* Memory Descriptor management routines
*
* Copyright (c) 2001-2003 Cluster File Systems, Inc.
- * Copyright (c) 2001-2002 Sandia National Laboratories
*
- * This file is part of Lustre, http://www.sf.net/projects/lustre/
+ * This file is part of Lustre, http://www.lustre.org
*
* Lustre is free software; you can redistribute it and/or
* modify it under the terms of version 2 of the GNU General Public
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*/
-#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_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) {
+ md->md_me = NULL;
+ me->me_md = NULL;
+ if (me->me_unlink == LNET_UNLINK)
+ lnet_me_unlink(me);
+ }
+
+ /* ensure all future handle lookups fail */
+ lnet_invalidate_handle(&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) {
+ md->md_eq->eq_refcount--;
+ LASSERT (md->md_eq->eq_refcount >= 0);
}
- if (md->eq != NULL)
- {
- md->eq->eq_refcount--;
- LASSERT (md->eq->eq_refcount >= 0);
- }
-
- lib_invalidate_handle (nal, &md->md_lh);
list_del (&md->md_list);
- lib_md_free(nal, md);
+ 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)
+/* must be called with LNET_LOCK held */
+static int
+lib_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;
+ lnet_eq_t *eq = NULL;
+ int i;
+ unsigned int niov;
+ int total_length = 0;
/* 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 we return success, caller may lnet_md_unlink() it.
+ * otherwise caller may only lnet_md_free() it.
*/
- if (!PtlHandleEqual (*eqh, PTL_EQ_NONE)) {
- eq = ptl_handle2eq(eqh, nal);
+ if (!LNetHandleIsEqual (umd->eq_handle, LNET_EQ_NONE)) {
+ eq = lnet_handle2eq(&umd->eq_handle);
if (eq == NULL)
- return PTL_INV_EQ;
+ return -ENOENT;
}
- /* Must check this _before_ allocation. Also, note that non-iov
- * MDs must set md_niov to 0. */
- LASSERT((md->options & (PTL_MD_IOV | PTL_MD_KIOV)) == 0 ||
- md->niov <= PTL_MD_MAX_IOV);
-
- 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++) {
+ /* 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);
+ */
+
+ 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 = eq;
+ 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 < 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 < 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 > CFS_PAGE_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);
- }
- }
+ 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 > umd->length)) // illegal max_size
+ return -EINVAL;
+ }
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);
+ lnet_initialise_handle (&lmd->md_lh, LNET_COOKIE_TYPE_MD);
+ list_add (&lmd->md_list, &the_lnet.ln_active_mds);
- return PTL_OK;
+ 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_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
+LNetMDAttach(lnet_handle_me_t meh, lnet_md_t umd,
+ lnet_unlink_t unlink, lnet_handle_md_t *handle)
{
- /*
- * 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_IOV)) != 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);
+ lnet_me_t *me;
+ lnet_libmd_t *md;
+ int rc;
+
+ LASSERT (the_lnet.ln_init);
+ LASSERT (the_lnet.ln_refcount > 0);
+
+ if ((umd.options & (LNET_MD_KIOV | LNET_MD_IOVEC)) != 0 &&
+ umd.length > LNET_MAX_IOV) /* too many fragments */
+ return -EINVAL;
+
+ md = lnet_md_alloc(&umd);
if (md == NULL)
- return (ret->rc = PTL_NOSPACE);
+ return -ENOMEM;
- state_lock(nal, &flags);
+ LNET_LOCK();
- me = ptl_handle2me(&args->me_in, nal);
+ me = lnet_handle2me(&meh);
if (me == NULL) {
- ret->rc = PTL_INV_ME;
- } else if (me->md != NULL) {
- ret->rc = PTL_INUSE;
+ rc = -ENOENT;
+ } else if (me->me_md != NULL) {
+ rc = -EBUSY;
} else {
- ret->rc = lib_md_build(nal, md, private, &args->md_in,
- &args->eq_in, args->unlink_in);
+ rc = lib_md_build(md, &umd, unlink);
+ if (rc == 0) {
+ me->me_md = md;
+ md->md_me = me;
- if (ret->rc == PTL_OK) {
- me->md = md;
- md->me = me;
+ lnet_md2handle(handle, md);
- ptl_md2handle(&ret->handle_out, md);
+ /* check if this MD matches any blocked msgs */
+ lnet_match_blocked_msg(md); /* expects LNET_LOCK held */
- state_unlock (nal, &flags);
- return (PTL_OK);
+ LNET_UNLOCK();
+ return (0);
}
}
- lib_md_free (nal, md);
+ lnet_md_free (md);
- state_unlock (nal, &flags);
- return (ret->rc);
+ LNET_UNLOCK();
+ return (rc);
}
-int do_PtlMDBind(nal_cb_t * nal, void *private, void *v_args, void *v_ret)
+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
- */
-
- PtlMDBind_in *args = v_args;
- PtlMDBind_out *ret = v_ret;
- lib_md_t *md;
- unsigned long flags;
+ lnet_libmd_t *md;
+ int rc;
- if ((args->md_in.options & (PTL_MD_KIOV | PTL_MD_IOV)) != 0 &&
- args->md_in.niov > PTL_MD_MAX_IOV) /* too many fragments */
- return (ret->rc = PTL_IOV_TOO_MANY);
+ LASSERT (the_lnet.ln_init);
+ LASSERT (the_lnet.ln_refcount > 0);
+
+ if ((umd.options & (LNET_MD_KIOV | LNET_MD_IOVEC)) != 0 &&
+ umd.length > LNET_MAX_IOV) /* too many fragments */
+ return -EINVAL;
- md = lib_md_alloc(nal, &args->md_in);
+ md = lnet_md_alloc(&umd);
if (md == NULL)
- return (ret->rc = PTL_NOSPACE);
+ return -ENOMEM;
- state_lock(nal, &flags);
+ LNET_LOCK();
- ret->rc = lib_md_build(nal, md, private,
- &args->md_in, &args->eq_in, PTL_UNLINK);
+ rc = lib_md_build(md, &umd, unlink);
- if (ret->rc == PTL_OK) {
- ptl_md2handle(&ret->handle_out, md);
+ if (rc == 0) {
+ lnet_md2handle(handle, md);
- state_unlock(nal, &flags);
- return (PTL_OK);
+ LNET_UNLOCK();
+ return (0);
}
- lib_md_free (nal, md);
+ lnet_md_free (md);
- state_unlock(nal, &flags);
- return (ret->rc);
+ LNET_UNLOCK();
+ return (rc);
}
-int do_PtlMDUnlink(nal_cb_t * nal, void *private, void *v_args, void *v_ret)
+int
+LNetMDUnlink (lnet_handle_md_t mdh)
{
- PtlMDUnlink_in *args = v_args;
- PtlMDUnlink_out *ret = v_ret;
- ptl_event_t ev;
- lib_md_t *md;
- unsigned long flags;
+ lnet_event_t ev;
+ lnet_libmd_t *md;
+
+ LASSERT (the_lnet.ln_init);
+ LASSERT (the_lnet.ln_refcount > 0);
- state_lock(nal, &flags);
+ LNET_LOCK();
- md = ptl_handle2md(&args->md_in, nal);
+ md = lnet_handle2md(&mdh);
if (md == NULL) {
- state_unlock(nal, &flags);
- return (ret->rc = PTL_INV_MD);
+ LNET_UNLOCK();
+ return -ENOENT;
}
- /* If the MD is busy, lib_md_unlink just marks it for deletion, and
+ /* 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... */
- if (md->eq != NULL &&
- md->pending == 0) {
- memset(&ev, 0, sizeof(ev));
-
- ev.type = PTL_EVENT_UNLINK;
- ev.status = PTL_OK;
- ev.unlinked = 1;
- lib_md_deconstruct(nal, md, &ev.mem_desc);
-
- lib_enq_event_locked(nal, private, md->eq, &ev);
+ if (md->md_eq != NULL &&
+ md->md_refcount == 0) {
+ lnet_build_unlink_event(md, &ev);
+ lnet_enq_event_locked(md->md_eq, &ev);
}
- lib_md_deconstruct(nal, md, &ret->status_out);
- lib_md_unlink(nal, md);
- ret->rc = PTL_OK;
-
- state_unlock(nal, &flags);
+ lnet_md_unlink(md);
- return (PTL_OK);
+ LNET_UNLOCK();
+ return 0;
}
-int do_PtlMDUpdate_internal(nal_cb_t * nal, void *private, void *v_args,
- void *v_ret)
-{
- /*
- * 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;
- }
-
- 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_IOV | PTL_MD_KIOV)) != 0) {
- ret->rc = PTL_INV_MD;
- 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 (!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;
- }
- }
-
- 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;
-
- // #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;
- }
-
- out:
- state_unlock(nal, &flags);
- return (ret->rc);
-}