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, 2017, 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 <lnet/lib-lnet.h>
41 /* must be called with lnet_res_lock held */
43 lnet_md_unlink(struct lnet_libmd *md)
45 if ((md->md_flags & LNET_MD_FLAG_ZOMBIE) == 0) {
46 /* first unlink attempt... */
47 struct lnet_me *me = md->md_me;
49 md->md_flags |= LNET_MD_FLAG_ZOMBIE;
51 /* Disassociate from ME (if any), and unlink it if it was created
54 /* detach MD from portal */
55 lnet_ptl_detach_md(me, md);
56 if (me->me_unlink == LNET_UNLINK)
60 /* ensure all future handle lookups fail */
61 lnet_res_lh_invalidate(&md->md_lh);
64 if (md->md_refcount != 0) {
65 CDEBUG(D_NET, "Queueing unlink of md %p\n", md);
69 CDEBUG(D_NET, "Unlinking md %p\n", md);
71 LASSERT(!list_empty(&md->md_list));
72 list_del_init(&md->md_list);
77 lnet_kvaddr_to_page(unsigned long vaddr)
79 if (is_vmalloc_addr((void *)vaddr))
80 return vmalloc_to_page((void *)vaddr);
84 #ifdef HAVE_KMAP_TO_PAGE
86 * This ifdef is added to handle the kernel versions
87 * which have kmap_to_page() function exported. If so,
88 * we should use it. Otherwise, remain with the legacy check.
90 return kmap_to_page((void *)vaddr);
93 if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) {
94 /* No highmem pages only used for bulk (kiov) I/O */
95 CERROR("find page for address in highmem\n");
98 return virt_to_page(vaddr);
99 #endif /* HAVE_KMAP_TO_PAGE */
102 return virt_to_page(vaddr);
103 #endif /* CONFIG_HIGHMEM */
105 EXPORT_SYMBOL(lnet_kvaddr_to_page);
108 lnet_cpt_of_md(struct lnet_libmd *md, unsigned int offset)
110 int cpt = CFS_CPT_ANY;
112 struct bio_vec *kiov;
115 * if the md_options has a bulk handle then we want to look at the
116 * bulk md because that's the data which we will be DMAing
118 if (md && (md->md_options & LNET_MD_BULK_HANDLE) != 0 &&
119 !LNetMDHandleIsInvalid(md->md_bulk_handle))
120 md = lnet_handle2md(&md->md_bulk_handle);
122 if (!md || md->md_niov == 0)
128 while (offset >= kiov->bv_len) {
129 offset -= kiov->bv_len;
133 CERROR("offset %d goes beyond kiov\n", offset);
138 cpt = cfs_cpt_of_node(lnet_cpt_table(),
139 page_to_nid(kiov->bv_page));
144 static struct lnet_libmd *
145 lnet_md_build(const struct lnet_md *umd, int unlink)
149 int total_length = 0;
150 struct lnet_libmd *lmd;
153 if (umd->options & LNET_MD_KIOV)
156 niov = DIV_ROUND_UP(offset_in_page(umd->start) + umd->length,
158 size = offsetof(struct lnet_libmd, md_kiov[niov]);
160 if (size <= LNET_SMALL_MD_SIZE) {
161 lmd = kmem_cache_zalloc(lnet_small_mds_cachep, GFP_NOFS);
164 "slab-alloced 'md' of size %u at %p.\n",
167 CDEBUG(D_MALLOC, "failed to allocate 'md' of size %u\n",
171 LIBCFS_ALLOC(lmd, size);
175 return ERR_PTR(-ENOMEM);
178 INIT_LIST_HEAD(&lmd->md_list);
180 lmd->md_start = umd->start;
182 lmd->md_max_size = umd->max_size;
183 lmd->md_options = umd->options;
184 lmd->md_user_ptr = umd->user_ptr;
185 lmd->md_handler = NULL;
186 lmd->md_threshold = umd->threshold;
187 lmd->md_refcount = 0;
188 lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
189 lmd->md_bulk_handle = umd->bulk_handle;
191 if (umd->options & LNET_MD_KIOV) {
192 memcpy(lmd->md_kiov, umd->start,
193 niov * sizeof(lmd->md_kiov[0]));
195 for (i = 0; i < (int)niov; i++) {
196 /* We take the page pointer on trust */
197 if (lmd->md_kiov[i].bv_offset +
198 lmd->md_kiov[i].bv_len > PAGE_SIZE) {
200 return ERR_PTR(-EINVAL); /* invalid length */
203 total_length += lmd->md_kiov[i].bv_len;
206 lmd->md_length = total_length;
208 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
209 (umd->max_size < 0 ||
210 umd->max_size > total_length)) { /* illegal max_size */
212 return ERR_PTR(-EINVAL);
214 } else { /* contiguous - split into pages */
215 void *pa = umd->start;
216 int len = umd->length;
218 lmd->md_length = len;
223 plen = min_t(int, len, PAGE_SIZE - offset_in_page(pa));
225 lmd->md_kiov[i].bv_page =
226 lnet_kvaddr_to_page((unsigned long) pa);
227 lmd->md_kiov[i].bv_offset = offset_in_page(pa);
228 lmd->md_kiov[i].bv_len = plen;
234 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
235 (umd->max_size < 0 ||
236 umd->max_size > (int)umd->length)) { /* illegal max_size */
238 return ERR_PTR(-EINVAL);
240 lmd->md_options |= LNET_MD_KIOV;
246 /* must be called with resource lock held */
248 lnet_md_link(struct lnet_libmd *md, lnet_handler_t handler, int cpt)
250 struct lnet_res_container *container = the_lnet.ln_md_containers[cpt];
252 /* NB we are passed an allocated, but inactive md.
253 * if we return success, caller may lnet_md_unlink() it.
254 * otherwise caller may only lnet_md_free() it.
256 /* This implementation doesn't know how to create START events or
257 * disable END events. Best to LASSERT our caller is compliant so
258 * we find out quickly... */
259 /* TODO - reevaluate what should be here in light of
260 * the removal of the start and end events
261 * maybe there we shouldn't even allow LNET_EQ_NONE!)
262 * LASSERT (handler != NULL);
264 md->md_handler = handler;
266 lnet_res_lh_initialize(container, &md->md_lh);
268 LASSERT(list_empty(&md->md_list));
269 list_add(&md->md_list, &container->rec_active);
274 void lnet_assert_handler_unused(lnet_handler_t handler)
276 struct lnet_res_container *container;
281 cfs_percpt_for_each(container, cpt, the_lnet.ln_md_containers) {
282 struct lnet_libmd *md;
285 list_for_each_entry(md, &container->rec_active, md_list)
286 LASSERT(md->md_handler != handler);
287 lnet_res_unlock(cpt);
290 EXPORT_SYMBOL(lnet_assert_handler_unused);
292 /* must be called with lnet_res_lock held */
294 lnet_md_deconstruct(struct lnet_libmd *lmd, struct lnet_event *ev)
296 ev->md_start = lmd->md_start;
297 ev->md_options = lmd->md_options;
298 ev->md_user_ptr = lmd->md_user_ptr;
302 lnet_md_validate(const struct lnet_md *umd)
304 if (umd->start == NULL && umd->length != 0) {
305 CERROR("MD start pointer can not be NULL with length %u\n",
310 if ((umd->options & LNET_MD_KIOV) &&
311 umd->length > LNET_MAX_IOV) {
312 CERROR("Invalid option: too many fragments %u, %d max\n",
313 umd->length, LNET_MAX_IOV);
321 * Create a memory descriptor and attach it to a ME
323 * \param me An ME to associate the new MD with.
324 * \param umd Provides initial values for the user-visible parts of a MD.
325 * Other than its use for initialization, there is no linkage between this
326 * structure and the MD maintained by the LNet.
327 * \param unlink A flag to indicate whether the MD is automatically unlinked
328 * when it becomes inactive, either because the operation threshold drops to
329 * zero or because the available memory becomes less than \a umd.max_size.
330 * (Note that the check for unlinking a MD only occurs after the completion
331 * of a successful operation on the MD.) The value LNET_UNLINK enables auto
332 * unlinking; the value LNET_RETAIN disables it.
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().
336 * \retval 0 On success.
337 * \retval -EINVAL If \a umd is not valid.
338 * \retval -ENOMEM If new MD cannot be allocated.
339 * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a
340 * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by
341 * calling LNetInvalidateHandle() on it.
342 * \retval -EBUSY If the ME pointed to by \a meh is already associated with
346 LNetMDAttach(struct lnet_me *me, const struct lnet_md *umd,
347 enum lnet_unlink unlink, struct lnet_handle_md *handle)
351 struct lnet_libmd *md;
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: no MD_OP set\n");
365 md = lnet_md_build(umd, unlink);
376 rc = lnet_md_link(md, umd->handler, cpt);
381 /* attach this MD to portal of ME and check if it matches any
382 * blocked msgs on this portal */
383 lnet_ptl_attach_md(me, md, &matches, &drops);
385 lnet_md2handle(handle, md);
387 lnet_res_unlock(cpt);
389 lnet_drop_delayed_msg_list(&drops, "Bad match");
390 lnet_recv_delayed_msg_list(&matches);
395 lnet_res_unlock(cpt);
399 EXPORT_SYMBOL(LNetMDAttach);
402 * Create a "free floating" memory descriptor - a MD that is not associated
403 * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
405 * \param umd,unlink See the discussion for LNetMDAttach().
406 * \param handle On successful returns, a handle to the newly created MD is
407 * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(),
408 * and LNetGet() operations.
410 * \retval 0 On success.
411 * \retval -EINVAL If \a umd is not valid.
412 * \retval -ENOMEM If new MD cannot be allocated.
413 * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that
414 * it's OK to supply a NULL \a umd.eq_handle by calling
415 * LNetInvalidateHandle() on it.
418 LNetMDBind(const struct lnet_md *umd, enum lnet_unlink unlink,
419 struct lnet_handle_md *handle)
421 struct lnet_libmd *md;
425 LASSERT(the_lnet.ln_refcount > 0);
427 if (lnet_md_validate(umd) != 0)
430 if ((umd->options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) != 0) {
431 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
435 md = lnet_md_build(umd, unlink);
439 if (md->md_length > LNET_MTU) {
440 CERROR("Invalid length: too big transfer size %u, %d max\n",
441 md->md_length, LNET_MTU);
446 cpt = lnet_res_lock_current();
448 rc = lnet_md_link(md, umd->handler, cpt);
452 lnet_md2handle(handle, md);
454 lnet_res_unlock(cpt);
458 lnet_res_unlock(cpt);
464 EXPORT_SYMBOL(LNetMDBind);
467 * Unlink the memory descriptor from any ME it may be linked to and release
468 * the internal resources associated with it. As a result, active messages
469 * associated with the MD may get aborted.
471 * This function does not free the memory region associated with the MD;
472 * i.e., the memory the user allocated for this MD. If the ME associated with
473 * this MD is not NULL and was created with auto unlink enabled, the ME is
474 * unlinked as well (see LNetMEAttach()).
476 * Explicitly unlinking a MD via this function call has the same behavior as
477 * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
478 * is generated in the latter case.
480 * An unlinked event can be reported in two ways:
481 * - If there's no pending operations on the MD, it's unlinked immediately
482 * and an LNET_EVENT_UNLINK event is logged before this function returns.
483 * - Otherwise, the MD is only marked for deletion when this function
484 * returns, and the unlinked event will be piggybacked on the event of
485 * the completion of the last operation by setting the unlinked field of
486 * the event. No dedicated LNET_EVENT_UNLINK event is generated.
488 * Note that in both cases the unlinked field of the event is always set; no
489 * more event will happen on the MD after such an event is logged.
491 * \param mdh A handle for the MD to be unlinked.
493 * \retval 0 On success.
494 * \retval -ENOENT If \a mdh does not point to a valid MD object.
497 LNetMDUnlink(struct lnet_handle_md mdh)
499 struct lnet_event ev;
500 struct lnet_libmd *md;
503 LASSERT(the_lnet.ln_refcount > 0);
505 cpt = lnet_cpt_of_cookie(mdh.cookie);
508 md = lnet_handle2md(&mdh);
510 lnet_res_unlock(cpt);
514 md->md_flags |= LNET_MD_FLAG_ABORTED;
515 /* If the MD is busy, lnet_md_unlink just marks it for deletion, and
516 * when the LND is done, the completion event flags that the MD was
517 * unlinked. Otherwise, we enqueue an event now... */
518 if (md->md_handler && md->md_refcount == 0) {
519 lnet_build_unlink_event(md, &ev);
523 if (md->md_rspt_ptr != NULL)
524 lnet_detach_rsp_tracker(md, cpt);
528 lnet_res_unlock(cpt);
531 EXPORT_SYMBOL(LNetMDUnlink);