On Thu 15-02-24 15:33:30, Kent Overstreet wrote:
> On Thu, Feb 15, 2024 at 09:22:07PM +0100, Vlastimil Babka wrote:
> > On 2/15/24 19:29, Kent Overstreet wrote:
> > > On Thu, Feb 15, 2024 at 08:47:59AM -0800, Suren Baghdasaryan wrote:
> > >> On Thu, Feb 15, 2024 at 8:45 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >> >
> > >> > On Thu 15-02-24 06:58:42, Suren Baghdasaryan wrote:
> > >> > > On Thu, Feb 15, 2024 at 1:22 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >> > > >
> > >> > > > On Mon 12-02-24 13:39:17, Suren Baghdasaryan wrote:
> > >> > > > [...]
> > >> > > > > @@ -423,4 +424,18 @@ void __show_mem(unsigned int filter, nodemask_t *nodemask, int max_zone_idx)
> > >> > > > > #ifdef CONFIG_MEMORY_FAILURE
> > >> > > > > printk("%lu pages hwpoisoned\n", atomic_long_read(&num_poisoned_pages));
> > >> > > > > #endif
> > >> > > > > +#ifdef CONFIG_MEM_ALLOC_PROFILING
> > >> > > > > + {
> > >> > > > > + struct seq_buf s;
> > >> > > > > + char *buf = kmalloc(4096, GFP_ATOMIC);
> > >> > > > > +
> > >> > > > > + if (buf) {
> > >> > > > > + printk("Memory allocations:\n");
> > >> > > > > + seq_buf_init(&s, buf, 4096);
> > >> > > > > + alloc_tags_show_mem_report(&s);
> > >> > > > > + printk("%s", buf);
> > >> > > > > + kfree(buf);
> > >> > > > > + }
> > >> > > > > + }
> > >> > > > > +#endif
> > >> > > >
> > >> > > > I am pretty sure I have already objected to this. Memory allocations in
> > >> > > > the oom path are simply no go unless there is absolutely no other way
> > >> > > > around that. In this case the buffer could be preallocated.
> > >> > >
> > >> > > Good point. We will change this to a smaller buffer allocated on the
> > >> > > stack and will print records one-by-one. Thanks!
> > >> >
> > >> > __show_mem could be called with a very deep call chains. A single
> > >> > pre-allocated buffer should just do ok.
> > >>
> > >> Ack. Will do.
> > >
> > > No, we're not going to permanently burn 4k here.
> > >
> > > It's completely fine if the allocation fails, there's nothing "unsafe"
> > > about doing a GFP_ATOMIC allocation here.
> >
> > Well, I think without __GFP_NOWARN it will cause a warning and thus
> > recursion into __show_mem(), potentially infinite? Which is of course
> > trivial to fix, but I'd myself rather sacrifice a bit of memory to get this
> > potentially very useful output, if I enabled the profiling. The necessary
> > memory overhead of page_ext and slabobj_ext makes the printing buffer
> > overhead negligible in comparison?
>
> __GFP_NOWARN is a good point, we should have that.
>
> But - and correct me if I'm wrong here - doesn't an OOM kick in well
> before GFP_ATOMIC 4k allocations are failing?
Not really, GFP_ATOMIC users can compete with reclaimers and consume
those reserves.
> I'd expect the system to
> be well and truly hosed at that point.
It is OOMed...
> If we want this report to be 100% reliable, then yes the preallocated
> buffer makes sense - but I don't think 100% makes sense here; I think we
> can accept ~99% and give back that 4k.
Think about that from the memory reserves consumers. The atomic reserve
is a scarse resource and now you want to use it for debugging purposes
for which you could have preallocated.
--
Michal Hocko
SUSE Labs
