cachepc-linux

Fork of AMDESE/linux with modifications for CachePC side-channel attack
git clone https://git.sinitax.com/sinitax/cachepc-linux
Log | Files | Refs | README | LICENSE | sfeed.txt

Transition-notes (6617B)


      1Lemma 1:
      2	If ps_tq is scheduled, ps_tq_active is 1.  ps_tq_int() can be called
      3	only when ps_tq_active is 1.
      4Proof:	All assignments to ps_tq_active and all scheduling of ps_tq happen
      5	under ps_spinlock.  There are three places where that can happen:
      6	one in ps_set_intr() (A) and two in ps_tq_int() (B and C).
      7	Consider the sequnce of these events.  A can not be preceded by
      8	anything except B, since it is under if (!ps_tq_active) under
      9	ps_spinlock.  C is always preceded by B, since we can't reach it
     10	other than through B and we don't drop ps_spinlock between them.
     11	IOW, the sequence is A?(BA|BC|B)*.  OTOH, number of B can not exceed
     12	the sum of numbers of A and C, since each call of ps_tq_int() is
     13	the result of ps_tq execution.  Therefore, the sequence starts with
     14	A and each B is preceded by either A or C.  Moments when we enter
     15	ps_tq_int() are sandwiched between {A,C} and B in that sequence,
     16	since at any time number of B can not exceed the number of these
     17	moments which, in turn, can not exceed the number of A and C.
     18	In other words, the sequence of events is (A or C set ps_tq_active to
     19	1 and schedule ps_tq, ps_tq is executed, ps_tq_int() is entered,
     20	B resets ps_tq_active)*.
     21
     22
     23consider the following area:
     24	* in do_pd_request1(): to calls of pi_do_claimed() and return in
     25	  case when pd_req is NULL.
     26	* in next_request(): to call of do_pd_request1()
     27	* in do_pd_read(): to call of ps_set_intr()
     28	* in do_pd_read_start(): to calls of pi_do_claimed(), next_request()
     29and ps_set_intr()
     30	* in do_pd_read_drq(): to calls of pi_do_claimed() and next_request()
     31	* in do_pd_write(): to call of ps_set_intr()
     32	* in do_pd_write_start(): to calls of pi_do_claimed(), next_request()
     33and ps_set_intr()
     34	* in do_pd_write_done(): to calls of pi_do_claimed() and next_request()
     35	* in ps_set_intr(): to check for ps_tq_active and to scheduling
     36	  ps_tq if ps_tq_active was 0.
     37	* in ps_tq_int(): from the moment when we get ps_spinlock() to the
     38	  return, call of con() or scheduling ps_tq.
     39	* in pi_schedule_claimed() when called from pi_do_claimed() called from
     40	  pd.c, everything until returning 1 or setting or setting ->claim_cont
     41	  on the path that returns 0
     42	* in pi_do_claimed() when called from pd.c, everything until the call
     43	  of pi_do_claimed() plus the everything until the call of cont() if
     44	  pi_do_claimed() has returned 1.
     45	* in pi_wake_up() called for PIA that belongs to pd.c, everything from
     46	  the moment when pi_spinlock has been acquired.
     47
     48Lemma 2:
     49	1) at any time at most one thread of execution can be in that area or
     50	be preempted there.
     51	2) When there is such a thread, pd_busy is set or pd_lock is held by
     52	that thread.
     53	3) When there is such a thread, ps_tq_active is 0 or ps_spinlock is
     54	held by that thread.
     55	4) When there is such a thread, all PIA belonging to pd.c have NULL
     56	->claim_cont or pi_spinlock is held by thread in question.
     57
     58Proof:	consider the first moment when the above is not true.
     59
     60(1) can become not true if some thread enters that area while another is there.
     61	a) do_pd_request1() can be called from next_request() or do_pd_request()
     62	   In the first case the thread was already in the area.  In the second,
     63	   the thread was holding pd_lock and found pd_busy not set, which would
     64	   mean that (2) was already not true.
     65	b) ps_set_intr() and pi_schedule_claimed() can be called only from the
     66	   area.
     67	c) pi_do_claimed() is called by pd.c only from the area.
     68	d) ps_tq_int() can enter the area only when the thread is holding
     69	   ps_spinlock and ps_tq_active is 1 (due to Lemma 1).  It means that
     70	   (3) was already not true.
     71	e) do_pd_{read,write}* could be called only from the area.  The only
     72	   case that needs consideration is call from pi_wake_up() and there
     73	   we would have to be called for the PIA that got ->claimed_cont
     74	   from pd.c.  That could happen only if pi_do_claimed() had been
     75	   called from pd.c for that PIA, which happens only for PIA belonging
     76	   to pd.c.
     77	f) pi_wake_up() can enter the area only when the thread is holding
     78	   pi_spinlock and ->claimed_cont is non-NULL for PIA belonging to
     79	   pd.c.  It means that (4) was already not true.
     80
     81(2) can become not true only when pd_lock is released by the thread in question.
     82	Indeed, pd_busy is reset only in the area and thread that resets
     83	it is holding pd_lock.	The only place within the area where we
     84	release pd_lock is in pd_next_buf() (called from within the area).
     85	But that code does not reset pd_busy, so pd_busy would have to be
     86	0 when pd_next_buf() had acquired pd_lock.  If it become 0 while
     87	we were acquiring the lock, (1) would be already false, since
     88	the thread that had reset it would be in the area simulateously.
     89	If it was 0 before we tried to acquire pd_lock, (2) would be
     90	already false.
     91
     92For similar reasons, (3) can become not true only when ps_spinlock is released
     93by the thread in question.  However, all such places within the area are right
     94after resetting ps_tq_active to 0.
     95
     96(4) is done the same way - all places where we release pi_spinlock within
     97the area are either after resetting ->claimed_cont to NULL while holding
     98pi_spinlock, or after not tocuhing ->claimed_cont since acquiring pi_spinlock
     99also in the area.  The only place where ->claimed_cont is made non-NULL is
    100in the area, under pi_spinlock and we do not release it until after leaving
    101the area.
    102
    103QED.
    104
    105
    106Corollary 1: ps_tq_active can be killed.  Indeed, the only place where we
    107check its value is in ps_set_intr() and if it had been non-zero at that
    108point, we would have violated either (2.1) (if it was set while ps_set_intr()
    109was acquiring ps_spinlock) or (2.3) (if it was set when we started to
    110acquire ps_spinlock).
    111
    112Corollary 2: ps_spinlock can be killed.  Indeed, Lemma 1 and Lemma 2 show
    113that the only possible contention is between scheduling ps_tq followed by
    114immediate release of spinlock and beginning of execution of ps_tq on
    115another CPU.
    116
    117Corollary 3: assignment to pd_busy in do_pd_read_start() and do_pd_write_start()
    118can be killed.  Indeed, we are not holding pd_lock and thus pd_busy is already
    1191 here.
    120
    121Corollary 4: in ps_tq_int() uses of con can be replaced with uses of
    122ps_continuation, since the latter is changed only from the area.
    123We don't need to reset it to NULL, since we are guaranteed that there
    124will be a call of ps_set_intr() before we look at ps_continuation again.
    125We can remove the check for ps_continuation being NULL for the same
    126reason - the value is guaranteed to be set by the last ps_set_intr() and
    127we never pass it NULL.  Assignements in the beginning of ps_set_intr()
    128can be taken to callers as long as they remain within the area.