diff mbox

Tidy up locking for libgomp OpenACC entry points

Message ID 20150424181308.099d8a26@octopus
State New
Headers show

Commit Message

Julian Brown April 24, 2015, 5:13 p.m. UTC
On Thu, 23 Apr 2015 18:41:34 +0200
Thomas Schwinge <thomas@codesourcery.com> wrote:

> Hi!
> 
> On Wed, 22 Apr 2015 19:42:43 +0100, Julian Brown
> <julian@codesourcery.com> wrote:
> > This patch is an attempt to fix some potential race conditions with
> > accesses to shared data structures from multiple concurrent threads
> > in libgomp's OpenACC entry points. The main change is to move
> > locking out of lookup_host and lookup_dev in oacc-mem.c and into
> > their callers (which can then hold the locks for the whole
> > operation that they are performing).
> 
> Yeah, that makes sense I guess.
> 
> > Also missing locking has been added for gomp_acc_insert_pointer.
> > 
> > Tests look OK (with offloading to NVidia PTX).
> 
> How did you test to get some confidence in the locking being
> sufficient?

Merely by running the existing tests and via inspection, sadly. I'm not
sure how much value we'd get from implementing an exhaustive threading
testsuite at this stage: I guess testcases will be easier to come by in
the future if/when people start to use e.g. OpenMP and OpenACC together.

> Going further (separate patch?), a few more comments:
> 
> Is it OK that oacc-init.c:cached_base_dev is accessed without locking?
> 
> Generally, we have to keep in mind that the same device may be
> accessed in parallel through both OpenACC and OpenMP interfaces.  For
> this, for example, in oacc-init.c, even though acc_device_lock is
> held, is it OK to call gomp_init_device(D) without D->lock being
> locked?  (Compare to target.c code.)
> 
> Please document what exactly oacc-init.c:acc_device_lock is to guard.
> I'm not sure I'm understanding this correctly.

I've attached a follow-on patch that documents the purpose of
acc_device_lock -- and also fixes some places that should have been
holding the lock, but were not.

I've also added locking (with dev->lock) when calling gomp_init_device
and gomp_fini_device from the OpenACC initialisation/finalisation code.

> Should oacc-init.c:acc_shutdown_1 release goacc_thread_lock before any
> gomp_fatal calls?  (That seems to be the general policy in libgomp.)

I added this to the first patch.

> > --- a/libgomp/oacc-mem.c
> > +++ b/libgomp/oacc-mem.c
> 
> > @@ -120,25 +116,32 @@ acc_free (void *d)
> >  {
> >    splay_tree_key k;
> >    struct goacc_thread *thr = goacc_thread ();
> > +  struct gomp_device_descr *acc_dev = thr->dev;
> >  
> >    if (!d)
> >      return;
> >  
> >    assert (thr && thr->dev);
> >  
> > +  gomp_mutex_lock (&acc_dev->lock);
> > +
> >    /* We don't have to call lazy open here, as the ptr value must
> > have been returned by acc_malloc.  It's not permitted to pass NULL
> > in (unless you got that null from acc_malloc).  */
> > -  if ((k = lookup_dev (thr->dev->openacc.data_environ, d, 1)))
> > -   {
> > -     void *offset;
> > +  if ((k = lookup_dev (acc_dev->openacc.data_environ, d, 1)))
> > +    {
> > +      void *offset;
> > +
> > +      offset = d - k->tgt->tgt_start + k->tgt_offset;
> >  
> > -     offset = d - k->tgt->tgt_start + k->tgt_offset;
> > +      gomp_mutex_unlock (&acc_dev->lock);
> >  
> > -     acc_unmap_data ((void *)(k->host_start + offset));
> > -   }
> > +      acc_unmap_data ((void *)(k->host_start + offset));
> > +    }
> > +  else
> > +    gomp_mutex_unlock (&acc_dev->lock);
> 
> Does it make sense to make the unlock unconditional, and move the
> acc_unmap_data after it, guarded by »if (k)«?

I've left this one -- just a stylistic tweak, but I think it's fine
as-is.

> > -  thr->dev->free_func (thr->dev->target_id, d);
> > +  acc_dev->free_func (acc_dev->target_id, d);
> >  }
> >  
> >  void
> > @@ -178,16 +181,24 @@ acc_deviceptr (void *h)
> >    goacc_lazy_initialize ();
> >  
> >    struct goacc_thread *thr = goacc_thread ();
> > +  struct gomp_device_descr *dev = thr->dev;
> > +
> > +  gomp_mutex_lock (&dev->lock);
> >  
> > -  n = lookup_host (thr->dev, h, 1);
> > +  n = lookup_host (dev, h, 1);
> >  
> >    if (!n)
> > -    return NULL;
> > +    {
> > +      gomp_mutex_unlock (&dev->lock);
> > +      return NULL;
> > +    }
> >  
> >    offset = h - n->host_start;
> >  
> >    d = n->tgt->tgt_start + n->tgt_offset + offset;
> >  
> > +  gomp_mutex_unlock (&dev->lock);
> > +
> >    return d;
> >  }
> 
> Do we need to retain the lock while working with n?  If not, the
> unlock could be placed right after the lookup_host, unconditionally.
> I'm confused -- it's commonly being done (retained) in target.c code,
> but not in the tgt_fn lookup in target.c:GOMP_target.

I think the difference can be explained as follows: a given mapping
(splay_key_tree_s) is essentially immutable after it is created (apart
from the refcounts). Thus it can be safely accessed *so long as we know
it will not be deallocated*.

Now, in some parts of target.c, we have an "active" target_mem_desc,
corresponding to a set of host-target mappings. So long as we are
holding that target_mem_desc (e.g. as we are in GOMP_target_data), we
know that none of the associated mappings' refcounts will fall to zero:
so, we can access them (read only) safely without explicitly holding the
lock.

But, that's *not* the case for e.g. acc_deviceptr: that can be called
at any point, in particular partway through deallocation of a set of
mappings. So we need the lock for that.

I didn't write the code of course, so E&OE :-).

> Should external functions (dev2host_func) be called without any locks
> held (to avoid issues with any call-backs); so move the unlocking
> before that?

I fixed this in the first patch.

Re-tested (libgomp), results look OK.

OK to apply now?

Julian

ChangeLog

    libgomp/
    * oacc-init.c (acc_shutdown_1): Call gomp_mutex_unlock for
    goacc_thread_lock on error paths.
    * oacc-mem.c (lookup_host): Remove locking from function. Note
    locking requirement for caller in function comment.
    (lookup_dev): Likewise.
    (acc_free, acc_deviceptr, acc_hostptr, acc_is_present)
    (acc_map_data, acc_unmap_data, present_create_copy, delete_copyout)
    (update_dev_host, gomp_acc_insert_pointer, gomp_acc_remove_pointer):
    Add locking.

ChangeLog

    libgomp/
    * oacc-init.c (acc_device_lock): Add explanatory comment.
    (resolve_device): Add comment about locking requirement.
    (acc_init_1, acc_shutdown_1): Likewise. Add locking around
    gomp_init_device and gomp_fini_device calls.
    (acc_get_num_devices, acc_set_device_type, acc_get_device_type)
    (acc_get_device_num, acc_set_device_num): Add locking around
    resolve_device and gomp_init_device calls.

Comments

Thomas Schwinge April 29, 2015, 5:18 p.m. UTC | #1
Hi Julian!

On Fri, 24 Apr 2015 18:13:08 +0100, Julian Brown <julian@codesourcery.com> wrote:
> On Thu, 23 Apr 2015 18:41:34 +0200
> Thomas Schwinge <thomas@codesourcery.com> wrote:
> > On Wed, 22 Apr 2015 19:42:43 +0100, Julian Brown
> > <julian@codesourcery.com> wrote:
> > > This patch is an attempt to fix some potential race conditions with
> > > accesses to shared data structures from multiple concurrent threads
> > > in libgomp's OpenACC entry points.

> > > Tests look OK (with offloading to NVidia PTX).
> > 
> > How did you test to get some confidence in the locking being
> > sufficient?
> 
> Merely by running the existing tests and via inspection, sadly. I'm not
> sure how much value we'd get from implementing an exhaustive threading
> testsuite at this stage: I guess testcases will be easier to come by in
> the future if/when people start to use e.g. OpenMP and OpenACC together.

;-) The poor person to be the first to actually exercise all that code...

Thanks for your explanations and the rework.

> Re-tested (libgomp), results look OK.
> 
> OK to apply now?

OK, thanks!


Grüße,
 Thomas
diff mbox

Patch

commit 8ae37f2c789e3a61fe4c62101fc13ea56cc1ee9f
Author: Julian Brown <julian@codesourcery.com>
Date:   Tue Apr 21 12:42:17 2015 -0700

    More locking in oacc-mem.c

diff --git a/libgomp/oacc-init.c b/libgomp/oacc-init.c
index 03db8ee..9d47a23 100644
--- a/libgomp/oacc-init.c
+++ b/libgomp/oacc-init.c
@@ -218,11 +218,17 @@  acc_shutdown_1 (acc_device_t d)
       /* This would mean the user is shutting down OpenACC in the middle of an
          "acc data" pragma.  Likely not intentional.  */
       if (walk->mapped_data)
-	gomp_fatal ("shutdown in 'acc data' region");
+	{
+	  gomp_mutex_unlock (&goacc_thread_lock);
+	  gomp_fatal ("shutdown in 'acc data' region");
+	}
 
       /* Similarly, if this happens then user code has done something weird.  */
       if (walk->saved_bound_dev)
-        gomp_fatal ("shutdown during host fallback");
+	{
+	  gomp_mutex_unlock (&goacc_thread_lock);
+	  gomp_fatal ("shutdown during host fallback");
+	}
 
       if (walk->dev)
 	{
diff --git a/libgomp/oacc-mem.c b/libgomp/oacc-mem.c
index 89ef5fc..90d43eb 100644
--- a/libgomp/oacc-mem.c
+++ b/libgomp/oacc-mem.c
@@ -35,7 +35,8 @@ 
 #include <stdint.h>
 #include <assert.h>
 
-/* Return block containing [H->S), or NULL if not contained.  */
+/* Return block containing [H->S), or NULL if not contained.  The device lock
+   for DEV must be locked on entry, and remains locked on exit.  */
 
 static splay_tree_key
 lookup_host (struct gomp_device_descr *dev, void *h, size_t s)
@@ -46,9 +47,7 @@  lookup_host (struct gomp_device_descr *dev, void *h, size_t s)
   node.host_start = (uintptr_t) h;
   node.host_end = (uintptr_t) h + s;
 
-  gomp_mutex_lock (&dev->lock);
   key = splay_tree_lookup (&dev->mem_map, &node);
-  gomp_mutex_unlock (&dev->lock);
 
   return key;
 }
@@ -56,7 +55,8 @@  lookup_host (struct gomp_device_descr *dev, void *h, size_t s)
 /* Return block containing [D->S), or NULL if not contained.
    The list isn't ordered by device address, so we have to iterate
    over the whole array.  This is not expected to be a common
-   operation.  */
+   operation.  The device lock associated with TGT must be locked on entry, and
+   remains locked on exit.  */
 
 static splay_tree_key
 lookup_dev (struct target_mem_desc *tgt, void *d, size_t s)
@@ -67,16 +67,12 @@  lookup_dev (struct target_mem_desc *tgt, void *d, size_t s)
   if (!tgt)
     return NULL;
 
-  gomp_mutex_lock (&tgt->device_descr->lock);
-
   for (t = tgt; t != NULL; t = t->prev)
     {
       if (t->tgt_start <= (uintptr_t) d && t->tgt_end >= (uintptr_t) d + s)
         break;
     }
 
-  gomp_mutex_unlock (&tgt->device_descr->lock);
-
   if (!t)
     return NULL;
 
@@ -119,26 +115,35 @@  void
 acc_free (void *d)
 {
   splay_tree_key k;
-  struct goacc_thread *thr = goacc_thread ();
 
   if (!d)
     return;
 
+  struct goacc_thread *thr = goacc_thread ();
+
   assert (thr && thr->dev);
 
+  struct gomp_device_descr *acc_dev = thr->dev;
+
+  gomp_mutex_lock (&acc_dev->lock);
+
   /* We don't have to call lazy open here, as the ptr value must have
      been returned by acc_malloc.  It's not permitted to pass NULL in
      (unless you got that null from acc_malloc).  */
-  if ((k = lookup_dev (thr->dev->openacc.data_environ, d, 1)))
-   {
-     void *offset;
+  if ((k = lookup_dev (acc_dev->openacc.data_environ, d, 1)))
+    {
+      void *offset;
 
-     offset = d - k->tgt->tgt_start + k->tgt_offset;
+      offset = d - k->tgt->tgt_start + k->tgt_offset;
 
-     acc_unmap_data ((void *)(k->host_start + offset));
-   }
+      gomp_mutex_unlock (&acc_dev->lock);
 
-  thr->dev->free_func (thr->dev->target_id, d);
+      acc_unmap_data ((void *)(k->host_start + offset));
+    }
+  else
+    gomp_mutex_unlock (&acc_dev->lock);
+
+  acc_dev->free_func (acc_dev->target_id, d);
 }
 
 void
@@ -178,16 +183,24 @@  acc_deviceptr (void *h)
   goacc_lazy_initialize ();
 
   struct goacc_thread *thr = goacc_thread ();
+  struct gomp_device_descr *dev = thr->dev;
+
+  gomp_mutex_lock (&dev->lock);
 
-  n = lookup_host (thr->dev, h, 1);
+  n = lookup_host (dev, h, 1);
 
   if (!n)
-    return NULL;
+    {
+      gomp_mutex_unlock (&dev->lock);
+      return NULL;
+    }
 
   offset = h - n->host_start;
 
   d = n->tgt->tgt_start + n->tgt_offset + offset;
 
+  gomp_mutex_unlock (&dev->lock);
+
   return d;
 }
 
@@ -204,16 +217,24 @@  acc_hostptr (void *d)
   goacc_lazy_initialize ();
 
   struct goacc_thread *thr = goacc_thread ();
+  struct gomp_device_descr *acc_dev = thr->dev;
 
-  n = lookup_dev (thr->dev->openacc.data_environ, d, 1);
+  gomp_mutex_lock (&acc_dev->lock);
+
+  n = lookup_dev (acc_dev->openacc.data_environ, d, 1);
 
   if (!n)
-    return NULL;
+    {
+      gomp_mutex_unlock (&acc_dev->lock);
+      return NULL;
+    }
 
   offset = d - n->tgt->tgt_start + n->tgt_offset;
 
   h = n->host_start + offset;
 
+  gomp_mutex_unlock (&acc_dev->lock);
+
   return h;
 }
 
@@ -232,6 +253,8 @@  acc_is_present (void *h, size_t s)
   struct goacc_thread *thr = goacc_thread ();
   struct gomp_device_descr *acc_dev = thr->dev;
 
+  gomp_mutex_lock (&acc_dev->lock);
+
   n = lookup_host (acc_dev, h, s);
 
   if (n && ((uintptr_t)h < n->host_start
@@ -239,6 +262,8 @@  acc_is_present (void *h, size_t s)
 	    || s > n->host_end - n->host_start))
     n = NULL;
 
+  gomp_mutex_unlock (&acc_dev->lock);
+
   return n != NULL;
 }
 
@@ -274,20 +299,32 @@  acc_map_data (void *h, void *d, size_t s)
 	gomp_fatal ("[%p,+%d]->[%p,+%d] is a bad map",
                     (void *)h, (int)s, (void *)d, (int)s);
 
+      gomp_mutex_lock (&acc_dev->lock);
+
       if (lookup_host (acc_dev, h, s))
-	gomp_fatal ("host address [%p, +%d] is already mapped", (void *)h,
-		    (int)s);
+        {
+	  gomp_mutex_unlock (&acc_dev->lock);
+	  gomp_fatal ("host address [%p, +%d] is already mapped", (void *)h,
+		      (int)s);
+	}
 
       if (lookup_dev (thr->dev->openacc.data_environ, d, s))
-	gomp_fatal ("device address [%p, +%d] is already mapped", (void *)d,
-		    (int)s);
+        {
+	  gomp_mutex_unlock (&acc_dev->lock);
+	  gomp_fatal ("device address [%p, +%d] is already mapped", (void *)d,
+		      (int)s);
+	}
+
+      gomp_mutex_unlock (&acc_dev->lock);
 
       tgt = gomp_map_vars (acc_dev, mapnum, &hostaddrs, &devaddrs, &sizes,
 			   &kinds, true, false);
     }
 
+  gomp_mutex_lock (&acc_dev->lock);
   tgt->prev = acc_dev->openacc.data_environ;
   acc_dev->openacc.data_environ = tgt;
+  gomp_mutex_unlock (&acc_dev->lock);
 }
 
 void
@@ -299,17 +336,26 @@  acc_unmap_data (void *h)
   /* No need to call lazy open, as the address must have been mapped.  */
 
   size_t host_size;
+
+  gomp_mutex_lock (&acc_dev->lock);
+
   splay_tree_key n = lookup_host (acc_dev, h, 1);
   struct target_mem_desc *t;
 
   if (!n)
-    gomp_fatal ("%p is not a mapped block", (void *)h);
+    {
+      gomp_mutex_unlock (&acc_dev->lock);
+      gomp_fatal ("%p is not a mapped block", (void *)h);
+    }
 
   host_size = n->host_end - n->host_start;
 
   if (n->host_start != (uintptr_t) h)
-    gomp_fatal ("[%p,%d] surrounds1 %p",
-		(void *) n->host_start, (int) host_size, (void *) h);
+    {
+      gomp_mutex_unlock (&acc_dev->lock);
+      gomp_fatal ("[%p,%d] surrounds %p",
+		  (void *) n->host_start, (int) host_size, (void *) h);
+    }
 
   t = n->tgt;
 
@@ -323,8 +369,6 @@  acc_unmap_data (void *h)
       t->tgt_end = 0;
       t->to_free = 0;
 
-      gomp_mutex_lock (&acc_dev->lock);
-
       for (tp = NULL, t = acc_dev->openacc.data_environ; t != NULL;
 	   tp = t, t = t->prev)
 	if (n->tgt == t)
@@ -336,10 +380,10 @@  acc_unmap_data (void *h)
 
 	    break;
 	  }
-
-      gomp_mutex_unlock (&acc_dev->lock);
     }
 
+  gomp_mutex_unlock (&acc_dev->lock);
+
   gomp_unmap_vars (t, true);
 }
 
@@ -361,6 +405,8 @@  present_create_copy (unsigned f, void *h, size_t s)
   struct goacc_thread *thr = goacc_thread ();
   struct gomp_device_descr *acc_dev = thr->dev;
 
+  gomp_mutex_lock (&acc_dev->lock);
+
   n = lookup_host (acc_dev, h, s);
   if (n)
     {
@@ -368,13 +414,22 @@  present_create_copy (unsigned f, void *h, size_t s)
       d = (void *) (n->tgt->tgt_start + n->tgt_offset);
 
       if (!(f & FLAG_PRESENT))
-        gomp_fatal ("[%p,+%d] already mapped to [%p,+%d]",
-            (void *)h, (int)s, (void *)d, (int)s);
+        {
+	  gomp_mutex_unlock (&acc_dev->lock);
+          gomp_fatal ("[%p,+%d] already mapped to [%p,+%d]",
+        	      (void *)h, (int)s, (void *)d, (int)s);
+	}
       if ((h + s) > (void *)n->host_end)
-        gomp_fatal ("[%p,+%d] not mapped", (void *)h, (int)s);
+	{
+	  gomp_mutex_unlock (&acc_dev->lock);
+	  gomp_fatal ("[%p,+%d] not mapped", (void *)h, (int)s);
+	}
+
+      gomp_mutex_unlock (&acc_dev->lock);
     }
   else if (!(f & FLAG_CREATE))
     {
+      gomp_mutex_unlock (&acc_dev->lock);
       gomp_fatal ("[%p,+%d] not mapped", (void *)h, (int)s);
     }
   else
@@ -389,6 +444,8 @@  present_create_copy (unsigned f, void *h, size_t s)
       else
 	kinds = GOMP_MAP_ALLOC;
 
+      gomp_mutex_unlock (&acc_dev->lock);
+
       tgt = gomp_map_vars (acc_dev, mapnum, &hostaddrs, NULL, &s, &kinds, true,
 			   false);
 
@@ -439,21 +496,31 @@  delete_copyout (unsigned f, void *h, size_t s)
   struct goacc_thread *thr = goacc_thread ();
   struct gomp_device_descr *acc_dev = thr->dev;
 
+  gomp_mutex_lock (&acc_dev->lock);
+
   n = lookup_host (acc_dev, h, s);
 
   /* No need to call lazy open, as the data must already have been
      mapped.  */
 
   if (!n)
-    gomp_fatal ("[%p,%d] is not mapped", (void *)h, (int)s);
+    {
+      gomp_mutex_unlock (&acc_dev->lock);
+      gomp_fatal ("[%p,%d] is not mapped", (void *)h, (int)s);
+    }
 
   d = (void *) (n->tgt->tgt_start + n->tgt_offset);
 
   host_size = n->host_end - n->host_start;
 
   if (n->host_start != (uintptr_t) h || host_size != s)
-    gomp_fatal ("[%p,%d] surrounds2 [%p,+%d]",
-		(void *) n->host_start, (int) host_size, (void *) h, (int) s);
+    {
+      gomp_mutex_unlock (&acc_dev->lock);
+      gomp_fatal ("[%p,%d] surrounds2 [%p,+%d]",
+		  (void *) n->host_start, (int) host_size, (void *) h, (int) s);
+    }
+
+  gomp_mutex_unlock (&acc_dev->lock);
 
   if (f & FLAG_COPYOUT)
     acc_dev->dev2host_func (acc_dev->target_id, h, d, s);
@@ -482,16 +549,23 @@  update_dev_host (int is_dev, void *h, size_t s)
   struct goacc_thread *thr = goacc_thread ();
   struct gomp_device_descr *acc_dev = thr->dev;
 
+  gomp_mutex_lock (&acc_dev->lock);
+
   n = lookup_host (acc_dev, h, s);
 
   /* No need to call lazy open, as the data must already have been
      mapped.  */
 
   if (!n)
-    gomp_fatal ("[%p,%d] is not mapped", h, (int)s);
+    {
+      gomp_mutex_unlock (&acc_dev->lock);
+      gomp_fatal ("[%p,%d] is not mapped", h, (int)s);
+    }
 
   d = (void *) (n->tgt->tgt_start + n->tgt_offset);
 
+  gomp_mutex_unlock (&acc_dev->lock);
+
   if (is_dev)
     acc_dev->host2dev_func (acc_dev->target_id, d, h, s);
   else
@@ -522,8 +596,11 @@  gomp_acc_insert_pointer (size_t mapnum, void **hostaddrs, size_t *sizes,
   tgt = gomp_map_vars (acc_dev, mapnum, hostaddrs,
 		       NULL, sizes, kinds, true, false);
   gomp_debug (0, "  %s: mappings prepared\n", __FUNCTION__);
+
+  gomp_mutex_lock (&acc_dev->lock);
   tgt->prev = acc_dev->openacc.data_environ;
   acc_dev->openacc.data_environ = tgt;
+  gomp_mutex_unlock (&acc_dev->lock);
 }
 
 void
@@ -535,10 +612,15 @@  gomp_acc_remove_pointer (void *h, bool force_copyfrom, int async, int mapnum)
   struct target_mem_desc *t;
   int minrefs = (mapnum == 1) ? 2 : 3;
 
+  gomp_mutex_lock (&acc_dev->lock);
+
   n = lookup_host (acc_dev, h, 1);
 
   if (!n)
-    gomp_fatal ("%p is not a mapped block", (void *)h);
+    {
+      gomp_mutex_unlock (&acc_dev->lock);
+      gomp_fatal ("%p is not a mapped block", (void *)h);
+    }
 
   gomp_debug (0, "  %s: restore mappings\n", __FUNCTION__);
 
@@ -546,8 +628,6 @@  gomp_acc_remove_pointer (void *h, bool force_copyfrom, int async, int mapnum)
 
   struct target_mem_desc *tp;
 
-  gomp_mutex_lock (&acc_dev->lock);
-
   if (t->refcount == minrefs)
     {
       /* This is the last reference, so pull the descriptor off the