X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lnet%2Flnet%2Flib-md.c;h=53a855751833a0c6589e582c039f32feeed23c5d;hb=97084f70a6c2956d49d8dc289aeb8ede84e5f7b4;hp=9a391cde5c94e43ec2599d62787977398ffb98ff;hpb=090c677210ee2946d99c71412e4ff762bb300f4f;p=fs%2Flustre-release.git diff --git a/lnet/lnet/lib-md.c b/lnet/lnet/lib-md.c index 9a391cd..53a8557 100644 --- a/lnet/lnet/lib-md.c +++ b/lnet/lnet/lib-md.c @@ -1,458 +1,558 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * lib/lib-md.c - * Memory Descriptor management routines + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * Copyright (c) 2001-2003 Cluster File Systems, Inc. - * Copyright (c) 2001-2002 Sandia National Laboratories + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * This file is part of Lustre, http://www.sf.net/projects/lustre/ + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.gnu.org/licenses/gpl-2.0.html + * + * GPL HEADER END + */ +/* + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * Copyright (c) 2012, 2017, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * lnet/lnet/lib-md.c * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * Memory Descriptor management routines */ -#ifndef __KERNEL__ -# include -#else -# define DEBUG_SUBSYSTEM S_PORTALS -# include -#endif +#define DEBUG_SUBSYSTEM S_LNET -#include -#include +#include -/* must be called with state lock held */ -void lib_md_unlink(nal_cb_t * nal, lib_md_t * md) +/* must be called with lnet_res_lock held */ +void +lnet_md_unlink(struct lnet_libmd *md) { - if ((md->md_flags & PTL_MD_FLAG_ZOMBIE) == 0) { - /* first unlink attempt... */ - lib_me_t *me = md->me; - - md->md_flags |= PTL_MD_FLAG_ZOMBIE; - - /* Disassociate from ME (if any), and unlink it if it was created - * with PTL_UNLINK */ - if (me != NULL) { - me->md = NULL; - if (me->unlink == PTL_UNLINK) - lib_me_unlink(nal, me); - } - - /* emsure all future handle lookups fail */ - lib_invalidate_handle(nal, &md->md_lh); - } - - if (md->pending != 0) { - CDEBUG(D_NET, "Queueing unlink of md %p\n", md); - return; - } - - CDEBUG(D_NET, "Unlinking md %p\n", md); - - if ((md->options & PTL_MD_KIOV) != 0) { - if (nal->cb_unmap_pages != NULL) - nal->cb_unmap_pages (nal, md->md_niov, md->md_iov.kiov, - &md->md_addrkey); - } else if (nal->cb_unmap != NULL) { - nal->cb_unmap (nal, md->md_niov, md->md_iov.iov, - &md->md_addrkey); - } - - if (md->eq != NULL) { - md->eq->eq_refcount--; - LASSERT (md->eq->eq_refcount >= 0); - } - - list_del (&md->md_list); - lib_md_free(nal, md); + if ((md->md_flags & LNET_MD_FLAG_ZOMBIE) == 0) { + /* first unlink attempt... */ + struct lnet_me *me = md->md_me; + + md->md_flags |= LNET_MD_FLAG_ZOMBIE; + + /* Disassociate from ME (if any), and unlink it if it was created + * with LNET_UNLINK */ + if (me != NULL) { + /* detach MD from portal */ + lnet_ptl_detach_md(me, md); + if (me->me_unlink == LNET_UNLINK) + lnet_me_unlink(me); + } + + /* ensure all future handle lookups fail */ + lnet_res_lh_invalidate(&md->md_lh); + } + + if (md->md_refcount != 0) { + CDEBUG(D_NET, "Queueing unlink of md %p\n", md); + return; + } + + CDEBUG(D_NET, "Unlinking md %p\n", md); + + if (md->md_eq != NULL) { + int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie); + + LASSERT(*md->md_eq->eq_refs[cpt] > 0); + (*md->md_eq->eq_refs[cpt])--; + } + + LASSERT(!list_empty(&md->md_list)); + list_del_init(&md->md_list); + lnet_md_free(md); } -/* must be called with state lock held */ -static int lib_md_build(nal_cb_t *nal, lib_md_t *new, void *private, - ptl_md_t *md, ptl_handle_eq_t *eqh, int unlink) +struct page * +lnet_kvaddr_to_page(unsigned long vaddr) { - lib_eq_t *eq = NULL; - int rc; - int i; - - /* NB we are passed an allocated, but uninitialised/active md. - * if we return success, caller may lib_md_unlink() it. - * otherwise caller may only lib_md_free() it. - */ - - if (!PtlHandleIsEqual (*eqh, PTL_EQ_NONE)) { - eq = ptl_handle2eq(eqh, nal); - if (eq == NULL) - return PTL_EQ_INVALID; - } - - /* Must check this _before_ allocation. Also, note that non-iov - * MDs must set md_niov to 0. */ - LASSERT((md->options & (PTL_MD_IOVEC | PTL_MD_KIOV)) == 0 || - md->niov <= PTL_MD_MAX_IOV); - - /* This implementation doesn't know how to create START events or - * disable END events. Best to LASSERT our caller is compliant so - * we find out quickly... */ - LASSERT (PtlHandleIsEqual (*eqh, PTL_EQ_NONE) || - ((md->options & PTL_MD_EVENT_START_DISABLE) != 0 && - (md->options & PTL_MD_EVENT_END_DISABLE) == 0)); - - if ((md->options & PTL_MD_MAX_SIZE) != 0 && /* max size used */ - (md->max_size < 0 || md->max_size > md->length)) // illegal max_size - return PTL_MD_INVALID; - - new->me = NULL; - new->start = md->start; - new->length = md->length; - new->offset = 0; - new->max_size = md->max_size; - new->options = md->options; - new->user_ptr = md->user_ptr; - new->eq = eq; - new->threshold = md->threshold; - new->pending = 0; - new->md_flags = (unlink == PTL_UNLINK) ? PTL_MD_FLAG_AUTO_UNLINK : 0; - - if ((md->options & PTL_MD_IOVEC) != 0) { - int total_length = 0; - - if ((md->options & PTL_MD_KIOV) != 0) /* Can't specify both */ - return PTL_MD_INVALID; - - new->md_niov = md->niov; - - if (nal->cb_read (nal, private, new->md_iov.iov, md->start, - md->niov * sizeof (new->md_iov.iov[0]))) - return PTL_SEGV; - - for (i = 0; i < new->md_niov; i++) { - /* We take the base address on trust */ - if (new->md_iov.iov[i].iov_len <= 0) /* invalid length */ - return PTL_VAL_FAILED; - - total_length += new->md_iov.iov[i].iov_len; - } - - if (md->length > total_length) - return PTL_IOV_TOO_SMALL; - - if (nal->cb_map != NULL) { - rc = nal->cb_map (nal, new->md_niov, new->md_iov.iov, - &new->md_addrkey); - if (rc != PTL_OK) - return (rc); - } - } else if ((md->options & PTL_MD_KIOV) != 0) { -#ifndef __KERNEL__ - return PTL_MD_INVALID; + if (is_vmalloc_addr((void *)vaddr)) + return vmalloc_to_page((void *)vaddr); + +#ifdef CONFIG_HIGHMEM + +#ifdef HAVE_KMAP_TO_PAGE + /* + * This ifdef is added to handle the kernel versions + * which have kmap_to_page() function exported. If so, + * we should use it. Otherwise, remain with the legacy check. + */ + return kmap_to_page((void *)vaddr); #else - int total_length = 0; - - /* Trap attempt to use paged I/O if unsupported early. */ - if (nal->cb_send_pages == NULL || - nal->cb_recv_pages == NULL) - return PTL_MD_INVALID; - - new->md_niov = md->niov; - - if (nal->cb_read (nal, private, new->md_iov.kiov, md->start, - md->niov * sizeof (new->md_iov.kiov[0]))) - return PTL_SEGV; - - for (i = 0; i < new->md_niov; i++) { - /* We take the page pointer on trust */ - if (new->md_iov.kiov[i].kiov_offset + - new->md_iov.kiov[i].kiov_len > PAGE_SIZE ) - return PTL_VAL_FAILED; /* invalid length */ - - total_length += new->md_iov.kiov[i].kiov_len; - } - - if (md->length > total_length) - return PTL_IOV_TOO_SMALL; - - if (nal->cb_map_pages != NULL) { - rc = nal->cb_map_pages (nal, new->md_niov, new->md_iov.kiov, - &new->md_addrkey); - if (rc != PTL_OK) - return (rc); - } -#endif - } else { /* contiguous */ - new->md_niov = 1; - new->md_iov.iov[0].iov_base = md->start; - new->md_iov.iov[0].iov_len = md->length; - - if (nal->cb_map != NULL) { - rc = nal->cb_map (nal, new->md_niov, new->md_iov.iov, - &new->md_addrkey); - if (rc != PTL_OK) - return (rc); - } - } - - if (eq != NULL) - eq->eq_refcount++; - - /* It's good; let handle2md succeed and add to active mds */ - lib_initialise_handle (nal, &new->md_lh, PTL_COOKIE_TYPE_MD); - list_add (&new->md_list, &nal->ni.ni_active_mds); - - return PTL_OK; + + if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) { + /* No highmem pages only used for bulk (kiov) I/O */ + CERROR("find page for address in highmem\n"); + LBUG(); + } + return virt_to_page(vaddr); +#endif /* HAVE_KMAP_TO_PAGE */ +#else + + return virt_to_page(vaddr); +#endif /* CONFIG_HIGHMEM */ } +EXPORT_SYMBOL(lnet_kvaddr_to_page); -/* must be called with state lock held */ -void lib_md_deconstruct(nal_cb_t * nal, lib_md_t * md, ptl_md_t * new) +int +lnet_cpt_of_md(struct lnet_libmd *md, unsigned int offset) { - /* NB this doesn't copy out all the iov entries so when a - * discontiguous MD is copied out, the target gets to know the - * original iov pointer (in start) and the number of entries it had - * and that's all. - */ - new->start = md->start; - new->length = md->length; - new->threshold = md->threshold; - new->max_size = md->max_size; - new->options = md->options; - new->user_ptr = md->user_ptr; - ptl_eq2handle(&new->eventq, md->eq); - new->niov = ((md->options & (PTL_MD_IOVEC | PTL_MD_KIOV)) == 0) ? 0 : md->md_niov; + int cpt = CFS_CPT_ANY; + unsigned int niov; + + /* + * if the md_options has a bulk handle then we want to look at the + * bulk md because that's the data which we will be DMAing + */ + if (md && (md->md_options & LNET_MD_BULK_HANDLE) != 0 && + !LNetMDHandleIsInvalid(md->md_bulk_handle)) + md = lnet_handle2md(&md->md_bulk_handle); + + if (!md || md->md_niov == 0) + return CFS_CPT_ANY; + + niov = md->md_niov; + + /* + * There are three cases to handle: + * 1. The MD is using lnet_kiov_t + * 2. The MD is using struct kvec + * 3. Contiguous buffer allocated via vmalloc + * + * in case 2 we can use virt_to_page() macro to get the page + * address of the memory kvec describes. + * + * in case 3 use is_vmalloc_addr() and vmalloc_to_page() + * + * The offset provided can be within the first iov/kiov entry or + * it could go beyond it. In that case we need to make sure to + * look at the page which actually contains the data that will be + * DMAed. + */ + if ((md->md_options & LNET_MD_KIOV) != 0) { + lnet_kiov_t *kiov = md->md_iov.kiov; + + while (offset >= kiov->kiov_len) { + offset -= kiov->kiov_len; + niov--; + kiov++; + if (niov == 0) { + CERROR("offset %d goes beyond kiov\n", offset); + goto out; + } + } + + cpt = cfs_cpt_of_node(lnet_cpt_table(), + page_to_nid(kiov->kiov_page)); + } else { + struct kvec *iov = md->md_iov.iov; + unsigned long vaddr; + struct page *page; + + while (offset >= iov->iov_len) { + offset -= iov->iov_len; + niov--; + iov++; + if (niov == 0) { + CERROR("offset %d goes beyond iov\n", offset); + goto out; + } + } + + vaddr = ((unsigned long)iov->iov_base) + offset; + page = lnet_kvaddr_to_page(vaddr); + if (!page) { + CERROR("Couldn't resolve vaddr 0x%lx to page\n", vaddr); + goto out; + } + cpt = cfs_cpt_of_node(lnet_cpt_table(), page_to_nid(page)); + } + +out: + return cpt; } -int do_PtlMDAttach(nal_cb_t * nal, void *private, void *v_args, void *v_ret) +static int +lnet_md_build(struct lnet_libmd *lmd, struct lnet_md *umd, int unlink) { - /* - * Incoming: - * ptl_handle_me_t current_in - * ptl_md_t md_in - * ptl_unlink_t unlink_in - * - * Outgoing: - * ptl_handle_md_t * handle_out - */ - - PtlMDAttach_in *args = v_args; - PtlMDAttach_out *ret = v_ret; - lib_me_t *me; - lib_md_t *md; - unsigned long flags; - - if ((args->md_in.options & (PTL_MD_KIOV | PTL_MD_IOVEC)) != 0 && - args->md_in.niov > PTL_MD_MAX_IOV) /* too many fragments */ - return (ret->rc = PTL_IOV_TOO_MANY); - - md = lib_md_alloc(nal, &args->md_in); - if (md == NULL) - return (ret->rc = PTL_NO_SPACE); - - state_lock(nal, &flags); - - me = ptl_handle2me(&args->me_in, nal); - if (me == NULL) { - ret->rc = PTL_ME_INVALID; - } else if (me->md != NULL) { - ret->rc = PTL_ME_IN_USE; - } else { - ret->rc = lib_md_build(nal, md, private, &args->md_in, - &args->eq_in, args->unlink_in); - - if (ret->rc == PTL_OK) { - me->md = md; - md->me = me; - - ptl_md2handle(&ret->handle_out, md); - - state_unlock (nal, &flags); - return (PTL_OK); - } - } - - lib_md_free (nal, md); - - state_unlock (nal, &flags); - return (ret->rc); + int i; + unsigned int niov; + int total_length = 0; + + lmd->md_me = NULL; + lmd->md_start = umd->start; + lmd->md_offset = 0; + lmd->md_max_size = umd->max_size; + lmd->md_options = umd->options; + lmd->md_user_ptr = umd->user_ptr; + lmd->md_eq = NULL; + lmd->md_threshold = umd->threshold; + lmd->md_refcount = 0; + lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0; + lmd->md_bulk_handle = umd->bulk_handle; + + if ((umd->options & LNET_MD_IOVEC) != 0) { + + if ((umd->options & LNET_MD_KIOV) != 0) /* Can't specify both */ + return -EINVAL; + + lmd->md_niov = niov = umd->length; + memcpy(lmd->md_iov.iov, umd->start, + niov * sizeof(lmd->md_iov.iov[0])); + + for (i = 0; i < (int)niov; i++) { + /* We take the base address on trust */ + if (lmd->md_iov.iov[i].iov_len <= 0) /* invalid length */ + return -EINVAL; + + total_length += lmd->md_iov.iov[i].iov_len; + } + + lmd->md_length = total_length; + + if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */ + (umd->max_size < 0 || + umd->max_size > total_length)) // illegal max_size + return -EINVAL; + + } else if ((umd->options & LNET_MD_KIOV) != 0) { + lmd->md_niov = niov = umd->length; + memcpy(lmd->md_iov.kiov, umd->start, + niov * sizeof(lmd->md_iov.kiov[0])); + + for (i = 0; i < (int)niov; i++) { + /* We take the page pointer on trust */ + if (lmd->md_iov.kiov[i].kiov_offset + + lmd->md_iov.kiov[i].kiov_len > PAGE_SIZE) + return -EINVAL; /* invalid length */ + + total_length += lmd->md_iov.kiov[i].kiov_len; + } + + lmd->md_length = total_length; + + if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */ + (umd->max_size < 0 || + umd->max_size > total_length)) // illegal max_size + return -EINVAL; + } else { /* contiguous */ + lmd->md_length = umd->length; + lmd->md_niov = niov = 1; + lmd->md_iov.iov[0].iov_base = umd->start; + lmd->md_iov.iov[0].iov_len = umd->length; + + if ((umd->options & LNET_MD_MAX_SIZE) != 0 && /* max size used */ + (umd->max_size < 0 || + umd->max_size > (int)umd->length)) // illegal max_size + return -EINVAL; + } + + return 0; } -int do_PtlMDBind(nal_cb_t * nal, void *private, void *v_args, void *v_ret) +/* must be called with resource lock held */ +static int +lnet_md_link(struct lnet_libmd *md, struct lnet_eq *eq, int cpt) { - /* - * Incoming: - * ptl_handle_ni_t ni_in - * ptl_md_t md_in - * - * Outgoing: - * ptl_handle_md_t * handle_out - */ + struct lnet_res_container *container = the_lnet.ln_md_containers[cpt]; + + /* NB we are passed an allocated, but inactive md. + * if we return success, caller may lnet_md_unlink() it. + * otherwise caller may only lnet_md_free() it. + */ + /* This implementation doesn't know how to create START events or + * disable END events. Best to LASSERT our caller is compliant so + * we find out quickly... */ + /* TODO - reevaluate what should be here in light of + * the removal of the start and end events + * maybe there we shouldn't even allow LNET_EQ_NONE!) + * LASSERT (eq == NULL); + */ + if (eq) { + md->md_eq = eq; + (*md->md_eq->eq_refs[cpt])++; + } + + lnet_res_lh_initialize(container, &md->md_lh); + + LASSERT(list_empty(&md->md_list)); + list_add(&md->md_list, &container->rec_active); + + return 0; +} - PtlMDBind_in *args = v_args; - PtlMDBind_out *ret = v_ret; - lib_md_t *md; - unsigned long flags; +/* must be called with lnet_res_lock held */ +void +lnet_md_deconstruct(struct lnet_libmd *lmd, struct lnet_md *umd) +{ + /* NB this doesn't copy out all the iov entries so when a + * discontiguous MD is copied out, the target gets to know the + * original iov pointer (in start) and the number of entries it had + * and that's all. + */ + umd->start = lmd->md_start; + umd->length = ((lmd->md_options & (LNET_MD_IOVEC | LNET_MD_KIOV)) == 0) ? + lmd->md_length : lmd->md_niov; + umd->threshold = lmd->md_threshold; + umd->max_size = lmd->md_max_size; + umd->options = lmd->md_options; + umd->user_ptr = lmd->md_user_ptr; +} - if ((args->md_in.options & (PTL_MD_KIOV | PTL_MD_IOVEC)) != 0 && - args->md_in.niov > PTL_MD_MAX_IOV) /* too many fragments */ - return (ret->rc = PTL_IOV_TOO_MANY); +static int +lnet_md_validate(struct lnet_md *umd) +{ + if (umd->start == NULL && umd->length != 0) { + CERROR("MD start pointer can not be NULL with length %u\n", + umd->length); + return -EINVAL; + } + + if ((umd->options & (LNET_MD_KIOV | LNET_MD_IOVEC)) != 0 && + umd->length > LNET_MAX_IOV) { + CERROR("Invalid option: too many fragments %u, %d max\n", + umd->length, LNET_MAX_IOV); + return -EINVAL; + } + + return 0; +} - md = lib_md_alloc(nal, &args->md_in); - if (md == NULL) - return (ret->rc = PTL_NO_SPACE); +/** + * Create a memory descriptor and attach it to a ME + * + * \param me An ME to associate the new MD with. + * \param umd Provides initial values for the user-visible parts of a MD. + * Other than its use for initialization, there is no linkage between this + * structure and the MD maintained by the LNet. + * \param unlink A flag to indicate whether the MD is automatically unlinked + * when it becomes inactive, either because the operation threshold drops to + * zero or because the available memory becomes less than \a umd.max_size. + * (Note that the check for unlinking a MD only occurs after the completion + * of a successful operation on the MD.) The value LNET_UNLINK enables auto + * unlinking; the value LNET_RETAIN disables it. + * \param handle On successful returns, a handle to the newly created MD is + * saved here. This handle can be used later in LNetMDUnlink(). + * + * \retval 0 On success. + * \retval -EINVAL If \a umd is not valid. + * \retval -ENOMEM If new MD cannot be allocated. + * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a + * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by + * calling LNetInvalidateHandle() on it. + * \retval -EBUSY If the ME pointed to by \a meh is already associated with + * a MD. + */ +int +LNetMDAttach(struct lnet_me *me, struct lnet_md umd, + enum lnet_unlink unlink, struct lnet_handle_md *handle) +{ + LIST_HEAD(matches); + LIST_HEAD(drops); + struct lnet_libmd *md; + int cpt; + int rc; + + LASSERT(the_lnet.ln_refcount > 0); + + if (lnet_md_validate(&umd) != 0) + return -EINVAL; + + if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) == 0) { + CERROR("Invalid option: no MD_OP set\n"); + return -EINVAL; + } - state_lock(nal, &flags); + md = lnet_md_alloc(&umd); + if (md == NULL) + return -ENOMEM; - ret->rc = lib_md_build(nal, md, private, &args->md_in, - &args->eq_in, args->unlink_in); + rc = lnet_md_build(md, &umd, unlink); + if (rc != 0) + goto out_free; - if (ret->rc == PTL_OK) { - ptl_md2handle(&ret->handle_out, md); + cpt = me->me_cpt; - state_unlock(nal, &flags); - return (PTL_OK); - } + lnet_res_lock(cpt); - lib_md_free (nal, md); + if (me->me_md) + rc = -EBUSY; + else + rc = lnet_md_link(md, umd.eq_handle, cpt); - state_unlock(nal, &flags); - return (ret->rc); + if (rc != 0) + goto out_unlock; + + /* attach this MD to portal of ME and check if it matches any + * blocked msgs on this portal */ + lnet_ptl_attach_md(me, md, &matches, &drops); + + lnet_md2handle(handle, md); + + lnet_res_unlock(cpt); + + lnet_drop_delayed_msg_list(&drops, "Bad match"); + lnet_recv_delayed_msg_list(&matches); + + return 0; + +out_unlock: + lnet_res_unlock(cpt); +out_free: + lnet_md_free(md); + return rc; } +EXPORT_SYMBOL(LNetMDAttach); -int do_PtlMDUnlink(nal_cb_t * nal, void *private, void *v_args, void *v_ret) +/** + * Create a "free floating" memory descriptor - a MD that is not associated + * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations. + * + * \param umd,unlink See the discussion for LNetMDAttach(). + * \param handle On successful returns, a handle to the newly created MD is + * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(), + * and LNetGet() operations. + * + * \retval 0 On success. + * \retval -EINVAL If \a umd is not valid. + * \retval -ENOMEM If new MD cannot be allocated. + * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that + * it's OK to supply a NULL \a umd.eq_handle by calling + * LNetInvalidateHandle() on it. + */ +int +LNetMDBind(struct lnet_md umd, enum lnet_unlink unlink, + struct lnet_handle_md *handle) { - PtlMDUnlink_in *args = v_args; - PtlMDUnlink_out *ret = v_ret; - ptl_event_t ev; - lib_md_t *md; - unsigned long flags; - - state_lock(nal, &flags); - - md = ptl_handle2md(&args->md_in, nal); - if (md == NULL) { - state_unlock(nal, &flags); - return (ret->rc = PTL_MD_INVALID); - } - - /* If the MD is busy, lib_md_unlink just marks it for deletion, and - * when the NAL is done, the completion event flags that the MD was - * unlinked. Otherwise, we enqueue an event now... */ - - if (md->eq != NULL && - md->pending == 0) { - memset(&ev, 0, sizeof(ev)); - - ev.type = PTL_EVENT_UNLINK; - ev.ni_fail_type = PTL_OK; - ev.unlinked = 1; - lib_md_deconstruct(nal, md, &ev.mem_desc); - - lib_enq_event_locked(nal, private, md->eq, &ev); - } - - lib_md_deconstruct(nal, md, &ret->status_out); - lib_md_unlink(nal, md); - ret->rc = PTL_OK; - - state_unlock(nal, &flags); - - return (PTL_OK); + struct lnet_libmd *md; + int cpt; + int rc; + + LASSERT(the_lnet.ln_refcount > 0); + + if (lnet_md_validate(&umd) != 0) + return -EINVAL; + + if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) != 0) { + CERROR("Invalid option: GET|PUT illegal on active MDs\n"); + return -EINVAL; + } + + md = lnet_md_alloc(&umd); + if (md == NULL) + return -ENOMEM; + + rc = lnet_md_build(md, &umd, unlink); + if (rc != 0) + goto out_free; + + if (md->md_length > LNET_MTU) { + CERROR("Invalid length: too big transfer size %u, %d max\n", + md->md_length, LNET_MTU); + rc = -EINVAL; + goto out_free; + } + + cpt = lnet_res_lock_current(); + + rc = lnet_md_link(md, umd.eq_handle, cpt); + if (rc != 0) + goto out_unlock; + + lnet_md2handle(handle, md); + + lnet_res_unlock(cpt); + return 0; + + out_unlock: + lnet_res_unlock(cpt); + + out_free: + lnet_md_free(md); + return rc; } +EXPORT_SYMBOL(LNetMDBind); -int do_PtlMDUpdate_internal(nal_cb_t * nal, void *private, void *v_args, - void *v_ret) +/** + * Unlink the memory descriptor from any ME it may be linked to and release + * the internal resources associated with it. As a result, active messages + * associated with the MD may get aborted. + * + * This function does not free the memory region associated with the MD; + * i.e., the memory the user allocated for this MD. If the ME associated with + * this MD is not NULL and was created with auto unlink enabled, the ME is + * unlinked as well (see LNetMEAttach()). + * + * Explicitly unlinking a MD via this function call has the same behavior as + * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK + * is generated in the latter case. + * + * An unlinked event can be reported in two ways: + * - If there's no pending operations on the MD, it's unlinked immediately + * and an LNET_EVENT_UNLINK event is logged before this function returns. + * - Otherwise, the MD is only marked for deletion when this function + * returns, and the unlinked event will be piggybacked on the event of + * the completion of the last operation by setting the unlinked field of + * the event. No dedicated LNET_EVENT_UNLINK event is generated. + * + * Note that in both cases the unlinked field of the event is always set; no + * more event will happen on the MD after such an event is logged. + * + * \param mdh A handle for the MD to be unlinked. + * + * \retval 0 On success. + * \retval -ENOENT If \a mdh does not point to a valid MD object. + */ +int +LNetMDUnlink(struct lnet_handle_md mdh) { - /* - * Incoming: - * ptl_handle_md_t md_in - * ptl_md_t * old_inout - * ptl_md_t * new_inout - * ptl_handle_eq_t testq_in - * ptl_seq_t sequence_in - * - * Outgoing: - * ptl_md_t * old_inout - * ptl_md_t * new_inout - */ - PtlMDUpdate_internal_in *args = v_args; - PtlMDUpdate_internal_out *ret = v_ret; - lib_md_t *md; - lib_eq_t *test_eq = NULL; - ptl_md_t *new = &args->new_inout; - unsigned long flags; - - state_lock(nal, &flags); - - md = ptl_handle2md(&args->md_in, nal); - if (md == NULL) { - ret->rc = PTL_MD_INVALID; - goto out; - } - - if (args->old_inout_valid) - lib_md_deconstruct(nal, md, &ret->old_inout); - - if (!args->new_inout_valid) { - ret->rc = PTL_OK; - goto out; - } - - /* XXX fttb, the new MD must be the same type wrt fragmentation */ - if (((new->options ^ md->options) & - (PTL_MD_IOVEC | PTL_MD_KIOV)) != 0) { - ret->rc = PTL_MD_INVALID; - goto out; - } - - if (new->niov > md->md_niov) { - ret->rc = PTL_IOV_TOO_MANY; - goto out; - } - - if (new->niov < md->md_niov) { - ret->rc = PTL_IOV_TOO_SMALL; - goto out; - } - - if (!PtlHandleIsEqual (args->testq_in, PTL_EQ_NONE)) { - test_eq = ptl_handle2eq(&args->testq_in, nal); - if (test_eq == NULL) { - ret->rc = PTL_EQ_INVALID; - goto out; - } - } - - if (md->pending != 0) { - ret->rc = PTL_MD_NO_UPDATE; - goto out; - } - - if (test_eq == NULL || - test_eq->sequence == args->sequence_in) { - lib_me_t *me = md->me; - int unlink = (md->md_flags & PTL_MD_FLAG_AUTO_UNLINK) ? - PTL_UNLINK : PTL_RETAIN; - - // #warning this does not track eq refcounts properly - ret->rc = lib_md_build(nal, md, private, - new, &new->eventq, unlink); - - md->me = me; - } else { - ret->rc = PTL_MD_NO_UPDATE; - } - - out: - state_unlock(nal, &flags); - return (ret->rc); + struct lnet_event ev; + struct lnet_libmd *md; + int cpt; + + LASSERT(the_lnet.ln_refcount > 0); + + cpt = lnet_cpt_of_cookie(mdh.cookie); + lnet_res_lock(cpt); + + md = lnet_handle2md(&mdh); + if (md == NULL) { + lnet_res_unlock(cpt); + return -ENOENT; + } + + md->md_flags |= LNET_MD_FLAG_ABORTED; + /* If the MD is busy, lnet_md_unlink just marks it for deletion, and + * when the LND is done, the completion event flags that the MD was + * unlinked. Otherwise, we enqueue an event now... */ + if (md->md_eq != NULL && md->md_refcount == 0) { + lnet_build_unlink_event(md, &ev); + lnet_eq_enqueue_event(md->md_eq, &ev); + } + + if (md->md_rspt_ptr != NULL) + lnet_detach_rsp_tracker(md, cpt); + + lnet_md_unlink(md); + + lnet_res_unlock(cpt); + return 0; } +EXPORT_SYMBOL(LNetMDUnlink);