(lnet_md_t umd, lnet_unlink_t unlink, lnet_handle_md_t *handle)
359 LASSERT(the_lnet.ln_refcount > 0);
361 if (lnet_md_validate(&umd))
364 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT))) {
365 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
369 md = lnet_md_alloc(&umd);
373 rc = lnet_md_build(md, &umd, unlink);
377 cpt = lnet_res_lock_current();
379 rc = lnet_md_link(md, umd.eq_handle, cpt);
383 lnet_md2handle(handle, md);
385 lnet_res_unlock(cpt);
389 lnet_res_unlock(cpt);
395 EXPORT_SYMBOL(LNetMDBind);
398 * Unlink the memory descriptor from any ME it may be linked to and release
399 * the internal resources associated with it. As a result, active messages
400 * associated with the MD may get aborted.
402 * This function does not free the memory region associated with the MD;
403 * i.e., the memory the user allocated for this MD. If the ME associated with
404 * this MD is not NULL and was created with auto unlink enabled, the ME is
405 * unlinked as well (see LNetMEAttach()).
407 * Explicitly unlinking a MD via this function call has the same behavior as
408 * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
409 * is generated in the latter case.
411 * An unlinked event can be reported in two ways:
412 * - If there's no pending operations on the MD, it's unlinked immediately
413 * and an LNET_EVENT_UNLINK event is logged before this function returns.
414 * - Otherwise, the MD is only marked for deletion when this function
415 * returns, and the unlinked event will be piggybacked on the event of
416 * the completion of the last operation by setting the unlinked field of
417 * the event. No dedicated LNET_EVENT_UNLINK event is generated.
419 * Note that in both cases the unlinked field of the event is always set; no
420 * more event will happen on the MD after such an event is logged.
422 * \param mdh A handle for the MD to be unlinked.
424 * \retval 0 On success.
425 * \retval -ENOENT If \a mdh does not point to a valid MD object.
428 LNetMDUnlink(lnet_handle_md_t mdh)
434 LASSERT(the_lnet.ln_refcount > 0);
436 cpt = lnet_cpt_of_cookie(mdh.cookie);
439 md = lnet_handle2md(&mdh);
441 lnet_res_unlock(cpt);
445 md->md_flags |= LNET_MD_FLAG_ABORTED;
447 * If the MD is busy, lnet_md_unlink just marks it for deletion, and
448 * when the LND is done, the completion event flags that the MD was
449 * unlinked. Otherwise, we enqueue an event now...
451 if (md->md_eq && !md->md_refcount) {
452 lnet_build_unlink_event(md, &ev);
453 lnet_eq_enqueue_event(md->md_eq, &ev);
458 lnet_res_unlock(cpt);
461 EXPORT_SYMBOL(LNetMDUnlink);