Re: [CHECKER] Race condition in i2o_core.c

["Randy.Dunlap" <rddunlap@osdl.org>]

On Thu, 01 Apr 2004 17:46:00 -0800 Ken Ashcraft wrote:

| Looks like there is a race condition in i2o_core_reply involving the 
| variable "evt_in".  Notice that the increment of evt_in is protected by the 
| lock, but the reads are not protected.  It looks like "events" should also 
| be protected by the lock.  If this is not a race condition, the increment 
| should not be inside the critical section.
| 
| Feedback is appreciated.
| 
| thanks,
| Ken Ashcraft
| -----------------------------------
| /home/kash/interface/linux-2.6.3/drivers/message/i2o/i2o_core.c:264:i2o_core_reply: 
| ERROR:RACE: 264:264:Possible race condition on variable "evt_in", No locks 
| held on read on line 264, Locks {&i2o_evt_lock } held on write on line 268 
| [COUNTER=i2o_handler.reply] [fit=1] [fit_fn=1] [fn_ex=0] [fn_counter=1] 
| [ex=1] [counter=1] [z = -2.91998558035372] [fn-z = -4.35889894354067]
| 
|                  return;
|          }
| 
|          if(m->function == I2O_CMD_UTIL_EVT_REGISTER)
|          {
| 
| Error --->
|                  memcpy(events[evt_in].msg, msg, (msg[0]>>16)<<2);

static int evt_in;
Access to <int> is (considered to be) atomic.
However, the MODINC() macro is nowhere close to atomic.
Does that help?

|                  events[evt_in].iop = c;
| 
|                  spin_lock(&i2o_evt_lock);
|                  MODINC(evt_in, I2O_EVT_Q_LEN);
|                  if(evt_q_len == I2O_EVT_Q_LEN)
|                          MODINC(evt_out, I2O_EVT_Q_LEN);
|                  else
|                          evt_q_len++;
|                  spin_unlock(&i2o_evt_lock);
| 
|                  up(&evt_sem);
|                  wake_up_interruptible(&evt_wait);
|                  return;


--
~Randy
(Again.  Sometimes I think ln -s /usr/src/linux/.config .signature)