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;
114 * if the md_options has a bulk handle then we want to look at the
115 * bulk md because that's the data which we will be DMAing
117 if (md && (md->md_options & LNET_MD_BULK_HANDLE) != 0 &&
118 !LNetMDHandleIsInvalid(md->md_bulk_handle))
119 md = lnet_handle2md(&md->md_bulk_handle);
121 if (!md || md->md_niov == 0)
127 * There are three cases to handle:
128 * 1. The MD is using struct bio_vec
129 * 2. The MD is using struct kvec
130 * 3. Contiguous buffer allocated via vmalloc
132 * in case 2 we can use virt_to_page() macro to get the page
133 * address of the memory kvec describes.
135 * in case 3 use is_vmalloc_addr() and vmalloc_to_page()
137 * The offset provided can be within the first iov/kiov entry or
138 * it could go beyond it. In that case we need to make sure to
139 * look at the page which actually contains the data that will be
142 if ((md->md_options & LNET_MD_KIOV) != 0) {
143 struct bio_vec *kiov = md->md_iov.kiov;
145 while (offset >= kiov->bv_len) {
146 offset -= kiov->bv_len;
150 CERROR("offset %d goes beyond kiov\n", offset);
155 cpt = cfs_cpt_of_node(lnet_cpt_table(),
156 page_to_nid(kiov->bv_page));
158 struct kvec *iov = md->md_iov.iov;
162 while (offset >= iov->iov_len) {
163 offset -= iov->iov_len;
167 CERROR("offset %d goes beyond iov\n", offset);
172 vaddr = ((unsigned long)iov->iov_base) + offset;
173 page = lnet_kvaddr_to_page(vaddr);
175 CERROR("Couldn't resolve vaddr 0x%lx to page\n", vaddr);
178 cpt = cfs_cpt_of_node(lnet_cpt_table(), page_to_nid(page));
185 static struct lnet_libmd *
186 lnet_md_build(struct lnet_md *umd, int unlink)
190 int total_length = 0;
191 struct lnet_libmd *lmd;
194 if ((umd->options & LNET_MD_KIOV) != 0) {
196 size = offsetof(struct lnet_libmd, md_iov.kiov[niov]);
199 size = offsetof(struct lnet_libmd, md_iov.iov[niov]);
202 if (size <= LNET_SMALL_MD_SIZE) {
203 lmd = kmem_cache_zalloc(lnet_small_mds_cachep, GFP_NOFS);
206 "slab-alloced 'md' of size %u at %p.\n",
209 CDEBUG(D_MALLOC, "failed to allocate 'md' of size %u\n",
213 LIBCFS_ALLOC(lmd, size);
217 return ERR_PTR(-ENOMEM);
220 INIT_LIST_HEAD(&lmd->md_list);
223 lmd->md_start = umd->start;
225 lmd->md_max_size = umd->max_size;
226 lmd->md_options = umd->options;
227 lmd->md_user_ptr = umd->user_ptr;
228 lmd->md_handler = NULL;
229 lmd->md_threshold = umd->threshold;
230 lmd->md_refcount = 0;
231 lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
232 lmd->md_bulk_handle = umd->bulk_handle;
234 if ((umd->options & LNET_MD_KIOV) != 0) {
235 lmd->md_niov = niov = umd->length;
236 memcpy(lmd->md_iov.kiov, umd->start,
237 niov * sizeof(lmd->md_iov.kiov[0]));
239 for (i = 0; i < (int)niov; i++) {
240 /* We take the page pointer on trust */
241 if (lmd->md_iov.kiov[i].bv_offset +
242 lmd->md_iov.kiov[i].bv_len > PAGE_SIZE) {
244 return ERR_PTR(-EINVAL); /* invalid length */
247 total_length += lmd->md_iov.kiov[i].bv_len;
250 lmd->md_length = total_length;
252 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
253 (umd->max_size < 0 ||
254 umd->max_size > total_length)) { /* illegal max_size */
256 return ERR_PTR(-EINVAL);
258 } else { /* contiguous */
259 lmd->md_length = umd->length;
260 lmd->md_niov = niov = 1;
261 lmd->md_iov.iov[0].iov_base = umd->start;
262 lmd->md_iov.iov[0].iov_len = umd->length;
264 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
265 (umd->max_size < 0 ||
266 umd->max_size > (int)umd->length)) { /* illegal max_size */
268 return ERR_PTR(-EINVAL);
275 /* must be called with resource lock held */
277 lnet_md_link(struct lnet_libmd *md, lnet_handler_t handler, int cpt)
279 struct lnet_res_container *container = the_lnet.ln_md_containers[cpt];
281 /* NB we are passed an allocated, but inactive md.
282 * if we return success, caller may lnet_md_unlink() it.
283 * otherwise caller may only lnet_md_free() it.
285 /* This implementation doesn't know how to create START events or
286 * disable END events. Best to LASSERT our caller is compliant so
287 * we find out quickly... */
288 /* TODO - reevaluate what should be here in light of
289 * the removal of the start and end events
290 * maybe there we shouldn't even allow LNET_EQ_NONE!)
291 * LASSERT (handler != NULL);
293 md->md_handler = handler;
295 lnet_res_lh_initialize(container, &md->md_lh);
297 LASSERT(list_empty(&md->md_list));
298 list_add(&md->md_list, &container->rec_active);
303 void lnet_assert_handler_unused(lnet_handler_t handler)
305 struct lnet_res_container *container;
310 cfs_percpt_for_each(container, cpt, the_lnet.ln_md_containers) {
311 struct lnet_libmd *md;
314 list_for_each_entry(md, &container->rec_active, md_list)
315 LASSERT(md->md_handler != handler);
316 lnet_res_unlock(cpt);
319 EXPORT_SYMBOL(lnet_assert_handler_unused);
321 /* must be called with lnet_res_lock held */
323 lnet_md_deconstruct(struct lnet_libmd *lmd, struct lnet_event *ev)
325 ev->md_start = lmd->md_start;
326 ev->md_options = lmd->md_options;
327 ev->md_user_ptr = lmd->md_user_ptr;
331 lnet_md_validate(struct lnet_md *umd)
333 if (umd->start == NULL && umd->length != 0) {
334 CERROR("MD start pointer can not be NULL with length %u\n",
339 if ((umd->options & LNET_MD_KIOV) &&
340 umd->length > LNET_MAX_IOV) {
341 CERROR("Invalid option: too many fragments %u, %d max\n",
342 umd->length, LNET_MAX_IOV);
350 * Create a memory descriptor and attach it to a ME
352 * \param me An ME to associate the new MD with.
353 * \param umd Provides initial values for the user-visible parts of a MD.
354 * Other than its use for initialization, there is no linkage between this
355 * structure and the MD maintained by the LNet.
356 * \param unlink A flag to indicate whether the MD is automatically unlinked
357 * when it becomes inactive, either because the operation threshold drops to
358 * zero or because the available memory becomes less than \a umd.max_size.
359 * (Note that the check for unlinking a MD only occurs after the completion
360 * of a successful operation on the MD.) The value LNET_UNLINK enables auto
361 * unlinking; the value LNET_RETAIN disables it.
362 * \param handle On successful returns, a handle to the newly created MD is
363 * saved here. This handle can be used later in LNetMDUnlink().
365 * \retval 0 On success.
366 * \retval -EINVAL If \a umd is not valid.
367 * \retval -ENOMEM If new MD cannot be allocated.
368 * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a
369 * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by
370 * calling LNetInvalidateHandle() on it.
371 * \retval -EBUSY If the ME pointed to by \a meh is already associated with
375 LNetMDAttach(struct lnet_me *me, struct lnet_md umd,
376 enum lnet_unlink unlink, struct lnet_handle_md *handle)
380 struct lnet_libmd *md;
384 LASSERT(the_lnet.ln_refcount > 0);
386 if (lnet_md_validate(&umd) != 0)
389 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) == 0) {
390 CERROR("Invalid option: no MD_OP set\n");
394 md = lnet_md_build(&umd, unlink);
405 rc = lnet_md_link(md, umd.handler, cpt);
410 /* attach this MD to portal of ME and check if it matches any
411 * blocked msgs on this portal */
412 lnet_ptl_attach_md(me, md, &matches, &drops);
414 lnet_md2handle(handle, md);
416 lnet_res_unlock(cpt);
418 lnet_drop_delayed_msg_list(&drops, "Bad match");
419 lnet_recv_delayed_msg_list(&matches);
424 lnet_res_unlock(cpt);
428 EXPORT_SYMBOL(LNetMDAttach);
431 * Create a "free floating" memory descriptor - a MD that is not associated
432 * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
434 * \param umd,unlink See the discussion for LNetMDAttach().
435 * \param handle On successful returns, a handle to the newly created MD is
436 * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(),
437 * and LNetGet() operations.
439 * \retval 0 On success.
440 * \retval -EINVAL If \a umd is not valid.
441 * \retval -ENOMEM If new MD cannot be allocated.
442 * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that
443 * it's OK to supply a NULL \a umd.eq_handle by calling
444 * LNetInvalidateHandle() on it.
447 LNetMDBind(struct lnet_md umd, enum lnet_unlink unlink,
448 struct lnet_handle_md *handle)
450 struct lnet_libmd *md;
454 LASSERT(the_lnet.ln_refcount > 0);
456 if (lnet_md_validate(&umd) != 0)
459 if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) != 0) {
460 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
464 md = lnet_md_build(&umd, unlink);
468 if (md->md_length > LNET_MTU) {
469 CERROR("Invalid length: too big transfer size %u, %d max\n",
470 md->md_length, LNET_MTU);
475 cpt = lnet_res_lock_current();
477 rc = lnet_md_link(md, umd.handler, cpt);
481 lnet_md2handle(handle, md);
483 lnet_res_unlock(cpt);
487 lnet_res_unlock(cpt);
493 EXPORT_SYMBOL(LNetMDBind);
496 * Unlink the memory descriptor from any ME it may be linked to and release
497 * the internal resources associated with it. As a result, active messages
498 * associated with the MD may get aborted.
500 * This function does not free the memory region associated with the MD;
501 * i.e., the memory the user allocated for this MD. If the ME associated with
502 * this MD is not NULL and was created with auto unlink enabled, the ME is
503 * unlinked as well (see LNetMEAttach()).
505 * Explicitly unlinking a MD via this function call has the same behavior as
506 * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
507 * is generated in the latter case.
509 * An unlinked event can be reported in two ways:
510 * - If there's no pending operations on the MD, it's unlinked immediately
511 * and an LNET_EVENT_UNLINK event is logged before this function returns.
512 * - Otherwise, the MD is only marked for deletion when this function
513 * returns, and the unlinked event will be piggybacked on the event of
514 * the completion of the last operation by setting the unlinked field of
515 * the event. No dedicated LNET_EVENT_UNLINK event is generated.
517 * Note that in both cases the unlinked field of the event is always set; no
518 * more event will happen on the MD after such an event is logged.
520 * \param mdh A handle for the MD to be unlinked.
522 * \retval 0 On success.
523 * \retval -ENOENT If \a mdh does not point to a valid MD object.
526 LNetMDUnlink(struct lnet_handle_md mdh)
528 struct lnet_event ev;
529 struct lnet_libmd *md;
532 LASSERT(the_lnet.ln_refcount > 0);
534 cpt = lnet_cpt_of_cookie(mdh.cookie);
537 md = lnet_handle2md(&mdh);
539 lnet_res_unlock(cpt);
543 md->md_flags |= LNET_MD_FLAG_ABORTED;
544 /* If the MD is busy, lnet_md_unlink just marks it for deletion, and
545 * when the LND is done, the completion event flags that the MD was
546 * unlinked. Otherwise, we enqueue an event now... */
547 if (md->md_handler && md->md_refcount == 0) {
548 lnet_build_unlink_event(md, &ev);
552 if (md->md_rspt_ptr != NULL)
553 lnet_detach_rsp_tracker(md, cpt);
557 lnet_res_unlock(cpt);
560 EXPORT_SYMBOL(LNetMDUnlink);