On Mon 26-08-24 14:38:40, Uladzislau Rezki wrote:
> On Mon, Aug 26, 2024 at 09:52:42AM +0200, Michal Hocko wrote:
> > On Fri 23-08-24 18:42:47, Uladzislau Rezki wrote:
> > [...]
> > > @@ -3666,7 +3655,16 @@ static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
> > > set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
> > > page_order = vm_area_page_order(area);
> > >
> > > - area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
> > > + /*
> > > + * Higher order nofail allocations are really expensive and
> > > + * potentially dangerous (pre-mature OOM, disruptive reclaim
> > > + * and compaction etc.
> > > + *
> > > + * Please note, the __vmalloc_node_range_noprof() falls-back
> > > + * to order-0 pages if high-order attempt has been unsuccessful.
> > > + */
> > > + area->nr_pages = vm_area_alloc_pages(page_order ?
> > > + gfp_mask &= ~__GFP_NOFAIL : gfp_mask | __GFP_NOWARN,
> > > node, page_order, nr_small_pages, area->pages);
> > >
> > > atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
> > > <snip>
> > >
> > > Is that aligned with your wish?
> >
> > I am not a great fan of modifying gfp_mask inside the ternary operator
> > like that. It makes the code harder to read. Is there any actual reason
> > to simply drop GFP_NOFAIL unconditionally and rely do the NOFAIL
> > handling for all orders at the same place?
> >
> 1. So, for bulk we have below:
>
> /* gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL; */
>
> I am not sure if we need it but it says it does not support it which
> is not clear for me why we have to drop __GFP_NOFAIL for bulk(). There
> is a fallback to a single page allocator. If passing __GFP_NOFAIL does
> not trigger any warning or panic a system, then i do not follow why
> we drop that flag.
>
> Is that odd?
I suspect this was a pre-caution more than anything.
> 2. High-order allocations. Do you think we should not care much about
> it when __GFP_NOFAIL is set? Same here, there is a fallback for order-0
> if "high" fails, it is more likely NO_FAIL succeed for order-0. Thus
> keeping NOFAIL for high-order sounds like not a good approach to me.
We should avoid high order allocations with GFP_NOFAIL at all cost.
> 3. "... at the same place?"
> Do you mean in the __vmalloc_node_range_noprof()?
>
> __vmalloc_node_range_noprof()
> -> __vmalloc_area_node(gfp_mask)
> -> vm_area_alloc_pages()
>
> if, so it is not straight forward, i.e. there is one more allocation:
>
> <snip>
> static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
> pgprot_t prot, unsigned int page_shift,
> int node)
> {
> ...
> /* Please note that the recursion is strictly bounded. */
> if (array_size > PAGE_SIZE) {
> area->pages = __vmalloc_node_noprof(array_size, 1, nested_gfp, node,
> area->caller);
> } else {
> area->pages = kmalloc_node_noprof(array_size, nested_gfp, node);
> }
> ...
> }
> <snip>
>
> whereas it is easier to do it inside of the __vmalloc_area_node().
Right. The allocation path is quite convoluted here. If it is just too
much of a hassle to implement NOFAIL at a single place then we should
aim at reducing that. Having that at 3 different layers is just begging
for inconsistences.
--
Michal Hocko
SUSE Labs
