riscv32 EXT4 splat, 6.8 regression?

[Nam Cao]

On 2024-04-16 Mike Rapoport wrote:
> Hi,
> 
> On Tue, Apr 16, 2024 at 01:02:20PM +0200, Björn Töpel wrote:
> > Christian Brauner <brauner at kernel.org> writes:
> > 
> > > [Adding Mike who's knowledgeable in this area]
> > 
> > >> > Further, it seems like riscv32 indeed inserts a page like that to the
> > >> > buddy allocator, when the memblock is free'd:
> > >> > 
> > >> >   | [<c024961c>] __free_one_page+0x2a4/0x3ea
> > >> >   | [<c024a448>] __free_pages_ok+0x158/0x3cc
> > >> >   | [<c024b1a4>] __free_pages_core+0xe8/0x12c
> > >> >   | [<c0c1435a>] memblock_free_pages+0x1a/0x22
> > >> >   | [<c0c17676>] memblock_free_all+0x1ee/0x278
> > >> >   | [<c0c050b0>] mem_init+0x10/0xa4
> > >> >   | [<c0c1447c>] mm_core_init+0x11a/0x2da
> > >> >   | [<c0c00bb6>] start_kernel+0x3c4/0x6de
> > >> > 
> > >> > Here, a page with VA 0xfffff000 is a added to the freelist. We were just
> > >> > lucky (unlucky?) that page was used for the page cache.
> > >> 
> > >> I just educated myself about memory mapping last night, so the below
> > >> may be complete nonsense. Take it with a grain of salt.
> > >> 
> > >> In riscv's setup_bootmem(), we have this line:
> > >> 	max_low_pfn = max_pfn = PFN_DOWN(phys_ram_end);
> > >> 
> > >> I think this is the root cause: max_low_pfn indicates the last page
> > >> to be mapped. Problem is: nothing prevents PFN_DOWN(phys_ram_end) from
> > >> getting mapped to the last page (0xfffff000). If max_low_pfn is mapped
> > >> to the last page, we get the reported problem.
> > >> 
> > >> There seems to be some code to make sure the last page is not used
> > >> (the call to memblock_set_current_limit() right above this line). It is
> > >> unclear to me why this still lets the problem slip through.
> > >> 
> > >> The fix is simple: never let max_low_pfn gets mapped to the last page.
> > >> The below patch fixes the problem for me. But I am not entirely sure if
> > >> this is the correct fix, further investigation needed.
> > >> 
> > >> Best regards,
> > >> Nam
> > >> 
> > >> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
> > >> index fa34cf55037b..17cab0a52726 100644
> > >> --- a/arch/riscv/mm/init.c
> > >> +++ b/arch/riscv/mm/init.c
> > >> @@ -251,7 +251,8 @@ static void __init setup_bootmem(void)
> > >>  	}
> > >>  
> > >>  	min_low_pfn = PFN_UP(phys_ram_base);
> > >> -	max_low_pfn = max_pfn = PFN_DOWN(phys_ram_end);
> > >> +	max_low_pfn = PFN_DOWN(memblock_get_current_limit());
> > >> +	max_pfn = PFN_DOWN(phys_ram_end);
> > >>  	high_memory = (void *)(__va(PFN_PHYS(max_low_pfn)));
> > >>  
> > >>  	dma32_phys_limit = min(4UL * SZ_1G, (unsigned long)PFN_PHYS(max_low_pfn));
> > 
> > Yeah, AFAIU memblock_set_current_limit() only limits the allocation from
> > memblock. The "forbidden" page (PA 0xc03ff000 VA 0xfffff000) will still
> > be allowed in the zone.
> > 
> > I think your patch requires memblock_set_current_limit() is
> > unconditionally called, which currently is not done.
> > 
> > The hack I tried was (which seems to work):
> > 
> > --
> > diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
> > index fe8e159394d8..3a1f25d41794 100644
> > --- a/arch/riscv/mm/init.c
> > +++ b/arch/riscv/mm/init.c
> > @@ -245,8 +245,10 @@ static void __init setup_bootmem(void)
> >          */
> >         if (!IS_ENABLED(CONFIG_64BIT)) {
> >                 max_mapped_addr = __pa(~(ulong)0);
> > -               if (max_mapped_addr == (phys_ram_end - 1))
> > +               if (max_mapped_addr == (phys_ram_end - 1)) {
> >                         memblock_set_current_limit(max_mapped_addr - 4096);
> > +                       phys_ram_end -= 4096;
> > +               }
> >         }
> 
> You can just memblock_reserve() the last page of the first gigabyte, e.g.

"last page of the first gigabyte" - why first gigabyte? Do you mean
last page of *last* gigabyte?

> 	if (!IS_ENABLED(CONFIG_64BIT)
> 		memblock_reserve(SZ_1G - PAGE_SIZE, PAGE_SIZE);
> 
> The page will still be mapped, but it will never make it to the page
> allocator.
> 
> The nice thing about it is, that memblock lets you to reserve regions that are
> not necessarily populated, so there's no need to check where the actual RAM
> ends.

I tried the suggested code, it didn't work. I think there are 2
mistakes:
 - last gigabyte, not first
 - memblock_reserve() takes physical addresses as arguments, not
   virtual addresses

The below patch fixes the problem. Is this what you really meant?

Best regards,
Nam

diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
index fa34cf55037b..ac7efdd77be8 100644
--- a/arch/riscv/mm/init.c
+++ b/arch/riscv/mm/init.c
@@ -245,6 +245,7 @@ static void __init setup_bootmem(void)
 	 * be done as soon as the kernel mapping base address is determined.
 	 */
 	if (!IS_ENABLED(CONFIG_64BIT)) {
+		memblock_reserve(__pa(-PAGE_SIZE), __pa(PAGE_SIZE));
 		max_mapped_addr = __pa(~(ulong)0);
 		if (max_mapped_addr == (phys_ram_end - 1))
 			memblock_set_current_limit(max_mapped_addr - 4096);