Re: kern/52331: ydc driver: sleep-under-spin-mutex bugs in yds_allocmem

[Robert Elz <kre%munnari.OZ.AU@localhost>]

The following reply was made to PR kern/52331; it has been noted by GNATS.

From: Robert Elz <kre%munnari.OZ.AU@localhost>
To: gnats-bugs%NetBSD.org@localhost
Cc: nat%netbsd.org@localhost
Subject: Re: kern/52331: ydc driver: sleep-under-spin-mutex bugs in yds_allocmem
Date: Sun, 25 Jun 2017 19:53:31 +0700

     Date:        Sun, 25 Jun 2017 08:45:00 +0000 (UTC)
     From:        baijiaju1990%163.com@localhost
     Message-ID:  <20170625084500.68FE47A2B0%mollari.NetBSD.org@localhost>
 
 While your analysis tool seems good at finding code worth reviewing, I
 am not sure your review of the code to determine if there is a bug or
 not in this case is quite up to it.
 
   | The driver may sleep in interrupt, and the function call path in file "sys/dev/pci/yds.c" in NetBSD-7.1 release is:
   | yds_resume [acquire the spin mutex]
   |   yds_init
   |     yds_allocate_slots
   |       yds_allocmem
   |         bus_dmamem_alloc(BUS_DMA_WAITOK) --> may sleep
   |         bus_dmamem_map(BUS_DMA_WAITOK) --> may sleep
   |         bus_dmamem_create(BUS_DMA_WAITOK) --> may sleep
   |         bus_dmamem_load(BUS_DMA_WAITOK) --> may sleep
   | 
 
 First:
 
   | The possible fix of this bug is to replace "BUS_DMA_WAITOK" with
   | "BUS_DMA_NOWAIT".
 
 while that would avoid a potential sleep it would not actually work (if
 the sleep was ever necessary) as then the resources would not be allocated.
 
 When yds_resume() calls yds_init() the driver must have already been
 initialised, yds_init() is first called in yds_attach(), and if
 it fails, the attach also fails - in that case the code never
 reaches the code (right at the end of yds_attach() which
 establishes yds_resume as the "switch back on" power handler.
 
 So yds_resume() cannot be called unless yds_init() has succeeded.
 
 One of the events that makes yds_init() fail is if yds_allocate_slots() fails.
 
 yds_allocate_slots() only calls yds_allocmem() if KERNADDR(p) is NULL,
 where p = &sc->sc_ctrldata; (KERNADDR is p->addr)
 
 If that happens, that is, if yds_allocmem() is called, yds_allocate_slots()
 fails if yds_allocmem() fails - once again, if that happens in the call
 that comes from yds_init() from yds_attach() the attach fails, and yds_resume
 can never be called.
 
 yds_allocmem() does call bus_dmamem_alloc() (etc) as your PR revealed,
 but remember is only called if p->addr == NULL.
 
 The second bus_dma*() call in yds_allocmem() is
 
 	        error = bus_dmamem_map(sc->sc_dmatag, p->segs, p->nsegs, p->size,
                                &p->addr, BUS_DMA_WAITOK|BUS_DMA_COHERENT);
 
 That sets p_addr (unless it fails, in which case yds_allocmem returns the
 resources it has already claimed, and fails, and when that happens,
 yds_allocate_slots() also fails, which causes yds_init() to fail, which
 causes yds_attach() to fail, and yds_resume can never be called.
 
 So we know that for yds_resume to be called, the yds_init() in yds_attach()
 must have succeeded, which means that yds_allocate_slots() succeeded, which
 means that yds_allocmem() succeeded, which means that p->addr != NULL when
 yds_attach() is finished with the yds_init() call.
 
 Any later call of yds_allocmem() will find p->addr != NULL, and never call
 yds_allocmem() again (or not until yds_freemem() called from yds_free() has
 returned it all - that is only called from audio.c, but it wull take someone
 more familiar with the code than I can ever be to know whether that is
 possible in a situation where the power management resume function might
 still later be called.)
 
 But what's more, when yds_freemem() actually releases the resources
 identified by p_addr (and the others allocated by yds_allocmem()) it
 never bothers to set the pointer(s) back to NULL, so even if it were
 possible that yds_free() might be called from audio.c, and the power
 handler resume function called later, I still don't see how yds_allocmem()
 can ever be called again.
 
 I have cc'd Nathanial Sloss <nat%netbsd.org@localhost> on this reply - Nat, do you
 want this PR, or can we just assume that the bug reported is not in fact
 possible, and close it?
 
 kre