On 18/10/14 00:46, leo@yuriev.ru wrote:

(...)

• ti->mti_readers[i].mr_pid = pid;
• ti->mti_readers[i].mr_tid = tid;

• r = &ti->mti_readers[i];
• r->mr_txnid = (txnid_t)-1;
• r->mr_tid = tid;
• r->mr_pid = pid; /* should be written last, see ITS#. */ if (i == nr) ti->mti_numreaders = ++nr; /* Save numreaders for un-mutexed mdb_env_close() */ env->me_numreaders = nr;

Actually that too is too early to set mr_pid: We must be sure env_close() will reset it. But it is also too late, we must get rid of any garbage value in md_pid at once to protect it from other processes' env_close(). Fixing.

Also, I'll rename the confusingly named me_numreaders to me_close_readers and drop some confused code using it.