-/* -*- 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.
+ *
+ * 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).
+ *
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
*
- * Copyright (c) 2001-2003 Cluster File Systems, Inc.
- * Copyright (c) 2001-2002 Sandia National Laboratories
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
*
- * This file is part of Lustre, http://www.sf.net/projects/lustre/
+ * 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, 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 <stdio.h>
-#else
-# define DEBUG_SUBSYSTEM S_PORTALS
-# include <linux/kp30.h>
-#endif
+#define DEBUG_SUBSYSTEM S_LNET
-#include <portals/lib-p30.h>
-#include <portals/arg-blocks.h>
+#include <lnet/lib-lnet.h>
-/*
- * 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(lnet_libmd_t *md)
{
- lib_me_t *me = md->me;
+ if ((md->md_flags & LNET_MD_FLAG_ZOMBIE) == 0) {
+ /* first unlink attempt... */
+ lnet_me_t *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->pending != 0) {
+ if (md->md_refcount != 0) {
CDEBUG(D_NET, "Queueing unlink of md %p\n", md);
- md->md_flags |= PTL_MD_FLAG_UNLINK;
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 (me) {
- me->md = NULL;
- if (me->unlink == PTL_UNLINK)
- lib_me_unlink(nal, me);
- }
+ if (md->md_eq != NULL) {
+ int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
- if (md->eq != NULL)
- {
- md->eq->eq_refcount--;
- LASSERT (md->eq->eq_refcount >= 0);
- }
+ LASSERT(*md->md_eq->eq_refs[cpt] > 0);
+ (*md->md_eq->eq_refs[cpt])--;
+ }
- lib_invalidate_handle (nal, &md->md_lh);
- list_del (&md->md_list);
- lib_md_free(nal, md);
+ LASSERT(!cfs_list_empty(&md->md_list));
+ cfs_list_del_init(&md->md_list);
+ lnet_md_free_locked(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)
+static int
+lnet_md_build(lnet_libmd_t *lmd, lnet_md_t *umd, int unlink)
{
- const int max_size_opts = PTL_MD_AUTO_UNLINK |
- PTL_MD_MAX_SIZE;
- lib_eq_t *eq = NULL;
- int rc;
- int i;
-
- /* NB we are passes 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 (!PtlHandleEqual (*eqh, PTL_EQ_NONE)) {
- eq = ptl_handle2eq(eqh, nal);
- if (eq == NULL)
- return PTL_INV_EQ;
- }
-
- if ((md->options & PTL_MD_IOV) != 0 && /* discontiguous MD */
- md->niov > PTL_MD_MAX_IOV) /* too many fragments */
- return PTL_IOV_TOO_MANY;
-
- if ((md->options & max_size_opts) != 0 && /* max size used */
- (md->max_size < 0 || md->max_size > md->length)) // illegal max_size
- return PTL_INV_MD;
-
- new->me = NULL;
- new->start = md->start;
- new->length = md->length;
- new->offset = 0;
- new->max_size = md->max_size;
- new->unlink = unlink;
- new->options = md->options;
- new->user_ptr = md->user_ptr;
- new->eq = eq;
- new->threshold = md->threshold;
- new->pending = 0;
- new->md_flags = 0;
-
- if ((md->options & PTL_MD_IOV) != 0) {
- int total_length = 0;
-
- if ((md->options & PTL_MD_KIOV) != 0) /* Can't specify both */
- return PTL_INV_MD;
-
- 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++) {
+ 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;
+
+ 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 (new->md_iov.iov[i].iov_len <= 0) /* invalid length */
- return PTL_VAL_FAILED;
+ if (lmd->md_iov.iov[i].iov_len <= 0) /* invalid length */
+ return -EINVAL;
- total_length += new->md_iov.iov[i].iov_len;
+ total_length += lmd->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) {
+ 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) {
#ifndef __KERNEL__
- return PTL_INV_MD;
+ return -EINVAL;
#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_INV_MD;
-
- 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++) {
+ 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 (new->md_iov.kiov[i].kiov_offset +
- new->md_iov.kiov[i].kiov_len > PAGE_SIZE )
- return PTL_VAL_FAILED; /* invalid length */
+ if (lmd->md_iov.kiov[i].kiov_offset +
+ lmd->md_iov.kiov[i].kiov_len > PAGE_CACHE_SIZE)
+ return -EINVAL; /* invalid length */
- total_length += new->md_iov.kiov[i].kiov_len;
+ total_length += lmd->md_iov.kiov[i].kiov_len;
}
- if (md->length > total_length)
- return PTL_IOV_TOO_SMALL;
+ lmd->md_length = total_length;
- 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);
- }
+ 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;
#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++;
+ 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;
+ }
- /* 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 0;
+}
- return PTL_OK;
+/* must be called with resource lock held */
+static int
+lnet_md_link(lnet_libmd_t *md, lnet_handle_eq_t eq_handle, int cpt)
+{
+ 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 (!LNetHandleIsInvalid(eq_handle)) {
+ md->md_eq = lnet_handle2eq(&eq_handle);
+
+ if (md->md_eq == NULL)
+ return -ENOENT;
+
+ (*md->md_eq->eq_refs[cpt])++;
+ }
+
+ lnet_res_lh_initialize(container, &md->md_lh);
+
+ LASSERT(cfs_list_empty(&md->md_list));
+ cfs_list_add(&md->md_list, &container->rec_active);
+
+ return 0;
}
-/* must be called with state lock held */
-void lib_md_deconstruct(nal_cb_t * nal, lib_md_t * md, ptl_md_t * new)
+/* must be called with lnet_res_lock held */
+void
+lnet_md_deconstruct(lnet_libmd_t *lmd, lnet_md_t *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.
*/
- 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_IOV | PTL_MD_KIOV)) == 0) ? 0 : md->md_niov;
+ 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;
+ lnet_eq2handle(&umd->eq_handle, lmd->md_eq);
}
-int do_PtlMDAttach(nal_cb_t * nal, void *private, void *v_args, void *v_ret)
+int
+lnet_md_validate(lnet_md_t *umd)
{
- /*
- * Incoming:
- * ptl_handle_me_t current_in
- * ptl_md_t md_in
- * ptl_unlink_t unlink_in
- *
- * Outgoing:
- * ptl_handle_md_t * handle_out
- */
+ 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;
+ }
- PtlMDAttach_in *args = v_args;
- PtlMDAttach_out *ret = v_ret;
- lib_me_t *me;
- lib_md_t *md;
- unsigned long flags;
+ return 0;
+}
- md = lib_md_alloc (nal);
+/**
+ * Create a memory descriptor and attach it to a ME
+ *
+ * \param meh A handle for a 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(lnet_handle_me_t meh, lnet_md_t umd,
+ lnet_unlink_t unlink, lnet_handle_md_t *handle)
+{
+ CFS_LIST_HEAD (matches);
+ CFS_LIST_HEAD (drops);
+ struct lnet_me *me;
+ struct lnet_libmd *md;
+ int cpt;
+ int rc;
+
+ LASSERT (the_lnet.ln_init);
+ 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;
+ }
+
+ md = lnet_md_alloc(&umd);
if (md == NULL)
- return (ret->rc = PTL_NOSPACE);
+ return -ENOMEM;
- state_lock(nal, &flags);
+ rc = lnet_md_build(md, &umd, unlink);
+ cpt = lnet_cpt_of_cookie(meh.cookie);
- me = ptl_handle2me(&args->me_in, nal);
- if (me == NULL) {
- ret->rc = PTL_INV_ME;
- } else if (me->md != NULL) {
- ret->rc = PTL_INUSE;
- } else {
- ret->rc = lib_md_build(nal, md, private, &args->md_in,
- &args->eq_in, args->unlink_in);
+ lnet_res_lock(cpt);
+ if (rc != 0)
+ goto failed;
- if (ret->rc == PTL_OK) {
- me->md = md;
- md->me = me;
+ me = lnet_handle2me(&meh);
+ if (me == NULL)
+ rc = -ENOENT;
+ else if (me->me_md != NULL)
+ rc = -EBUSY;
+ else
+ rc = lnet_md_link(md, umd.eq_handle, cpt);
- ptl_md2handle(&ret->handle_out, md);
+ if (rc != 0)
+ goto failed;
- state_unlock (nal, &flags);
- return (PTL_OK);
- }
- }
+ /* 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);
- lib_md_free (nal, md);
+ lnet_drop_delayed_msg_list(&drops, "Bad match");
+ lnet_recv_delayed_msg_list(&matches);
- state_unlock (nal, &flags);
- return (ret->rc);
+ return 0;
+
+ failed:
+ lnet_md_free_locked(md);
+
+ lnet_res_unlock(cpt);
+ return rc;
}
+EXPORT_SYMBOL(LNetMDAttach);
-int do_PtlMDBind(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(lnet_md_t umd, lnet_unlink_t unlink, lnet_handle_md_t *handle)
{
- /*
- * Incoming:
- * ptl_handle_ni_t ni_in
- * ptl_md_t md_in
- *
- * Outgoing:
- * ptl_handle_md_t * handle_out
- */
+ lnet_libmd_t *md;
+ int cpt;
+ int rc;
- PtlMDBind_in *args = v_args;
- PtlMDBind_out *ret = v_ret;
- lib_md_t *md;
- unsigned long flags;
+ LASSERT (the_lnet.ln_init);
+ LASSERT (the_lnet.ln_refcount > 0);
- md = lib_md_alloc (nal);
- if (md == NULL)
- return (ret->rc = PTL_NOSPACE);
+ if (lnet_md_validate(&umd) != 0)
+ return -EINVAL;
- state_lock(nal, &flags);
+ if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT)) != 0) {
+ CERROR("Invalid option: GET|PUT illegal on active MDs\n");
+ return -EINVAL;
+ }
- ret->rc = lib_md_build(nal, md, private,
- &args->md_in, &args->eq_in, PTL_UNLINK);
+ md = lnet_md_alloc(&umd);
+ if (md == NULL)
+ return -ENOMEM;
- if (ret->rc == PTL_OK) {
- ptl_md2handle(&ret->handle_out, md);
+ rc = lnet_md_build(md, &umd, unlink);
- state_unlock(nal, &flags);
- return (PTL_OK);
- }
+ cpt = lnet_res_lock_current();
+ if (rc != 0)
+ goto failed;
- lib_md_free (nal, md);
+ rc = lnet_md_link(md, umd.eq_handle, cpt);
+ if (rc != 0)
+ goto failed;
- state_unlock(nal, &flags);
- return (ret->rc);
-}
+ lnet_md2handle(handle, md);
-int do_PtlMDUnlink(nal_cb_t * nal, void *private, void *v_args, void *v_ret)
-{
- PtlMDUnlink_in *args = v_args;
- PtlMDUnlink_out *ret = v_ret;
-
- lib_md_t *md;
- unsigned long flags;
-
- state_lock(nal, &flags);
-
- md = ptl_handle2md(&args->md_in, nal);
- if (md == NULL) {
- ret->rc = PTL_INV_MD;
- } else if (md->pending != 0) { /* being filled/spilled */
- ret->rc = PTL_MD_INUSE;
- } else {
- /* Callers attempting to unlink a busy MD which will get
- * unlinked once the net op completes should see INUSE,
- * before completion and INV_MD thereafter. LASSERT we've
- * got that right... */
- LASSERT ((md->md_flags & PTL_MD_FLAG_UNLINK) == 0);
-
- lib_md_deconstruct(nal, md, &ret->status_out);
- lib_md_unlink(nal, md);
- ret->rc = PTL_OK;
- }
+ lnet_res_unlock(cpt);
+ return 0;
- state_unlock(nal, &flags);
+ failed:
+ lnet_md_free_locked(md);
- return (ret->rc);
+ lnet_res_unlock(cpt);
+ 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.
+ *
+ * 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 (lnet_handle_md_t 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_INV_MD;
- goto out;
- }
+ lnet_event_t ev;
+ lnet_libmd_t *md;
+ int cpt;
- if (args->old_inout_valid)
- lib_md_deconstruct(nal, md, &ret->old_inout);
+ LASSERT(the_lnet.ln_init);
+ LASSERT(the_lnet.ln_refcount > 0);
- if (!args->new_inout_valid) {
- ret->rc = PTL_OK;
- goto out;
- }
+ cpt = lnet_cpt_of_cookie(mdh.cookie);
+ lnet_res_lock(cpt);
- if (!PtlHandleEqual (args->testq_in, PTL_EQ_NONE)) {
- test_eq = ptl_handle2eq(&args->testq_in, nal);
- if (test_eq == NULL) {
- ret->rc = PTL_INV_EQ;
- goto out;
- }
+ md = lnet_handle2md(&mdh);
+ if (md == NULL) {
+ lnet_res_unlock(cpt);
+ return -ENOENT;
}
- if (md->pending != 0) {
- ret->rc = PTL_NOUPDATE;
- goto out;
- }
-
- if (test_eq == NULL ||
- test_eq->sequence == args->sequence_in) {
- lib_me_t *me = md->me;
+ /* If the MD is busy, lnet_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... */
- // #warning this does not track eq refcounts properly
- ret->rc = lib_md_build(nal, md, private,
- new, &new->eventq, md->unlink);
-
- md->me = me;
- } else {
- ret->rc = PTL_NOUPDATE;
+ if (md->md_eq != NULL &&
+ md->md_refcount == 0) {
+ lnet_build_unlink_event(md, &ev);
+ lnet_eq_enqueue_event(md->md_eq, &ev);
}
- out:
- state_unlock(nal, &flags);
- return (ret->rc);
+ lnet_md_unlink(md);
+
+ lnet_res_unlock(cpt);
+ return 0;
}
+EXPORT_SYMBOL(LNetMDUnlink);