Unsafe reads of the av_entry_pool because the util_av implementation does not require a lock to be held in some cases

[sunkuamzn]

The util_av implementation requires a mutex (will be a genlock after #10803) to be held for when writing to the av_entry_pool

libfabric/prov/util/src/util_av.c

Line 305 in a07fe39

assert(ofi_mutex_held(&av->lock));

libfabric/prov/util/src/util_av.c

Line 278 in a07fe39

assert(ofi_mutex_held(&av->lock));

libfabric/prov/util/src/util_av.c

Line 337 in a07fe39

assert(ofi_mutex_held(&av->lock));

but not when reading from the av_entry_pool

libfabric/prov/util/src/util_av.c

Lines 243 to 249 in a07fe39

	void *ofi_av_get_addr(struct util_av *av, fi_addr_t fi_addr)
	{
	struct util_av_entry *entry;
	entry = ofi_bufpool_get_ibuf(av->av_entry_pool, fi_addr);
	return entry->data;
	}

libfabric/prov/util/src/util_av.c

Lines 419 to 424 in a07fe39

	size_t ofi_av_size(struct util_av *av)
	{
	return av->av_entry_pool->entry_cnt ?
	av->av_entry_pool->entry_cnt :
	av->av_entry_pool->attr.chunk_cnt;
	}

except for ofi_av_lookup_fi_addr

libfabric/prov/util/src/util_av.c

Lines 362 to 364 in a07fe39

	ofi_mutex_lock(&av->lock);
	fi_addr = ofi_av_lookup_fi_addr_unsafe(av, addr);
	ofi_mutex_unlock(&av->lock);

SHM, SM2 (and maybe other providers) are performing unsafe reads of the av_entry_pool

libfabric/prov/shm/src/smr_ep.c

Line 215 in a07fe39

id = smr_addr_lookup(ep->util_ep.av, fi_addr);

libfabric/prov/shm/src/smr_av.c

Line 242 in a07fe39

id = smr_addr_lookup(util_av, fi_addr);

libfabric/prov/sm2/src/sm2_ep.c

Line 169 in a07fe39

*gid = *((sm2_gid_t *) ofi_av_get_addr(ep->util_ep.av, fi_addr));