OpenACC reference count consistency checking

Message ID 1543578624-1511-1-git-send-email-julian@codesourcery.com
State New
Headers show
Series
  • OpenACC reference count consistency checking
Related show

Commit Message

Julian Brown Nov. 30, 2018, 11:50 a.m.
This is a trunk-compatible version of the patch posted here:

  https://gcc.gnu.org/ml/gcc-patches/2018-11/msg02365.html

I understand it may not be suitable for committing (especially not
outside stage 1 -- though it's "obviously harmless" in its dormant state),
but it might be helpful for review purposes for the main attach/detach
patch, i.e.:

  https://gcc.gnu.org/ml/gcc-patches/2018-11/msg02556.html

For convenience, I will copy the blurb from the og8 submission of the
patch here.

[...] The model used for checking is as follows.

 1. Each splay tree key that references a target memory descriptor
    increases that descriptor's refcount by 1.

 2. Each variable listed in a target memory descriptor that links back to a
    splay tree key increases that key's refcount by 1. Each target memory
    descriptor's variable list is counted only once, even if multiple
    splay tree keys point to it (via their "tgt" field).

 3. Additional ("real") target memory descriptors may be present
    representing data mapped through "acc data" or "acc parallel/kernels"
    blocks.  These descriptors have their refcount bumped, and the
    variables linked through such blocks have their refcounts bumped also
    (again, with "once only" semantics).

 4. Asynchronous operations "artificially" bump the reference counts for
    referenced target memory descriptors (but *not* for linked
    variables/splay tree keys), in order to delay freeing mapped device
    memory until the asynchronous operation has completed.  We model this,
    for checking purposes only, using an off-side linked list.

 5. "Virtual" reference counts ("virtual_refcount") cannot be checked
    purely statically, so we add the incoming value to each key's
    statically-determined reference count ("refcount_chk"), and make
    sure that the total matches the incoming reference count ("refcount").

Thanks,

Julian

ChangeLog

	libgomp/
	* libgomp.h (RC_CHECKING): New macro, disabled by default, guarding all
	hunks in this patch.
	(target_mem_desc): Add forward declaration.
	(async_tgt_use): New struct.
	(target_mem_desc): Add refcount_chk, mark fields.
	(acc_dispatch_t): Add tgt_uses, au_lock fields.
	(dump_tgt, gomp_rc_check): Add prototypes.
	* oacc-async (goacc_async_unmap_tgt): Add refcount self-check code.
	(goacc_async_copyout_unmap_vars): Likewise.
	(goacc_remove_var_async): Likewise.
	* oacc-parallel.c (GOACC_parallel_keyed_internal): Add refcount
	self-check code.
	(GOACC_data_start, GOACC_data_end, GOACC_enter_exit_data): Likewise.
	* target.c (stdio.h): Include.
	(dump_tgt, rc_check_clear, rc_check_count, rc_check_verify)
	(gomp_rc_check): New functions to consistency-check reference counts.
	(gomp_target_init): Initialise self-check-related device fields.
---
 libgomp/libgomp.h       |   31 +++++++
 libgomp/oacc-async.c    |   46 +++++++++++
 libgomp/oacc-parallel.c |   33 ++++++++
 libgomp/target.c        |  199 +++++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 309 insertions(+), 0 deletions(-)

Comments

Thomas Schwinge Jan. 30, 2020, 3:21 p.m. | #1
Hi Julian!

Notwithstanding the open question about how to implement this checking in
libgomp in a non-intrusive (performance-wise) yet maintainable (avoid
'#if 0') way, I have two more questions.


Is there a specific reason why this checking isn't also enabled for
libgomp OpenMP 'target' entry points?


Can you please explain (textually?) how this checking (design per your
textual description below) is working in context of mixed OpenACC
structured ("S") and dynamic ("D") reference counts?  For example:

    // S: 0, D: 0
    
    #pragma acc enter data copyin ([data]) // copyin; S: 0, D: 1
    
    acc_copyin ([data]) // no-op; S: 0, D: 2
    
    #pragma acc data copyout ([data]) // no-op; S: 1, D: 2
      {
        acc_create ([data]) // no-op; S: 1, D: 3
        
        #pragma acc data create ([data]) // no-op; S: 2, D: 3
          {
            #pragma acc parallel copyout ([data]) // no-op; S: 3, D: 3
              {
              } // no-op; S: 2, D: 3
    
            acc_delete_finalize ([data]) // no-op; S: 2, D: 0
    
            acc_create ([data]) // no-op; S: 2, D: 1
          } // no-op; S: 1, D: 1
    
        #pragma acc exit data delete ([data]) // no-op; S: 1, D: 0
      } // copyout; S: 0, D: 0
    
    assert (!acc_is_present ([data]));

(Haven't compiled but I'm reasonably sure that the nesting and my manual
"[action]; [S], [D]" annotations are correct.  But please verify, if
course.)


Grüße
 Thomas


On 2018-11-30T03:50:24-0800, Julian Brown <julian@codesourcery.com> wrote:
> The model used for checking is as follows.

>

>  1. Each splay tree key that references a target memory descriptor

>     increases that descriptor's refcount by 1.

>

>  2. Each variable listed in a target memory descriptor that links back to a

>     splay tree key increases that key's refcount by 1. Each target memory

>     descriptor's variable list is counted only once, even if multiple

>     splay tree keys point to it (via their "tgt" field).

>

>  3. Additional ("real") target memory descriptors may be present

>     representing data mapped through "acc data" or "acc parallel/kernels"

>     blocks.  These descriptors have their refcount bumped, and the

>     variables linked through such blocks have their refcounts bumped also

>     (again, with "once only" semantics).

>

>  4. Asynchronous operations "artificially" bump the reference counts for

>     referenced target memory descriptors (but *not* for linked

>     variables/splay tree keys), in order to delay freeing mapped device

>     memory until the asynchronous operation has completed.  We model this,

>     for checking purposes only, using an off-side linked list.

>

>  5. "Virtual" reference counts ("virtual_refcount") cannot be checked

>     purely statically, so we add the incoming value to each key's

>     statically-determined reference count ("refcount_chk"), and make

>     sure that the total matches the incoming reference count ("refcount").

>

> Thanks,

>

> Julian

>

> ChangeLog

>

> 	libgomp/

> 	* libgomp.h (RC_CHECKING): New macro, disabled by default, guarding all

> 	hunks in this patch.

> 	(target_mem_desc): Add forward declaration.

> 	(async_tgt_use): New struct.

> 	(target_mem_desc): Add refcount_chk, mark fields.

> 	(acc_dispatch_t): Add tgt_uses, au_lock fields.

> 	(dump_tgt, gomp_rc_check): Add prototypes.

> 	* oacc-async (goacc_async_unmap_tgt): Add refcount self-check code.

> 	(goacc_async_copyout_unmap_vars): Likewise.

> 	(goacc_remove_var_async): Likewise.

> 	* oacc-parallel.c (GOACC_parallel_keyed_internal): Add refcount

> 	self-check code.

> 	(GOACC_data_start, GOACC_data_end, GOACC_enter_exit_data): Likewise.

> 	* target.c (stdio.h): Include.

> 	(dump_tgt, rc_check_clear, rc_check_count, rc_check_verify)

> 	(gomp_rc_check): New functions to consistency-check reference counts.

> 	(gomp_target_init): Initialise self-check-related device fields.

> ---

>  libgomp/libgomp.h       |   31 +++++++

>  libgomp/oacc-async.c    |   46 +++++++++++

>  libgomp/oacc-parallel.c |   33 ++++++++

>  libgomp/target.c        |  199 +++++++++++++++++++++++++++++++++++++++++++++++

>  4 files changed, 309 insertions(+), 0 deletions(-)

>

> diff --git a/libgomp/libgomp.h b/libgomp/libgomp.h

> index df49c1b..24cbddd 100644

> --- a/libgomp/libgomp.h

> +++ b/libgomp/libgomp.h

> @@ -874,9 +874,26 @@ struct target_var_desc {

>    uintptr_t length;

>  };

>  

> +/* Uncomment to enable reference-count consistency checking (for development

> +   use only).  */

> +/*#define RC_CHECKING 1*/

> +

> +#ifdef RC_CHECKING

> +struct target_mem_desc;

> +

> +struct async_tgt_use {

> +  struct target_mem_desc *tgt;

> +  struct async_tgt_use *next;

> +};

> +#endif

> +

>  struct target_mem_desc {

>    /* Reference count.  */

>    uintptr_t refcount;

> +#ifdef RC_CHECKING

> +  uintptr_t refcount_chk;

> +  bool mark;

> +#endif

>    /* All the splay nodes allocated together.  */

>    splay_tree_node array;

>    /* Start of the target region.  */

> @@ -925,6 +942,10 @@ struct splay_tree_key_s {

>       "present increment" operations (via "acc enter data") refering to the same

>       host-memory block.  */

>    uintptr_t virtual_refcount;

> +#ifdef RC_CHECKING

> +  /* The recalculated reference count, for verification.  */

> +  uintptr_t refcount_chk;

> +#endif

>    /* For a block with attached pointers, the attachment counters for each.  */

>    unsigned short *attach_count;

>    /* Pointer to the original mapping of "omp declare target link" object.  */

> @@ -958,6 +979,10 @@ typedef struct acc_dispatch_t

>      int nasyncqueue;

>      struct goacc_asyncqueue **asyncqueue;

>      struct goacc_asyncqueue_list *active;

> +#ifdef RC_CHECKING

> +    struct async_tgt_use *tgt_uses;

> +    gomp_mutex_t au_lock;

> +#endif

>  

>      __typeof (GOMP_OFFLOAD_openacc_async_construct) *construct_func;

>      __typeof (GOMP_OFFLOAD_openacc_async_destruct) *destruct_func;

> @@ -1085,6 +1110,12 @@ extern void gomp_detach_pointer (struct gomp_device_descr *,

>  				 struct goacc_asyncqueue *, splay_tree_key,

>  				 uintptr_t, bool, struct gomp_coalesce_buf *);

>  

> +#ifdef RC_CHECKING

> +extern void dump_tgt (const char *, struct target_mem_desc *);

> +extern void gomp_rc_check (struct gomp_device_descr *,

> +			   struct target_mem_desc *);

> +#endif

> +

>  extern struct target_mem_desc *gomp_map_vars (struct gomp_device_descr *,

>  					      size_t, void **, void **,

>  					      size_t *, void *, bool,

> diff --git a/libgomp/oacc-async.c b/libgomp/oacc-async.c

> index 077e28f..8b0f228 100644

> --- a/libgomp/oacc-async.c

> +++ b/libgomp/oacc-async.c

> @@ -243,6 +243,29 @@ goacc_async_unmap_tgt (void *ptr)

>  {

>    struct target_mem_desc *tgt = (struct target_mem_desc *) ptr;

>  

> +#ifdef RC_CHECKING

> +  {

> +    struct gomp_device_descr *devicep = tgt->device_descr;

> +    struct async_tgt_use *aup, *au;

> +    gomp_mutex_lock (&devicep->openacc.async.au_lock);

> +    /* Remove tgt from asynchronous-use list.  */

> +    for (aup = NULL, au = devicep->openacc.async.tgt_uses; au;

> +	 aup = au, au = au->next)

> +      if (au->tgt == tgt)

> +	{

> +	  if (aup)

> +	    aup->next = au->next;

> +	  else

> +	    devicep->openacc.async.tgt_uses = au->next;

> +	  free (au);

> +	  break;

> +	}

> +    if (!au)

> +      gomp_fatal ("can't find tgt %p to remove in async list", tgt);

> +    gomp_mutex_unlock (&devicep->openacc.async.au_lock);

> +  }

> +#endif

> +

>    if (tgt->refcount > 1)

>      tgt->refcount--;

>    else

> @@ -258,6 +281,18 @@ goacc_async_copyout_unmap_vars (struct target_mem_desc *tgt,

>    /* Increment reference to delay freeing of device memory until callback

>       has triggered.  */

>    tgt->refcount++;

> +

> +#ifdef RC_CHECKING

> +  {

> +    struct async_tgt_use *au = malloc (sizeof (struct async_tgt_use));

> +    gomp_mutex_lock (&devicep->openacc.async.au_lock);

> +    /* Record the asynchronous use of this target_mem_desc.  */

> +    au->next = devicep->openacc.async.tgt_uses;

> +    au->tgt = tgt;

> +    devicep->openacc.async.tgt_uses = au;

> +    gomp_mutex_unlock (&devicep->openacc.async.au_lock);

> +  }

> +#endif

>    gomp_unmap_vars_async (tgt, true, aq);

>    devicep->openacc.async.queue_callback_func (aq, goacc_async_unmap_tgt,

>  					      (void *) tgt);

> @@ -276,6 +311,17 @@ goacc_remove_var_async (struct gomp_device_descr *devicep, splay_tree_key n,

>    struct target_mem_desc *tgt = n->tgt;

>    assert (tgt);

>    tgt->refcount++;

> +#ifdef RC_CHECKING

> +  {

> +    gomp_mutex_lock (&devicep->openacc.async.au_lock);

> +    struct async_tgt_use *au = malloc (sizeof (struct async_tgt_use));

> +    /* Record the asynchronous use of this target_mem_desc.  */

> +    au->next = devicep->openacc.async.tgt_uses;

> +    au->tgt = tgt;

> +    devicep->openacc.async.tgt_uses = au;

> +    gomp_mutex_unlock (&devicep->openacc.async.au_lock);

> +  }

> +#endif

>    gomp_remove_var (devicep, n);

>    devicep->openacc.async.queue_callback_func (aq, goacc_async_unmap_tgt,

>                                               (void *) tgt);

> diff --git a/libgomp/oacc-parallel.c b/libgomp/oacc-parallel.c

> index 0e9a3e8..22aafb2 100644

> --- a/libgomp/oacc-parallel.c

> +++ b/libgomp/oacc-parallel.c

> @@ -253,6 +253,15 @@ GOACC_parallel_keyed (int device, void (*fn) (void *),

>    tgt = gomp_map_vars_async (acc_dev, aq, mapnum, hostaddrs, NULL, sizes, kinds,

>  			     true, GOMP_MAP_VARS_OPENACC);

>  

> +#ifdef RC_CHECKING

> +  gomp_mutex_lock (&acc_dev->lock);

> +  assert (tgt);

> +  dump_tgt (__FUNCTION__, tgt);

> +  tgt->prev = thr->mapped_data;

> +  gomp_rc_check (acc_dev, tgt);

> +  gomp_mutex_unlock (&acc_dev->lock);

> +#endif

> +

>    devaddrs = gomp_alloca (sizeof (void *) * mapnum);

>    for (i = 0; i < mapnum; i++)

>      devaddrs[i] = (void *) gomp_map_val (tgt, hostaddrs, i);

> @@ -270,6 +279,12 @@ GOACC_parallel_keyed (int device, void (*fn) (void *),

>  					dims, tgt, aq);

>        goacc_async_copyout_unmap_vars (tgt, aq);

>      }

> +

> +#ifdef RC_CHECKING

> +  gomp_mutex_lock (&acc_dev->lock);

> +  gomp_rc_check (acc_dev, thr->mapped_data);

> +  gomp_mutex_unlock (&acc_dev->lock);

> +#endif

>  }

>  

>  /* Legacy entry point, only provide host execution.  */

> @@ -324,6 +339,12 @@ GOACC_data_start (int device, size_t mapnum,

>    gomp_debug (0, "  %s: mappings prepared\n", __FUNCTION__);

>    tgt->prev = thr->mapped_data;

>    thr->mapped_data = tgt;

> +

> +#ifdef RC_CHECKING

> +  gomp_mutex_lock (&acc_dev->lock);

> +  gomp_rc_check (acc_dev, thr->mapped_data);

> +  gomp_mutex_unlock (&acc_dev->lock);

> +#endif

>  }

>  

>  void

> @@ -336,6 +357,12 @@ GOACC_data_end (void)

>    thr->mapped_data = tgt->prev;

>    gomp_unmap_vars (tgt, true);

>    gomp_debug (0, "  %s: mappings restored\n", __FUNCTION__);

> +

> +#ifdef RC_CHECKING

> +  gomp_mutex_lock (&thr->dev->lock);

> +  gomp_rc_check (thr->dev, thr->mapped_data);

> +  gomp_mutex_unlock (&thr->dev->lock);

> +#endif

>  }

>  

>  void

> @@ -624,6 +651,12 @@ GOACC_enter_exit_data (int device, size_t mapnum,

>  	    }

>  	}

>      }

> +

> +#ifdef RC_CHECKING

> +  gomp_mutex_lock (&acc_dev->lock);

> +  gomp_rc_check (acc_dev, thr->mapped_data);

> +  gomp_mutex_unlock (&acc_dev->lock);

> +#endif

>  }

>  

>  static void

> diff --git a/libgomp/target.c b/libgomp/target.c

> index 6e115d1..b9f8ce8 100644

> --- a/libgomp/target.c

> +++ b/libgomp/target.c

> @@ -40,6 +40,9 @@

>  #include <assert.h>

>  #include <errno.h>

>  #include <limits.h>

> +#ifdef RC_CHECKING

> +#include <stdio.h>

> +#endif

>  

>  #ifdef PLUGIN_SUPPORT

>  #include <dlfcn.h>

> @@ -360,6 +363,198 @@ gomp_free_device_memory (struct gomp_device_descr *devicep, void *devptr)

>      }

>  }

>  

> +#ifdef RC_CHECKING

> +void

> +dump_tgt (const char *where, struct target_mem_desc *tgt)

> +{

> +  if (!getenv ("GOMP_DEBUG_TGT"))

> +    return;

> +

> +  fprintf (stderr, "%s: %s: tgt=%p\n", __FUNCTION__, where, (void*) tgt);

> +  fprintf (stderr, "refcount=%d\n", (int) tgt->refcount);

> +  fprintf (stderr, "tgt_start=%p\n", (void*) tgt->tgt_start);

> +  fprintf (stderr, "tgt_end=%p\n", (void*) tgt->tgt_end);

> +  fprintf (stderr, "to_free=%p\n", tgt->to_free);

> +  fprintf (stderr, "list_count=%d\n", (int) tgt->list_count);

> +  for (int i = 0; i < tgt->list_count; i++)

> +    {

> +      fprintf (stderr, "list item %d:\n", i);

> +      fprintf (stderr, "  key: %p\n", (void*) tgt->list[i].key);

> +      if (tgt->list[i].key)

> +	{

> +	  fprintf (stderr, "  key.host_start=%p\n",

> +		   (void*) tgt->list[i].key->host_start);

> +	  fprintf (stderr, "  key.host_end=%p\n",

> +		   (void*) tgt->list[i].key->host_end);

> +	  fprintf (stderr, "  key.tgt=%p\n", (void*) tgt->list[i].key->tgt);

> +	  fprintf (stderr, "  key.offset=%d\n",

> +		   (int) tgt->list[i].key->tgt_offset);

> +	  fprintf (stderr, "  key.refcount=%d\n",

> +		   (int) tgt->list[i].key->refcount);

> +	  fprintf (stderr, "  key.virtual_refcount=%d\n",

> +		   (int) tgt->list[i].key->virtual_refcount);

> +	  fprintf (stderr, "  key.attach_count=%p\n",

> +		   (void*) tgt->list[i].key->attach_count);

> +	  fprintf (stderr, "  key.link_key=%p\n",

> +		   (void*) tgt->list[i].key->link_key);

> +	}

> +    }

> +  fprintf (stderr, "\n");

> +}

> +

> +static void

> +rc_check_clear (splay_tree_node node)

> +{

> +  splay_tree_key k = &node->key;

> +

> +  k->refcount_chk = 0;

> +  k->tgt->refcount_chk = 0;

> +  k->tgt->mark = false;

> +

> +  if (node->left)

> +    rc_check_clear (node->left);

> +  if (node->right)

> +    rc_check_clear (node->right);

> +}

> +

> +static void

> +rc_check_count (splay_tree_node node)

> +{

> +  splay_tree_key k = &node->key;

> +  struct target_mem_desc *t;

> +

> +  /* Add virtual reference counts ("acc enter data", etc.) for this key.  */

> +  k->refcount_chk += k->virtual_refcount;

> +

> +  t = k->tgt;

> +  t->refcount_chk++;

> +

> +  if (!t->mark)

> +    {

> +      for (int i = 0; i < t->list_count; i++)

> +	if (t->list[i].key)

> +	  t->list[i].key->refcount_chk++;

> +

> +      t->mark = true;

> +    }

> +

> +  if (node->left)

> +    rc_check_count (node->left);

> +  if (node->right)

> +    rc_check_count (node->right);

> +}

> +

> +static bool

> +rc_check_verify (splay_tree_node node, bool noisy, bool errors)

> +{

> +  splay_tree_key k = &node->key;

> +  struct target_mem_desc *t;

> +

> +  if (k->refcount != REFCOUNT_INFINITY)

> +    {

> +      if (noisy)

> +	fprintf (stderr, "key %p (%p..+%d): rc=%d/%d, virt_rc=%d\n", k,

> +		 (void *) k->host_start, (int) (k->host_end - k->host_start),

> +		 (int) k->refcount, (int) k->refcount_chk,

> +		 (int) k->virtual_refcount);

> +

> +      if (k->refcount != k->refcount_chk)

> +	{

> +	  if (noisy)

> +	    fprintf (stderr, "  -- key refcount mismatch!\n");

> +	  errors = true;

> +	}

> +

> +      t = k->tgt;

> +

> +      if (noisy)

> +	fprintf (stderr, "tgt %p: rc=%d/%d\n", t, (int) t->refcount,

> +		 (int) t->refcount_chk);

> +

> +      if (t->refcount != t->refcount_chk)

> +	{

> +	  if (noisy)

> +	    fprintf (stderr,

> +		     "  -- target memory descriptor refcount mismatch!\n");

> +	  errors = true;

> +	}

> +    }

> +

> +  if (node->left)

> +    errors |= rc_check_verify (node->left, noisy, errors);

> +  if (node->right)

> +    errors |= rc_check_verify (node->right, noisy, errors);

> +

> +  return errors;

> +}

> +

> +/* Call with device locked.  */

> +

> +attribute_hidden void

> +gomp_rc_check (struct gomp_device_descr *devicep, struct target_mem_desc *tgt)

> +{

> +  splay_tree sp = &devicep->mem_map;

> +

> +  bool noisy = getenv ("GOMP_DEBUG_TGT") != 0;

> +

> +  if (noisy)

> +    fprintf (stderr, "\n*** GOMP_RC_CHECK ***\n\n");

> +

> +  if (sp->root)

> +    {

> +      gomp_mutex_lock (&devicep->openacc.async.au_lock);

> +      struct async_tgt_use *async_uses = devicep->openacc.async.tgt_uses;

> +

> +      rc_check_clear (sp->root);

> +

> +      for (struct target_mem_desc *t = tgt; t; t = t->prev)

> +	{

> +	  t->refcount_chk = 0;

> +	  t->mark = false;

> +	}

> +      for (struct async_tgt_use *au = async_uses; au; au = au->next)

> +	{

> +	  struct target_mem_desc *t = au->tgt;

> +	  t->refcount_chk = 0;

> +	  t->mark = false;

> +	}

> +

> +      /* Add references for interconnected splay-tree keys.  */

> +      rc_check_count (sp->root);

> +

> +      /* Add references for the tgt for a currently-executing kernel and/or

> +	 any enclosing data directives.  */

> +      for (struct target_mem_desc *t = tgt; t; t = t->prev)

> +	{

> +	  t->refcount_chk++;

> +

> +	  if (!t->mark)

> +	    {

> +	      for (int i = 0; i < t->list_count; i++)

> +		if (t->list[i].key)

> +		  t->list[i].key->refcount_chk++;

> +

> +	      t->mark = true;

> +	    }

> +	}

> +

> +      /* Add references from in-progress asynchronous operations.  */

> +      for (struct async_tgt_use *au = async_uses; au; au = au->next)

> +	{

> +	  struct target_mem_desc *t = au->tgt;

> +	  t->refcount_chk++;

> +	}

> +

> +      if (rc_check_verify (sp->root, noisy, false))

> +	{

> +	  gomp_mutex_unlock (&devicep->lock);

> +	  gomp_fatal ("refcount checking failure");

> +	}

> +      gomp_mutex_unlock (&devicep->openacc.async.au_lock);

> +    }

> +}

> +#endif

> +

>  /* Handle the case where gomp_map_lookup, splay_tree_lookup or

>     gomp_map_0len_lookup found oldn for newn.

>     Helper function of gomp_map_vars.  */

> @@ -3274,6 +3469,10 @@ gomp_target_init (void)

>  		current_device.type = current_device.get_type_func ();

>  		current_device.mem_map.root = NULL;

>  		current_device.state = GOMP_DEVICE_UNINITIALIZED;

> +#ifdef RC_CHECKING

> +		current_device.openacc.async.tgt_uses = NULL;

> +		gomp_mutex_init (&current_device.openacc.async.au_lock);

> +#endif

>  

>  		/* Augment DEVICES and NUM_DEVICES.  */

>  		devices = gomp_realloc (devices,
Julian Brown May 7, 2020, 4:11 p.m. | #2
Sorry about the delay replying to this email!

On Thu, 30 Jan 2020 16:21:20 +0100
Thomas Schwinge <thomas@codesourcery.com> wrote:

> Hi Julian!

> 

> Notwithstanding the open question about how to implement this

> checking in libgomp in a non-intrusive (performance-wise) yet

> maintainable (avoid '#if 0') way, I have two more questions.

> 

> 

> Is there a specific reason why this checking isn't also enabled for

> libgomp OpenMP 'target' entry points?


Just that it was developed in the context of adding manual deep-copy
support to OpenACC -- OpenMP wasn't my focus at that point. So, I
didn't try adding checking for OpenMP also. It might be interesting to
see how that goes though, particularly with regards to dynamic data
lifetimes in OpenMP.

> Can you please explain (textually?) how this checking (design per your

> textual description below) is working in context of mixed OpenACC

> structured ("S") and dynamic ("D") reference counts?  For example:

> 

>     // S: 0, D: 0

>     

>     #pragma acc enter data copyin ([data]) // copyin; S: 0, D: 1

>     

>     acc_copyin ([data]) // no-op; S: 0, D: 2


Unfortunately it's not quite that simple. The "refcount" fields (in
either splay tree keys or target_mem_descs) do not really represent
program-level reference counts, but rather references in the linked
splay tree structure within libgomp. That's correct: the refcounts are
used so as to know when data is still live, and when it can be freed.

Structured data mapping operations ("acc data", "acc parallel", etc.)
always create a target_mem_desc, with a list of target_var_descs that
describe data mapped in that structured block. That target_mem_desc
either "owns" a block of target memory corresponding to the structured
data block, or it doesn't.

We might have something like this (excuse ASCII art!):

   +===================+        +=================+
   | TARGET_MEM_DESC 1 |   ,-->	| TARGET_VAR_DESC |
   +-------------------+   |	+-----------------+
   | tgt_start...      |   |    | splay_tree_key  | --> ... 
   +-------------------+   |    +=================+
   | target_var_desc 0 | --' 
   | target_var_desc 1 | ---.  	+=================+
   | target_var_desc 2 | -. `-> | TARGET_VAR_DESC |
   +===================+  |     +-----------------+
			  |	| splay_tree_key  | --> ...  
			  |	+=================+  
			  |			     
			  |    	+=================+  
			  `--->	| TARGET_VAR_DESC |  
				+-----------------+  
   +=================+   .-----	| splay_tree_key  |
   | SPLAY_TREE_KEY  | <-'	+=================+
   +-----------------+	 
   | target_mem_desc | -.       +===================+
   +=================+	'-----> | TARGET_MEM_DESC 2 |
				+-------------------+
				| tgt_start...      |
				+-------------------+
				| target_var_desc   |
				+===================+

(Non-virtual/non-dynamic) reference counts correspond to the arrows
between blocks in the diagram (for the pointed-to block --
target_mem_desc or splay tree key).

For a structured data mapping, say "TARGET_MEM_DESC 1" is the descriptor
returned from gomp_map_vars.

Now, "TARGET_MEM_DESC 1" and "TARGET_MEM_DESC 2" can be the same block,
or different blocks. (Each of the TARGET_MEM_DESCs linked from splay
tree keys, linked from TARGET_VAR_DESCs, can be a mix of such
identical or different blocks for each of the splay tree keys linked
from TARGET_VAR_DESCs.) In the case where they're different blocks, and
TARGET_MEM_DESC 2 (etc.) owns its own mapped memory, TARGET_MEM_DESC 1
may have a NULL tgt_start -- thus, not own a target data block itself.

In the case of a dynamic mapping, this subtlety is especially
important. A target_mem_desc being returned from
gomp_map_vars{_internal} with a refcount of zero -- one which no splay
tree keys link back to, because it does not own its own block of target
memory -- is discarded before the function returns.

So, the first time a dynamic data mapping takes place for DATA, we have:

>     // S: 0, D: 0

>     

>     #pragma acc enter data copyin ([data]) // copyin; S: 1, D: 0


This is because the target_mem_desc created to describe on-target
memory for DATA will "own" that data: nothing has referred to it
beforehand. So there's a "real" link from the splay tree key for DATA's
host region to the target_mem_desc we just created. (Yes, the
splay tree key's reference counts look just like a structured data
mapping. That was a subject for another patch.)

>     acc_copyin ([data]) // no-op; S: 2, D: 1


So now we have another dynamic mapping. This time, we already have a
target_mem_desc describing DATA on the target. The
gomp_map_vars_internal function will return NULL -- but before it does
that, it realises that it will "lose" references in doing so. Those are
the ones linked via the discarded target_mem_desc's variable list to
splay tree keys that are referred to in the dynamic mapping operation.

For OpenACC, that's where the "virtual" refcount comes in -- to keep
track of those "lost" dynamic references. In particular, the "virtual"
refcount is the count by which the structured reference count must be
decremented when we hit an OpenACC "finalize" operation. Without that
(cf. OpenMP), we probably wouldn't need it.

>     #pragma acc data copyout ([data]) // no-op; S: 1, D: 2

>       {

>         acc_create ([data]) // no-op; S: 1, D: 3

>         

>         #pragma acc data create ([data]) // no-op; S: 2, D: 3

>           {

>             #pragma acc parallel copyout ([data]) // no-op; S: 3, D: 3

>               {

>               } // no-op; S: 2, D: 3

>     

>             acc_delete_finalize ([data]) // no-op; S: 2, D: 0

>     

>             acc_create ([data]) // no-op; S: 2, D: 1

>           } // no-op; S: 1, D: 1

>     

>         #pragma acc exit data delete ([data]) // no-op; S: 1, D: 0

>       } // copyout; S: 0, D: 0

>     

>     assert (!acc_is_present ([data]));

> 

> (Haven't compiled but I'm reasonably sure that the nesting and my

> manual "[action]; [S], [D]" annotations are correct.  But please

> verify, if course.)


I'm sure to make a mistake if I try to work through the rest of the
reference counts :-).

Let me know if that helps.

Thanks,

Julian

> On 2018-11-30T03:50:24-0800, Julian Brown <julian@codesourcery.com>

> wrote:

> > The model used for checking is as follows.

> >

> >  1. Each splay tree key that references a target memory descriptor

> >     increases that descriptor's refcount by 1.

> >

> >  2. Each variable listed in a target memory descriptor that links

> > back to a splay tree key increases that key's refcount by 1. Each

> > target memory descriptor's variable list is counted only once, even

> > if multiple splay tree keys point to it (via their "tgt" field).

> >

> >  3. Additional ("real") target memory descriptors may be present

> >     representing data mapped through "acc data" or "acc

> > parallel/kernels" blocks.  These descriptors have their refcount

> > bumped, and the variables linked through such blocks have their

> > refcounts bumped also (again, with "once only" semantics).

> >

> >  4. Asynchronous operations "artificially" bump the reference

> > counts for referenced target memory descriptors (but *not* for

> > linked variables/splay tree keys), in order to delay freeing mapped

> > device memory until the asynchronous operation has completed.  We

> > model this, for checking purposes only, using an off-side linked

> > list.

> >

> >  5. "Virtual" reference counts ("virtual_refcount") cannot be

> > checked purely statically, so we add the incoming value to each

> > key's statically-determined reference count ("refcount_chk"), and

> > make sure that the total matches the incoming reference count

> > ("refcount").
Thomas Schwinge May 8, 2020, 2:18 p.m. | #3
Hi Julian!

On 2020-05-07T17:11:09+0100, Julian Brown <julian@codesourcery.com> wrote:
> Sorry about the delay replying to this email!


No worries, I had other things to do, too.  ;-)


> On Thu, 30 Jan 2020 16:21:20 +0100

> Thomas Schwinge <thomas@codesourcery.com> wrote:

>> Notwithstanding the open question about how to implement this

>> checking in libgomp in a non-intrusive (performance-wise) yet

>> maintainable (avoid '#if 0') way, I have two more questions.



>> Is there a specific reason why this checking isn't also enabled for

>> libgomp OpenMP 'target' entry points?

>

> Just that it was developed in the context of adding manual deep-copy

> support to OpenACC -- OpenMP wasn't my focus at that point. So, I

> didn't try adding checking for OpenMP also. It might be interesting to

> see how that goes though, particularly with regards to dynamic data

> lifetimes in OpenMP.


ACK.


>> Can you please explain (textually?) how this checking (design per your

>> textual description below) is working in context of mixed OpenACC

>> structured ("S") and dynamic ("D") reference counts?  For example:

>>

>>     // S: 0, D: 0

>>

>>     #pragma acc enter data copyin ([data]) // copyin; S: 0, D: 1

>>

>>     acc_copyin ([data]) // no-op; S: 0, D: 2

>

> Unfortunately it's not quite that simple.


Does "not quite that simple" apply to (a) your reference count
consistency checking specifically, or to (b) libgomp implementation
peculiarities, or to (c) OpenACC reference counting semantics?

The latter (c) certainly are meant to be that simple (see OpenACC 3.0,
2.6.7. "Reference Counters", etc.), and these are what my example
illustrated.

Remember the conceptually simple implementation that we had before your
commit 378da98fcc907d05002bcd3d6ff7951f0cf485e5 "OpenACC reference count
overhaul", which to the best of my knowledge would be explained as
follows:

  - The OpenACC structured reference count corresponds to libgomp
    'key->refcount - key->dynamic_refcount'.
  - The OpenACC dynamic reference count corresponds to libgomp
    'key->dynamic_refcount'.
  - Thus, libgomp 'key->refcount' corresponds to the sum of OpenACC
    structured and dynamic reference counts.

..., and this seemed to have worked fine?  (... aside from a few specific
bugs that we fixed.)

Doing it like this meant that 'libgomp/target.c' didn't have to care
about the OpenACC-specific 'key->dynamic_refcount' at all.  Of course, we
could instead have implemented it as follows:

  - The OpenACC structured reference count corresponds to libgomp
    'key->refcount'.
  - The OpenACC dynamic reference count corresponds to libgomp
    'key->dynamic_refcount'.

..., which would've make some things simpler in 'libgomp/oacc-mem.c', but
'libgomp/target.c' then would've had to care about the OpenACC-specific
'key->dynamic_refcount'.

Now, explicitly asking the other way round: with your "overhaul", have we
now made the libgomp implementation different (and more complicated) just
to make it amenable to your reference count consistency checking (a), or
are there any actual "functional" reasons (b) that we'd not yet
considered in the old scheme?


I'm working through your explanation below -- it'll take me some more
time, but many thanks already!


Grüße
 Thomas


> The "refcount" fields (in

> either splay tree keys or target_mem_descs) do not really represent

> program-level reference counts, but rather references in the linked

> splay tree structure within libgomp. That's correct: the refcounts are

> used so as to know when data is still live, and when it can be freed.

>

> Structured data mapping operations ("acc data", "acc parallel", etc.)

> always create a target_mem_desc, with a list of target_var_descs that

> describe data mapped in that structured block. That target_mem_desc

> either "owns" a block of target memory corresponding to the structured

> data block, or it doesn't.

>

> We might have something like this (excuse ASCII art!):

>

>    +===================+        +=================+

>    | TARGET_MEM_DESC 1 |   ,-->       | TARGET_VAR_DESC |

>    +-------------------+   |  +-----------------+

>    | tgt_start...      |   |    | splay_tree_key  | --> ...

>    +-------------------+   |    +=================+

>    | target_var_desc 0 | --'

>    | target_var_desc 1 | ---.         +=================+

>    | target_var_desc 2 | -. `-> | TARGET_VAR_DESC |

>    +===================+  |     +-----------------+

>                         |     | splay_tree_key  | --> ...

>                         |     +=================+

>                         |

>                         |     +=================+

>                         `---> | TARGET_VAR_DESC |

>                               +-----------------+

>    +=================+   .-----       | splay_tree_key  |

>    | SPLAY_TREE_KEY  | <-'    +=================+

>    +-----------------+

>    | target_mem_desc | -.       +===================+

>    +=================+        '-----> | TARGET_MEM_DESC 2 |

>                               +-------------------+

>                               | tgt_start...      |

>                               +-------------------+

>                               | target_var_desc   |

>                               +===================+

>

> (Non-virtual/non-dynamic) reference counts correspond to the arrows

> between blocks in the diagram (for the pointed-to block --

> target_mem_desc or splay tree key).

>

> For a structured data mapping, say "TARGET_MEM_DESC 1" is the descriptor

> returned from gomp_map_vars.

>

> Now, "TARGET_MEM_DESC 1" and "TARGET_MEM_DESC 2" can be the same block,

> or different blocks. (Each of the TARGET_MEM_DESCs linked from splay

> tree keys, linked from TARGET_VAR_DESCs, can be a mix of such

> identical or different blocks for each of the splay tree keys linked

> from TARGET_VAR_DESCs.) In the case where they're different blocks, and

> TARGET_MEM_DESC 2 (etc.) owns its own mapped memory, TARGET_MEM_DESC 1

> may have a NULL tgt_start -- thus, not own a target data block itself.

>

> In the case of a dynamic mapping, this subtlety is especially

> important. A target_mem_desc being returned from

> gomp_map_vars{_internal} with a refcount of zero -- one which no splay

> tree keys link back to, because it does not own its own block of target

> memory -- is discarded before the function returns.

>

> So, the first time a dynamic data mapping takes place for DATA, we have:

>

>>     // S: 0, D: 0

>>

>>     #pragma acc enter data copyin ([data]) // copyin; S: 1, D: 0

>

> This is because the target_mem_desc created to describe on-target

> memory for DATA will "own" that data: nothing has referred to it

> beforehand. So there's a "real" link from the splay tree key for DATA's

> host region to the target_mem_desc we just created. (Yes, the

> splay tree key's reference counts look just like a structured data

> mapping. That was a subject for another patch.)

>

>>     acc_copyin ([data]) // no-op; S: 2, D: 1

>

> So now we have another dynamic mapping. This time, we already have a

> target_mem_desc describing DATA on the target. The

> gomp_map_vars_internal function will return NULL -- but before it does

> that, it realises that it will "lose" references in doing so. Those are

> the ones linked via the discarded target_mem_desc's variable list to

> splay tree keys that are referred to in the dynamic mapping operation.

>

> For OpenACC, that's where the "virtual" refcount comes in -- to keep

> track of those "lost" dynamic references. In particular, the "virtual"

> refcount is the count by which the structured reference count must be

> decremented when we hit an OpenACC "finalize" operation. Without that

> (cf. OpenMP), we probably wouldn't need it.

>

>>     #pragma acc data copyout ([data]) // no-op; S: 1, D: 2

>>       {

>>         acc_create ([data]) // no-op; S: 1, D: 3

>>

>>         #pragma acc data create ([data]) // no-op; S: 2, D: 3

>>           {

>>             #pragma acc parallel copyout ([data]) // no-op; S: 3, D: 3

>>               {

>>               } // no-op; S: 2, D: 3

>>

>>             acc_delete_finalize ([data]) // no-op; S: 2, D: 0

>>

>>             acc_create ([data]) // no-op; S: 2, D: 1

>>           } // no-op; S: 1, D: 1

>>

>>         #pragma acc exit data delete ([data]) // no-op; S: 1, D: 0

>>       } // copyout; S: 0, D: 0

>>

>>     assert (!acc_is_present ([data]));

>>

>> (Haven't compiled but I'm reasonably sure that the nesting and my

>> manual "[action]; [S], [D]" annotations are correct.  But please

>> verify, if course.)

>

> I'm sure to make a mistake if I try to work through the rest of the

> reference counts :-).

>

> Let me know if that helps.

>

> Thanks,

>

> Julian

>

>> On 2018-11-30T03:50:24-0800, Julian Brown <julian@codesourcery.com>

>> wrote:

>> > The model used for checking is as follows.

>> >

>> >  1. Each splay tree key that references a target memory descriptor

>> >     increases that descriptor's refcount by 1.

>> >

>> >  2. Each variable listed in a target memory descriptor that links

>> > back to a splay tree key increases that key's refcount by 1. Each

>> > target memory descriptor's variable list is counted only once, even

>> > if multiple splay tree keys point to it (via their "tgt" field).

>> >

>> >  3. Additional ("real") target memory descriptors may be present

>> >     representing data mapped through "acc data" or "acc

>> > parallel/kernels" blocks.  These descriptors have their refcount

>> > bumped, and the variables linked through such blocks have their

>> > refcounts bumped also (again, with "once only" semantics).

>> >

>> >  4. Asynchronous operations "artificially" bump the reference

>> > counts for referenced target memory descriptors (but *not* for

>> > linked variables/splay tree keys), in order to delay freeing mapped

>> > device memory until the asynchronous operation has completed.  We

>> > model this, for checking purposes only, using an off-side linked

>> > list.

>> >

>> >  5. "Virtual" reference counts ("virtual_refcount") cannot be

>> > checked purely statically, so we add the incoming value to each

>> > key's statically-determined reference count ("refcount_chk"), and

>> > make sure that the total matches the incoming reference count

>> > ("refcount").

-----------------
Mentor Graphics (Deutschland) GmbH, Arnulfstraße 201, 80634 München / Germany
Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Alexander Walter
Julian Brown May 8, 2020, 4:25 p.m. | #4
On Fri, 8 May 2020 16:18:34 +0200
Thomas Schwinge <thomas@codesourcery.com> wrote:

> >> Can you please explain (textually?) how this checking (design per

> >> your textual description below) is working in context of mixed

> >> OpenACC structured ("S") and dynamic ("D") reference counts?  For

> >> example:

> >> 

> >>     // S: 0, D: 0

> >>     

> >>     #pragma acc enter data copyin ([data]) // copyin; S: 0, D: 1

> >>     

> >>     acc_copyin ([data]) // no-op; S: 0, D: 2  

> >

> > Unfortunately it's not quite that simple.  

> 

> Does "not quite that simple" apply to (a) your reference count

> consistency checking specifically, or to (b) libgomp implementation

> peculiarities, or to (c) OpenACC reference counting semantics?


The OpenACC semantics (c) are indeed simple. I was referring to (b).
Having (a), i.e. machine-checkable invariants, is nice but I suppose
not vital. We do need to know exactly what the reference counting model
is, though!

> The latter (c) certainly are meant to be that simple (see OpenACC 3.0,

> 2.6.7. "Reference Counters", etc.), and these are what my example

> illustrated.


Aha, right.

> Remember the conceptually simple implementation that we had before

> your commit 378da98fcc907d05002bcd3d6ff7951f0cf485e5 "OpenACC

> reference count overhaul", which to the best of my knowledge would be

> explained as follows:

> 

>   - The OpenACC structured reference count corresponds to libgomp

>     'key->refcount - key->dynamic_refcount'.

>   - The OpenACC dynamic reference count corresponds to libgomp

>     'key->dynamic_refcount'.

>   - Thus, libgomp 'key->refcount' corresponds to the sum of OpenACC

>     structured and dynamic reference counts.

> 

> ..., and this seemed to have worked fine?  (... aside from a few

> specific bugs that we fixed.)


Certain things weren't right though, but that only showed up "in anger"
during the development of the manual deep-copy support. IIRC in
particular, for dynamic data mappings, "group mappings" would only try
to track the reference count for the first mapping in the group (e.g.
those comprising GOMP_MAP_TO/FROM, GOMP_MAP_TO_PSET and then
GOMP_MAP_POINTER would only try to refcount the GOMP_MAP_TO/FROM).

I'm not at all sure that the old implementation got the "subtle case"
of the target_mem_desc returned from gomp_map_vars "owning" the block
of device memory -- or not -- correct, at all. Hence writing the
verification code, to figure out what the invariants actually were
supposed to be.

Maybe the questions to ask are:

 1. With the old scheme, how do you calculate how much to decrement the
    "structured" key->refcount by when you see a finalize operation? Is
    that always correct? Do you need to know which target_mem_descs'
    variable lists refer back to this key?

 2. What do you do with the target_mem_desc returned from
    gomp_map_vars{_internal} in the dynamic data-mapping case? Is it
    then always freed at the right point? (We used to record it in an
    off-side linked list, but that scheme had its own problems.)

> Doing it like this meant that 'libgomp/target.c' didn't have to care

> about the OpenACC-specific 'key->dynamic_refcount' at all.  Of

> course, we could instead have implemented it as follows:

> 

>   - The OpenACC structured reference count corresponds to libgomp

>     'key->refcount'.

>   - The OpenACC dynamic reference count corresponds to libgomp

>     'key->dynamic_refcount'.

> 

> ..., which would've make some things simpler in 'libgomp/oacc-mem.c',

> but 'libgomp/target.c' then would've had to care about the

> OpenACC-specific 'key->dynamic_refcount'.


Yeah, I think that implies quite heavy surgery. In terms of the
key->refcount field, the overhauled OpenACC implementation is
more-or-less compatible with the way OpenMP dynamic data mapping works
with that counter, FWIW, and switching to a pure split between
key->refcount for structured and key->dynamic_refcount for dynamic
mappings would lose that. Unless we switched OpenMP over to the new
scheme too, of course.

> Now, explicitly asking the other way round: with your "overhaul",

> have we now made the libgomp implementation different (and more

> complicated) just to make it amenable to your reference count

> consistency checking (a), or are there any actual "functional"

> reasons (b) that we'd not yet considered in the old scheme?


The overhaul wasn't just done for arbitrary reasons -- trying to
rationalise the reference counting was vital in getting the manual
deep-copy support working properly. Indeed some bugs have been fixed
since then, but I'm not sure if that's all that was missing.

Are you advocating sticking with/switching back to the "old" scheme? If
so, we'd have to try to make that work with the manual deep-copy
implementation. There are still some awkward corners (particularly wrt.
unhandled cases of mixing structured & dynamic reference counts), but I
don't think those would be any easier with the "old" scheme -- and
indeed may be harder to track down without the ability to do automated
RC checking.

Thanks,

Julian

Patch

diff --git a/libgomp/libgomp.h b/libgomp/libgomp.h
index df49c1b..24cbddd 100644
--- a/libgomp/libgomp.h
+++ b/libgomp/libgomp.h
@@ -874,9 +874,26 @@  struct target_var_desc {
   uintptr_t length;
 };
 
+/* Uncomment to enable reference-count consistency checking (for development
+   use only).  */
+/*#define RC_CHECKING 1*/
+
+#ifdef RC_CHECKING
+struct target_mem_desc;
+
+struct async_tgt_use {
+  struct target_mem_desc *tgt;
+  struct async_tgt_use *next;
+};
+#endif
+
 struct target_mem_desc {
   /* Reference count.  */
   uintptr_t refcount;
+#ifdef RC_CHECKING
+  uintptr_t refcount_chk;
+  bool mark;
+#endif
   /* All the splay nodes allocated together.  */
   splay_tree_node array;
   /* Start of the target region.  */
@@ -925,6 +942,10 @@  struct splay_tree_key_s {
      "present increment" operations (via "acc enter data") refering to the same
      host-memory block.  */
   uintptr_t virtual_refcount;
+#ifdef RC_CHECKING
+  /* The recalculated reference count, for verification.  */
+  uintptr_t refcount_chk;
+#endif
   /* For a block with attached pointers, the attachment counters for each.  */
   unsigned short *attach_count;
   /* Pointer to the original mapping of "omp declare target link" object.  */
@@ -958,6 +979,10 @@  typedef struct acc_dispatch_t
     int nasyncqueue;
     struct goacc_asyncqueue **asyncqueue;
     struct goacc_asyncqueue_list *active;
+#ifdef RC_CHECKING
+    struct async_tgt_use *tgt_uses;
+    gomp_mutex_t au_lock;
+#endif
 
     __typeof (GOMP_OFFLOAD_openacc_async_construct) *construct_func;
     __typeof (GOMP_OFFLOAD_openacc_async_destruct) *destruct_func;
@@ -1085,6 +1110,12 @@  extern void gomp_detach_pointer (struct gomp_device_descr *,
 				 struct goacc_asyncqueue *, splay_tree_key,
 				 uintptr_t, bool, struct gomp_coalesce_buf *);
 
+#ifdef RC_CHECKING
+extern void dump_tgt (const char *, struct target_mem_desc *);
+extern void gomp_rc_check (struct gomp_device_descr *,
+			   struct target_mem_desc *);
+#endif
+
 extern struct target_mem_desc *gomp_map_vars (struct gomp_device_descr *,
 					      size_t, void **, void **,
 					      size_t *, void *, bool,
diff --git a/libgomp/oacc-async.c b/libgomp/oacc-async.c
index 077e28f..8b0f228 100644
--- a/libgomp/oacc-async.c
+++ b/libgomp/oacc-async.c
@@ -243,6 +243,29 @@  goacc_async_unmap_tgt (void *ptr)
 {
   struct target_mem_desc *tgt = (struct target_mem_desc *) ptr;
 
+#ifdef RC_CHECKING
+  {
+    struct gomp_device_descr *devicep = tgt->device_descr;
+    struct async_tgt_use *aup, *au;
+    gomp_mutex_lock (&devicep->openacc.async.au_lock);
+    /* Remove tgt from asynchronous-use list.  */
+    for (aup = NULL, au = devicep->openacc.async.tgt_uses; au;
+	 aup = au, au = au->next)
+      if (au->tgt == tgt)
+	{
+	  if (aup)
+	    aup->next = au->next;
+	  else
+	    devicep->openacc.async.tgt_uses = au->next;
+	  free (au);
+	  break;
+	}
+    if (!au)
+      gomp_fatal ("can't find tgt %p to remove in async list", tgt);
+    gomp_mutex_unlock (&devicep->openacc.async.au_lock);
+  }
+#endif
+
   if (tgt->refcount > 1)
     tgt->refcount--;
   else
@@ -258,6 +281,18 @@  goacc_async_copyout_unmap_vars (struct target_mem_desc *tgt,
   /* Increment reference to delay freeing of device memory until callback
      has triggered.  */
   tgt->refcount++;
+
+#ifdef RC_CHECKING
+  {
+    struct async_tgt_use *au = malloc (sizeof (struct async_tgt_use));
+    gomp_mutex_lock (&devicep->openacc.async.au_lock);
+    /* Record the asynchronous use of this target_mem_desc.  */
+    au->next = devicep->openacc.async.tgt_uses;
+    au->tgt = tgt;
+    devicep->openacc.async.tgt_uses = au;
+    gomp_mutex_unlock (&devicep->openacc.async.au_lock);
+  }
+#endif
   gomp_unmap_vars_async (tgt, true, aq);
   devicep->openacc.async.queue_callback_func (aq, goacc_async_unmap_tgt,
 					      (void *) tgt);
@@ -276,6 +311,17 @@  goacc_remove_var_async (struct gomp_device_descr *devicep, splay_tree_key n,
   struct target_mem_desc *tgt = n->tgt;
   assert (tgt);
   tgt->refcount++;
+#ifdef RC_CHECKING
+  {
+    gomp_mutex_lock (&devicep->openacc.async.au_lock);
+    struct async_tgt_use *au = malloc (sizeof (struct async_tgt_use));
+    /* Record the asynchronous use of this target_mem_desc.  */
+    au->next = devicep->openacc.async.tgt_uses;
+    au->tgt = tgt;
+    devicep->openacc.async.tgt_uses = au;
+    gomp_mutex_unlock (&devicep->openacc.async.au_lock);
+  }
+#endif
   gomp_remove_var (devicep, n);
   devicep->openacc.async.queue_callback_func (aq, goacc_async_unmap_tgt,
                                              (void *) tgt);
diff --git a/libgomp/oacc-parallel.c b/libgomp/oacc-parallel.c
index 0e9a3e8..22aafb2 100644
--- a/libgomp/oacc-parallel.c
+++ b/libgomp/oacc-parallel.c
@@ -253,6 +253,15 @@  GOACC_parallel_keyed (int device, void (*fn) (void *),
   tgt = gomp_map_vars_async (acc_dev, aq, mapnum, hostaddrs, NULL, sizes, kinds,
 			     true, GOMP_MAP_VARS_OPENACC);
 
+#ifdef RC_CHECKING
+  gomp_mutex_lock (&acc_dev->lock);
+  assert (tgt);
+  dump_tgt (__FUNCTION__, tgt);
+  tgt->prev = thr->mapped_data;
+  gomp_rc_check (acc_dev, tgt);
+  gomp_mutex_unlock (&acc_dev->lock);
+#endif
+
   devaddrs = gomp_alloca (sizeof (void *) * mapnum);
   for (i = 0; i < mapnum; i++)
     devaddrs[i] = (void *) gomp_map_val (tgt, hostaddrs, i);
@@ -270,6 +279,12 @@  GOACC_parallel_keyed (int device, void (*fn) (void *),
 					dims, tgt, aq);
       goacc_async_copyout_unmap_vars (tgt, aq);
     }
+
+#ifdef RC_CHECKING
+  gomp_mutex_lock (&acc_dev->lock);
+  gomp_rc_check (acc_dev, thr->mapped_data);
+  gomp_mutex_unlock (&acc_dev->lock);
+#endif
 }
 
 /* Legacy entry point, only provide host execution.  */
@@ -324,6 +339,12 @@  GOACC_data_start (int device, size_t mapnum,
   gomp_debug (0, "  %s: mappings prepared\n", __FUNCTION__);
   tgt->prev = thr->mapped_data;
   thr->mapped_data = tgt;
+
+#ifdef RC_CHECKING
+  gomp_mutex_lock (&acc_dev->lock);
+  gomp_rc_check (acc_dev, thr->mapped_data);
+  gomp_mutex_unlock (&acc_dev->lock);
+#endif
 }
 
 void
@@ -336,6 +357,12 @@  GOACC_data_end (void)
   thr->mapped_data = tgt->prev;
   gomp_unmap_vars (tgt, true);
   gomp_debug (0, "  %s: mappings restored\n", __FUNCTION__);
+
+#ifdef RC_CHECKING
+  gomp_mutex_lock (&thr->dev->lock);
+  gomp_rc_check (thr->dev, thr->mapped_data);
+  gomp_mutex_unlock (&thr->dev->lock);
+#endif
 }
 
 void
@@ -624,6 +651,12 @@  GOACC_enter_exit_data (int device, size_t mapnum,
 	    }
 	}
     }
+
+#ifdef RC_CHECKING
+  gomp_mutex_lock (&acc_dev->lock);
+  gomp_rc_check (acc_dev, thr->mapped_data);
+  gomp_mutex_unlock (&acc_dev->lock);
+#endif
 }
 
 static void
diff --git a/libgomp/target.c b/libgomp/target.c
index 6e115d1..b9f8ce8 100644
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -40,6 +40,9 @@ 
 #include <assert.h>
 #include <errno.h>
 #include <limits.h>
+#ifdef RC_CHECKING
+#include <stdio.h>
+#endif
 
 #ifdef PLUGIN_SUPPORT
 #include <dlfcn.h>
@@ -360,6 +363,198 @@  gomp_free_device_memory (struct gomp_device_descr *devicep, void *devptr)
     }
 }
 
+#ifdef RC_CHECKING
+void
+dump_tgt (const char *where, struct target_mem_desc *tgt)
+{
+  if (!getenv ("GOMP_DEBUG_TGT"))
+    return;
+
+  fprintf (stderr, "%s: %s: tgt=%p\n", __FUNCTION__, where, (void*) tgt);
+  fprintf (stderr, "refcount=%d\n", (int) tgt->refcount);
+  fprintf (stderr, "tgt_start=%p\n", (void*) tgt->tgt_start);
+  fprintf (stderr, "tgt_end=%p\n", (void*) tgt->tgt_end);
+  fprintf (stderr, "to_free=%p\n", tgt->to_free);
+  fprintf (stderr, "list_count=%d\n", (int) tgt->list_count);
+  for (int i = 0; i < tgt->list_count; i++)
+    {
+      fprintf (stderr, "list item %d:\n", i);
+      fprintf (stderr, "  key: %p\n", (void*) tgt->list[i].key);
+      if (tgt->list[i].key)
+	{
+	  fprintf (stderr, "  key.host_start=%p\n",
+		   (void*) tgt->list[i].key->host_start);
+	  fprintf (stderr, "  key.host_end=%p\n",
+		   (void*) tgt->list[i].key->host_end);
+	  fprintf (stderr, "  key.tgt=%p\n", (void*) tgt->list[i].key->tgt);
+	  fprintf (stderr, "  key.offset=%d\n",
+		   (int) tgt->list[i].key->tgt_offset);
+	  fprintf (stderr, "  key.refcount=%d\n",
+		   (int) tgt->list[i].key->refcount);
+	  fprintf (stderr, "  key.virtual_refcount=%d\n",
+		   (int) tgt->list[i].key->virtual_refcount);
+	  fprintf (stderr, "  key.attach_count=%p\n",
+		   (void*) tgt->list[i].key->attach_count);
+	  fprintf (stderr, "  key.link_key=%p\n",
+		   (void*) tgt->list[i].key->link_key);
+	}
+    }
+  fprintf (stderr, "\n");
+}
+
+static void
+rc_check_clear (splay_tree_node node)
+{
+  splay_tree_key k = &node->key;
+
+  k->refcount_chk = 0;
+  k->tgt->refcount_chk = 0;
+  k->tgt->mark = false;
+
+  if (node->left)
+    rc_check_clear (node->left);
+  if (node->right)
+    rc_check_clear (node->right);
+}
+
+static void
+rc_check_count (splay_tree_node node)
+{
+  splay_tree_key k = &node->key;
+  struct target_mem_desc *t;
+
+  /* Add virtual reference counts ("acc enter data", etc.) for this key.  */
+  k->refcount_chk += k->virtual_refcount;
+
+  t = k->tgt;
+  t->refcount_chk++;
+
+  if (!t->mark)
+    {
+      for (int i = 0; i < t->list_count; i++)
+	if (t->list[i].key)
+	  t->list[i].key->refcount_chk++;
+
+      t->mark = true;
+    }
+
+  if (node->left)
+    rc_check_count (node->left);
+  if (node->right)
+    rc_check_count (node->right);
+}
+
+static bool
+rc_check_verify (splay_tree_node node, bool noisy, bool errors)
+{
+  splay_tree_key k = &node->key;
+  struct target_mem_desc *t;
+
+  if (k->refcount != REFCOUNT_INFINITY)
+    {
+      if (noisy)
+	fprintf (stderr, "key %p (%p..+%d): rc=%d/%d, virt_rc=%d\n", k,
+		 (void *) k->host_start, (int) (k->host_end - k->host_start),
+		 (int) k->refcount, (int) k->refcount_chk,
+		 (int) k->virtual_refcount);
+
+      if (k->refcount != k->refcount_chk)
+	{
+	  if (noisy)
+	    fprintf (stderr, "  -- key refcount mismatch!\n");
+	  errors = true;
+	}
+
+      t = k->tgt;
+
+      if (noisy)
+	fprintf (stderr, "tgt %p: rc=%d/%d\n", t, (int) t->refcount,
+		 (int) t->refcount_chk);
+
+      if (t->refcount != t->refcount_chk)
+	{
+	  if (noisy)
+	    fprintf (stderr,
+		     "  -- target memory descriptor refcount mismatch!\n");
+	  errors = true;
+	}
+    }
+
+  if (node->left)
+    errors |= rc_check_verify (node->left, noisy, errors);
+  if (node->right)
+    errors |= rc_check_verify (node->right, noisy, errors);
+
+  return errors;
+}
+
+/* Call with device locked.  */
+
+attribute_hidden void
+gomp_rc_check (struct gomp_device_descr *devicep, struct target_mem_desc *tgt)
+{
+  splay_tree sp = &devicep->mem_map;
+
+  bool noisy = getenv ("GOMP_DEBUG_TGT") != 0;
+
+  if (noisy)
+    fprintf (stderr, "\n*** GOMP_RC_CHECK ***\n\n");
+
+  if (sp->root)
+    {
+      gomp_mutex_lock (&devicep->openacc.async.au_lock);
+      struct async_tgt_use *async_uses = devicep->openacc.async.tgt_uses;
+
+      rc_check_clear (sp->root);
+
+      for (struct target_mem_desc *t = tgt; t; t = t->prev)
+	{
+	  t->refcount_chk = 0;
+	  t->mark = false;
+	}
+      for (struct async_tgt_use *au = async_uses; au; au = au->next)
+	{
+	  struct target_mem_desc *t = au->tgt;
+	  t->refcount_chk = 0;
+	  t->mark = false;
+	}
+
+      /* Add references for interconnected splay-tree keys.  */
+      rc_check_count (sp->root);
+
+      /* Add references for the tgt for a currently-executing kernel and/or
+	 any enclosing data directives.  */
+      for (struct target_mem_desc *t = tgt; t; t = t->prev)
+	{
+	  t->refcount_chk++;
+
+	  if (!t->mark)
+	    {
+	      for (int i = 0; i < t->list_count; i++)
+		if (t->list[i].key)
+		  t->list[i].key->refcount_chk++;
+
+	      t->mark = true;
+	    }
+	}
+
+      /* Add references from in-progress asynchronous operations.  */
+      for (struct async_tgt_use *au = async_uses; au; au = au->next)
+	{
+	  struct target_mem_desc *t = au->tgt;
+	  t->refcount_chk++;
+	}
+
+      if (rc_check_verify (sp->root, noisy, false))
+	{
+	  gomp_mutex_unlock (&devicep->lock);
+	  gomp_fatal ("refcount checking failure");
+	}
+      gomp_mutex_unlock (&devicep->openacc.async.au_lock);
+    }
+}
+#endif
+
 /* Handle the case where gomp_map_lookup, splay_tree_lookup or
    gomp_map_0len_lookup found oldn for newn.
    Helper function of gomp_map_vars.  */
@@ -3274,6 +3469,10 @@  gomp_target_init (void)
 		current_device.type = current_device.get_type_func ();
 		current_device.mem_map.root = NULL;
 		current_device.state = GOMP_DEVICE_UNINITIALIZED;
+#ifdef RC_CHECKING
+		current_device.openacc.async.tgt_uses = NULL;
+		gomp_mutex_init (&current_device.openacc.async.au_lock);
+#endif
 
 		/* Augment DEVICES and NUM_DEVICES.  */
 		devices = gomp_realloc (devices,