'make -j16 universe' gives SIReset

[Alan Cox]

On 07/01/2011 19:23, Marius Strobl wrote:
> On Fri, Jul 01, 2011 at 08:17:52AM +1000, Peter Jeremy wrote:
>> [Moving back on-list]
>>
>> On 2011-Jun-30 06:30:08 +0800, Marius Strobl<marius at alchemy.franken.de>  wrote:
>>> On Thu, Jun 30, 2011 at 08:00:10AM +1000, Peter Jeremy wrote:
>>>> On 2011-Jun-29 19:54:44 +0200, Marius Strobl<marius at alchemy.franken.de>  wrote:
>>>>> On Wed, Jun 29, 2011 at 12:54:33PM +1000, Peter Jeremy wrote:
>>>>>> My V890 has been running "make -j32 buildworld" in a loop for a
>>>>>> week now without problems so I think that was the problem.
>>>> OTOH, a V440 that has been running similar load for a similar period
>>>> died overnight with:
>>>>
>>>> panic: uma_small_alloc: free page still has mappings!
>>>> VNASSERT failed
>>>> cpuid = 3
>>>> 0xfffff800079643c0: KDB: enter: panic
>> ...
>>>> I'm fairly sure that is the same kernel but will double-check and
>>>> investigate that panic further.
>> FWIW, that kernel didn't have the latest patchset (adding Zeus support).
> That shouldn't make a difference; the later version only adds the
> SPARC64 bits as you already noticed and adjusts the boot loader to
> compile again. I made no changes to the existing parts apart from
> fixing a comment. Besides I see no connection between fixing the
> gross user TLB flushing and the below problem so far.
>
>>> Ok, this appears to be an unrelated problem though. Alan, do you
>>> have an idea what could be causing this?
>> I managed to get the same panic (though different traceback) on the
>> V890 after about an hour of pho@'s stress test with INCARNATIONS=150:
>>
>> panic: uma_small_alloc: free page still has mappings!
>> cpuid = 1
>> KDB: enter: panic
>> [ thread pid 142 tid 100196 ]
>> Stopped at      kdb_enter+0x80: ta              %xcc, 1
>> db>  where
>> Tracing pid 142 tid 100196 td 0xfffff8a016ace880
>> panic() at panic+0x20c
>> uma_small_alloc() at uma_small_alloc+0xe8
>> keg_alloc_slab() at keg_alloc_slab+0xc8
>> keg_fetch_slab() at keg_fetch_slab+0x218
>> zone_fetch_slab() at zone_fetch_slab+0x44
>> uma_zalloc_arg() at uma_zalloc_arg+0x60c
>> m_getm2() at m_getm2+0x134
>> m_uiotombuf() at m_uiotombuf+0x4c
>> sosend_generic() at sosend_generic+0x420
>> sosend() at sosend+0x2c
>> soo_write() at soo_write+0x3c
>> dofilewrite() at dofilewrite+0x7c
>> kern_writev() at kern_writev+0x38
>> write() at write+0x4c
>> syscallenter() at syscallenter+0x270
>> syscall() at syscall+0x74
>> -- syscall (4, FreeBSD ELF64, write) %o7=0x101db4 --
>> userland() at 0x405936c8
>> user trace: trap %o7=0x101db4
>> pc 0x405936c8, sp 0x7fdffffd8a1
>> pc 0x101f44, sp 0x7fdffffd9a1
>> pc 0x104604, sp 0x7fdffffda81
>> pc 0x1046f0, sp 0x7fdffffdb51
>> pc 0x104994, sp 0x7fdffffdc21
>> pc 0x104d90, sp 0x7fdffffdd01
>> pc 0x101610, sp 0x7fdffffde41
>> pc 0x4020cff4, sp 0x7fdffffdf01
>> done
>> db>
>>
>> I've got a crashdump on the V440 but discovered that gdb reports
>> "GDB can't read core files on this machine." so it isn't much use.
>> Any suggestions on how to debug this?
> The VM and its interaction with the MD code are beyond me, I hope
> Alan can chime in here. Reading through the code I see a possible
> path which could lead to this though; tsb_tte_enter(), which is
> the only place where TD_PV ever is set and also only in case of
> managed pages, always calls pmap_cache_enter(), which together
> with pmap_cache_remove() does the page color handling. In
> pmap_remove_all() however, pmap_cache_remove() is only called for
> managed pages, so for unmanaged pages we might miss the removal
> of the mapping from the the color used. I've no idea though if
> this actually is relevant, i.e. whether the VM ever calls
> pmap_remove_all() for unmanaged pages.

In HEAD, it does not.  Other architectures have an assertion forbidding 
pmap_remove_all() calls on unmanaged pages.  (Btw, I'm happy to add this 
assertion to sparc64's pmap if you like.)  In older versions, calling 
pmap_remove_all() on unmanaged pages is expected to be a harmless NOP 
that's just a waste of cycles.

With unmanaged pages, it is expected that pmap_remove() is used to 
destroy mappings before the page is freed.

For years, vm_page_free{,_toq}() has asserted that the page has no 
managed mappings:

         if ((m->flags & PG_UNMANAGED) == 0) {
                 vm_page_lock_assert(m, MA_OWNED);
                 KASSERT(!pmap_page_is_mapped(m),
                     ("vm_page_free_toq: freeing mapped page %p", m));
         }

As a debugging aid, you might want to add an additional check here on 
colors.

> ... Tentatively I'd say it
> doesn't, in which case the only solution I see is to exclude
> unmanaged pages from the page color handling and caching, which
> I don't know whether it's safe (besides impacting performance).
> Unfortunately, with my gear I can't reproduce this. Could you
> please try the below patch? I've no idea whether it's correct
> but might give another datapoint.
>
> Marius
>
> Index: pmap.c
> ===================================================================
> --- pmap.c	(revision 223705)
> +++ pmap.c	(working copy)
> @@ -1382,21 +1385,21 @@ pmap_remove_all(vm_page_t m)
>   	vm_page_lock_queues();
>   	for (tp = TAILQ_FIRST(&m->md.tte_list); tp != NULL; tp = tpn) {
>   		tpn = TAILQ_NEXT(tp, tte_link);
> -		if ((tp->tte_data&  TD_PV) == 0)
> -			continue;
>   		pm = TTE_GET_PMAP(tp);
>   		va = TTE_GET_VA(tp);
>   		PMAP_LOCK(pm);
>   		if ((tp->tte_data&  TD_WIRED) != 0)
>   			pm->pm_stats.wired_count--;
> -		if ((tp->tte_data&  TD_REF) != 0)
> -			vm_page_flag_set(m, PG_REFERENCED);
> -		if ((tp->tte_data&  TD_W) != 0)
> -			vm_page_dirty(m);
> +		if ((tp->tte_data&  TD_PV) != 0) {
> +			if ((tp->tte_data&  TD_REF) != 0)
> +				vm_page_flag_set(m, PG_REFERENCED);
> +			if ((tp->tte_data&  TD_W) != 0)
> +				vm_page_dirty(m);
> +			pm->pm_stats.resident_count--;
> +		}
>   		tp->tte_data&= ~TD_V;
>   		tlb_page_demap(pm, va);
>   		TAILQ_REMOVE(&m->md.tte_list, tp, tte_link);
> -		pm->pm_stats.resident_count--;
>   		pmap_cache_remove(m, va);
>   		TTE_ZERO(tp);
>   		PMAP_UNLOCK(pm);
>