On Thu, May 13, 2021 at 11:44:00AM +0800, Kefeng Wang wrote:
> On 2021/5/12 16:26, Mike Rapoport wrote:
> > On Wed, May 12, 2021 at 11:08:14AM +0800, Kefeng Wang wrote:
> > >
> > > On 2021/5/11 16:48, Mike Rapoport wrote:
> > > > On Mon, May 10, 2021 at 11:10:20AM +0800, Kefeng Wang wrote:
> > > > >
> > > > > > > The memory is not continuous, see MEMBLOCK:
> > > > > > > memory size = 0x4c0fffff reserved size = 0x027ef058
> > > > > > > memory.cnt = 0xa
> > > > > > > memory[0x0] [0x80a00000-0x855fffff], 0x04c00000 bytes flags: 0x0
> > > > > > > memory[0x1] [0x86a00000-0x87dfffff], 0x01400000 bytes flags: 0x0
> > > > > > > memory[0x2] [0x8bd00000-0x8c4fffff], 0x00800000 bytes flags: 0x0
> > > > > > > memory[0x3] [0x8e300000-0x8ecfffff], 0x00a00000 bytes flags: 0x0
> > > > > > > memory[0x4] [0x90d00000-0xbfffffff], 0x2f300000 bytes flags: 0x0
> > > > > > > memory[0x5] [0xcc000000-0xdc9fffff], 0x10a00000 bytes flags: 0x0
> > > > > > > memory[0x6] [0xde700000-0xde9fffff], 0x00300000 bytes flags: 0x0
> > > > > > > ...
> > > > > > >
> > > > > > > The pfn_range [0xde600,0xde700] => addr_range [0xde600000,0xde700000]
> > > > > > > is not available memory, and we won't create memmap , so with or without
> > > > > > > your patch, we can't see the range in free_memmap(), right?
> > > > > >
> > > > > > This is not available memory and we won't see the reange in free_memmap(),
> > > > > > but we still should create memmap for it and that's what my patch tried to
> > > > > > do.
> > > > > >
> > > > > > There are a lot of places in core mm that operate on pageblocks and
> > > > > > free_unused_memmap() should make sure that any pageblock has a valid memory
> > > > > > map.
> > > > > >
> > > > > > Currently, that's not the case when SPARSEMEM=y and my patch tried to fix
> > > > > > it.
> > > > > >
> > > > > > Can you please send log with my patch applied and with the printing of
> > > > > > ranges that are freed in free_unused_memmap() you've used in previous
> > > > > > mails?
> > > >
> > > > > with your patch[1] and debug print in free_memmap,
> > > > > ----> free_memmap, start_pfn = 85800, 85800000 end_pfn = 86800, 86800000
> > > > > ----> free_memmap, start_pfn = 8c800, 8c800000 end_pfn = 8e000, 8e000000
> > > > > ----> free_memmap, start_pfn = 8f000, 8f000000 end_pfn = 90000, 90000000
> > > > > ----> free_memmap, start_pfn = dcc00, dcc00000 end_pfn = de400, de400000
> > > > > ----> free_memmap, start_pfn = dec00, dec00000 end_pfn = e0000, e0000000
> > > > > ----> free_memmap, start_pfn = e0c00, e0c00000 end_pfn = e4000, e4000000
> > > > > ----> free_memmap, start_pfn = f7000, f7000000 end_pfn = f8000, f8000000
> > > >
> > > > It seems that freeing of the memory map is suboptimal still because that
> > > > code was not designed for memory layout that has more holes than Swiss
> > > > cheese.
> > > >
> > > > Still, the range [0xde600,0xde700] is not freed and there should be struct
> > > > pages for this range.
> > > >
> > > > Can you add
> > > >
> > > > dump_page(pfn_to_page(0xde600), "");
> > > >
> > > > say, in the end of memblock_free_all()?
> > > >
> > > The range [0xde600,0xde700] is not memory, so it won't create struct page
> > > for it when sparse_init?
> >
> > sparse_init() indeed does not create memory map for unpopulated memory, but
> > it has pretty coarse granularity, i.e. 64M in your configuration. A hole
> > should be at least 64M in order to skip allocation of the memory map for
> > it.
> >
> > For example, your memory layout has a hole of 192M at pfn 0xc0000 and this
> > hole won't have the memory map.
> >
> > However the hole 0xdca00 - 0xde70 will still have a memory map in the
> > section that covers 0xdc000 - 0xe0000.
> >
> > I've tried outline this in a sketch below, hope it helps.
> >
> > Memory:
> > c0000 cc000 dca00
> > --------------------------+ +--------------------------+ +----+
> > memory bank |<- hole ->| memory bank | | mb |
> > --------------------------+ +--------------------------+ +----+
> > de700 dea00
> >
> > Memory map:
> >
> > b0000 b4000 c0000 cc000 d0000 d8000 dc000
> > +--------+--------+- ... -+ +--------+- ... -+--------+---------+
> > | memmap | memmap | ... |<- hole ->| memmap | ... | memmap | memmap |
> > +--------+--------+- ... -+ +--------+- ... -+--------+---------+
> >
> >
> Thanks for the sketch, it is more clear,
>
> > > After apply patch[1], the dump_page log,
> > >
> > > page:ef3cc000 is uninitialized and poisoned
> > > raw: ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff
> > > page dumped because:
> >
> > This means that there is a memory map entry, and it got poisoned during the
> > initialization and never got reinitialized to sensible values, which would
> > be PageReserved() in this case.
> >
> > I believe this was fixed by commit 0740a50b9baa ("mm/page_alloc.c: refactor
> > initialization of struct page for holes in memory layout") in the mainline
> > tree.
> >
> > Can you backport it to your 5.10 tree and check if it helps?
> Hi Mike, the 0740a50b9baa is already in 5.10, tags/v5.10.24~5
Ah, you are using stable 5.10.y.
> commit 4c84191cbc3eff49568d3c5cccb628fa382cf7fb
> Author: Mike Rapoport <rppt@xxxxxxxxxx>
> Date: Fri Mar 12 21:07:12 2021 -0800
>
> mm/page_alloc.c: refactor initialization of struct page for holes in
> memory layout
>
> commit 0740a50b9baa4472cfb12442df4b39e2712a64a4 upstream.
>
> but check init_unavailable_range(), we need deal with the hole in the
> range of one pageblock.
>
> For our scene, pageblock range: 0xde600,0xde7ff, but the available pfn begin
> with 0xde700.
>
> If pfn(eg, 0xde600) is not valid, the step in init_unavailable_range is
> pageblock_nr_pages, and ALIGN_DOWN(pfn, pageblock_nr_pages) from 0xde600
> to 0xde700 is same, so the page range for pfn [0xde600,0xde700] won't be
> initialized.
The pfn 0xde600 is valid in the sense that there is a memory map for that
pfn. Yet, with ARM's custom pfn_valid() will treat it as invalid because
there is a hole.
> After add the following patch, the oom test could passed,
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index aaa1655cf682..0c7e04f86f9f 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -6484,13 +6484,14 @@ static u64 __meminit init_unavailable_range(unsigned
> long spfn,
> unsigned long epfn,
> int zone, int node)
> {
> - unsigned long pfn;
> + unsigned long pfn, pfn_down;
> + unsigned long epfn_down = ALIGN_DOWN(epfn, pageblock_nr_pages);
> u64 pgcnt = 0;
>
> for (pfn = spfn; pfn < epfn; pfn++) {
> - if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
> - pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
> - + pageblock_nr_pages - 1;
> + pfn_down = ALIGN_DOWN(pfn, pageblock_nr_pages);
> + if (!pfn_valid(pfn_down) && pfn_down != epfn_down) {
> + pfn = pfn_down + pageblock_nr_pages - 1;
> continue;
> }
> __init_single_page(pfn_to_page(pfn), pfn, zone, node);
I'd rather prefer to keep init_unavailable_range() and the assumption that
the memory map always covers an entire pageblock.
Can you please try the below hack. Essentially, it makes arm with SPARSEMEM
to use the generic pfn_valid() and updates the freeing of the memory map to
have the entire pageblocks covered.
If this works I'll send formal patches for those changes.
diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
index 24804f11302d..86ee711a3fdb 100644
--- a/arch/arm/Kconfig
+++ b/arch/arm/Kconfig
@@ -73,7 +73,7 @@ config ARM
select HAVE_ARCH_KGDB if !CPU_ENDIAN_BE32 && MMU
select HAVE_ARCH_KASAN if MMU && !XIP_KERNEL
select HAVE_ARCH_MMAP_RND_BITS if MMU
- select HAVE_ARCH_PFN_VALID
+# select HAVE_ARCH_PFN_VALID
select HAVE_ARCH_SECCOMP
select HAVE_ARCH_SECCOMP_FILTER if AEABI && !OABI_COMPAT
select HAVE_ARCH_THREAD_STRUCT_WHITELIST
diff --git a/mm/memblock.c b/mm/memblock.c
index 504435753259..0d7bef1b49c3 100644
--- a/mm/memblock.c
+++ b/mm/memblock.c
@@ -1928,9 +1928,11 @@ static void __init free_unused_memmap(void)
unsigned long start, end, prev_end = 0;
int i;
+#ifndef CONFIG_ARM
if (!IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) ||
IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
return;
+#endif
/*
* This relies on each bank being in address order.
@@ -1943,14 +1945,13 @@ static void __init free_unused_memmap(void)
* due to SPARSEMEM sections which aren't present.
*/
start = min(start, ALIGN(prev_end, PAGES_PER_SECTION));
-#else
+#endif
/*
* Align down here since the VM subsystem insists that the
* memmap entries are valid from the bank start aligned to
* MAX_ORDER_NR_PAGES.
*/
start = round_down(start, MAX_ORDER_NR_PAGES);
-#endif
/*
* If we had a previous bank, and there is a space
> Before:
> On node 0 totalpages: 311551
> Normal zone: 1230 pages used for memmap
> Normal zone: 0 pages reserved
> Normal zone: 157440 pages, LIFO batch:31
> Normal zone: 16384 pages in unavailable ranges
> HighMem zone: 154111 pages, LIFO batch:31
> HighMem zone: 1 pages in unavailable ranges
>
> page:ef3cc000 is uninitialized and poisoned
> raw: ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff
>
> After:
> On node 0 totalpages: 311551
> Normal zone: 1230 pages used for memmap
> Normal zone: 0 pages reserved
> Normal zone: 157440 pages, LIFO batch:31
> Normal zone: 17152 pages in unavailable ranges
> HighMem zone: 154111 pages, LIFO batch:31
> HighMem zone: 513 pages in unavailable ranges
> ...
> page:(ptrval) refcount:1 mapcount:0 mapping:00000000 index:0x0 pfn:0xde600
> flags: 0xdd001000(reserved)
> raw: dd001000 ef3cc004 ef3cc004 00000000 00000000 00000000 ffffffff 00000001
>
--
Sincerely yours,
Mike.
