Re: [PATCH V1 0/3] Revert huge-paged linear mapping and its related fixups

From: Song Shuai
Date: Tue Jun 27 2023 - 11:13:28 EST

Hi Alex,

在 2023/6/27 19:47, Alexandre Ghiti 写道:
On Sun, Jun 25, 2023 at 10:36 PM Alexandre Ghiti <alexghiti@xxxxxxxxxxxx> wrote:

Hi Song,

On Sun, Jun 25, 2023 at 4:10 PM Song Shuai <songshuaishuai@xxxxxxxxxxx> wrote:

We have encountered these two issues about huge-paged linear mapping since v6.4-rc1:

1. Bug report: kernel paniced when system hibernates[1]

OpenSbi [v0.8,v1.3) set the PMP regions as !no-map, and the commit 3335068f8721
("riscv: Use PUD/P4D/PGD pages for the linear mapping") mapped them in linear mapping.
The hibernation process attempted to save/restore these mapped regions resulting in access fault.

This issue was temporarily fixed by commit ed309ce52218 ("RISC-V: mark hibernation as nonportable").
But as Alex's RFC and Rob's response stats in another thread [2] ,
"Hibernation is only one case. Speculative accesses could also occur."
So this fixing commit seems not the perfect answer to this issue.


2. Bug report: kernel paniced while booting (with UEFI )[3]

During the booting with UEFI, UEFI Memory Mapping overwrote the memblock.
The phys_ram_base was set as the end address of mmoderes0 (like 0x80040000 for 256 KiB mmoderes0@80000000),
which resulted the VA based on 2M-aligned PA was not 2M-aligned using va_pa_offset
(PAGE_OFFSET - phys_ram_base) to translate.

The best_map_size() from commit 3335068f8721 didn't check the virtual alignment
before choosing a map size. and then a "VA hole" was created where page faults always occurred.

This issue was fixed by commit 49a0a3731596 ("riscv: Check the virtual alignment before choosing a map size"),
But this fixing commit has a side-effect ("the possible third one" as Alex said in this thread).
There are numerous PTE allocations slowing down the boot time and consuming some system memory when UEFI booting
(Note that it's not involved when booting directly with OpenSbi, where phys_ram_base is the 2M-aligned base of DRAM).

In my test, compared with/out reverting both commit 49a0a3731596 and commit 3335068f8721,
I must wait ~20s for the linear mapping creation and mem_init_print_info() reported ~8M extra reserved memory.

Indeed, phys_ram_base is not aligned on a 2MB boundary when booting
with EDK2, IIRC that's because the first chunk of memory at
0x8000_0000 is marked as UC and is then completely evicted.


To eliminate this side-effect, We should find a way to align VA and PA on a 2MB boundary.
The simplest way is reverting the commit 3335068f8721 ("riscv: Use PUD/P4D/PGD pages for the linear mapping").


I disagree, the simplest way is to align phys_ram_base on a 2MB
boundary, either by aligning to the upper boundary (and then wastes
memory, like we used to) or by aligning to the lower boundary (but I
want to make sure it works).

I'll come up with something tomorrow.

@Song Shuai : can you test the following please?

commit a35b5e5e3f29e415f54fea064176315e79e21fb7 (HEAD ->
dev/alex/align_va_pa_v1)
Author: Alexandre Ghiti <alexghiti@xxxxxxxxxxxx>
Date: Mon Jun 5 14:26:55 2023 +0000

riscv: Start of DRAM should at least be aligned on PMD size for
the direct mapping

So that we do not end up mapping the whole linear mapping using 4K
pages, which is slow at boot time, and also very likely at runtime.

So make sure we align the start of DRAM on a PMD boundary.

Signed-off-by: Alexandre Ghiti <alexghiti@xxxxxxxxxxxx>

diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
index 4fa420faa780..4a43ec275c6d 100644
--- a/arch/riscv/mm/init.c
+++ b/arch/riscv/mm/init.c
@@ -214,8 +214,13 @@ static void __init setup_bootmem(void)
memblock_reserve(vmlinux_start, vmlinux_end - vmlinux_start);

phys_ram_end = memblock_end_of_DRAM();
+
+ /*
+ * Make sure we align the start of the memory on a PMD boundary so that
+ * at worst, we map the linear mapping with PMD mappings.
+ */
if (!IS_ENABLED(CONFIG_XIP_KERNEL))
- phys_ram_base = memblock_start_of_DRAM();
+ phys_ram_base = memblock_start_of_DRAM() & PMD_MASK;

/*
* In 64-bit, any use of __va/__pa before this point is wrong as we

I tested your patch and it really fixed what I concerned :

`There are numerous PTE allocations slowing down the boot time and consuming some system memory when UEFI booting.`

FYI, I posted the `ptdmp` and the available memory with/out the patch from my test:

```
>> v6.4

---[ Linear mapping ]---
0xff60000000000000-0xff600000001c0000 0x0000000080040000 1792K PTE D A G . . W R V
0xff600000001c0000-0xff60000000bc0000 0x0000000080200000 10M PTE D A G . . . R V
0xff60000000bc0000-0xff60000000fc0000 0x0000000080c00000 4M PTE D A G . . W R V
0xff60000000fc0000-0xff600000015c0000 0x0000000081000000 6M PTE D A G . . . R V
0xff600000015c0000-0xff600000ffdfd000 0x0000000081600000 4169972K PTE D A G . . W R V
0xff600000fffbf000-0xff600000fffc0000 0x000000017ffff000 4K PTE D A G . . W R V

>> v6.4 with the patch

---[ Linear mapping ]---
0xff60000000040000-0xff60000000200000 0x0000000080040000 1792K PTE D A G . . W R V
0xff60000000200000-0xff60000000c00000 0x0000000080200000 10M PMD D A G . . . R V
0xff60000000c00000-0xff60000001000000 0x0000000080c00000 4M PMD D A G . . W R V
0xff60000001000000-0xff60000001600000 0x0000000081000000 6M PMD D A G . . . R V
0xff60000001600000-0xff60000040000000 0x0000000081600000 1002M PMD D A G . . W R V
0xff60000040000000-0xff600000c0000000 0x00000000c0000000 2G PUD D A G . . W R V
0xff600000c0000000-0xff600000ffe00000 0x0000000140000000 1022M PMD D A G . . W R V
0xff600000ffe00000-0xff600000ffe3d000 0x000000017fe00000 244K PTE D A G . . W R V
0xff600000fffff000-0xff60000100000000 0x000000017ffff000 4K PTE D A G . . W R V
```

```
>> v6.4

Memory: 3945340K/4194048K available (8391K kernel code, 4959K rwdata, 4096K rodata, 2195K init, 476K bss, 248708K reserved, 0K cma-reserved)

>> v6.4 with the patch

Memory: 3953524K/4194048K available (8391K kernel code, 4959K rwdata, 4096K rodata, 2195K init, 476K bss, 240524K reserved, 0K cma-reserved)
```
An example of output when phys_ram_base is not aligned on a 2MB boundary:

---[ Linear mapping ]---
0xffffaf8000008000-0xffffaf8000200000 0x0000000080008000 2016K
PTE D A G . . W R V
0xffffaf8000200000-0xffffaf8000e00000 0x0000000080200000 12M
PMD D A G . . . R V
0xffffaf8000e00000-0xffffaf8001200000 0x0000000080e00000 4M
PMD D A G . . W R V
0xffffaf8001200000-0xffffaf8001a00000 0x0000000081200000 8M
PMD D A G . . . R V
0xffffaf8001a00000-0xffffaf807e600000 0x0000000081a00000 1996M
PMD D A G . . W R V
0xffffaf807e600000-0xffffaf807e714000 0x00000000fe600000 1104K
PTE D A G . . W R V
0xffffaf807e715000-0xffffaf807e718000 0x00000000fe715000 12K
PTE D A G . . W R V
0xffffaf807e71b000-0xffffaf807e71c000 0x00000000fe71b000 4K
PTE D A G . . W R V
0xffffaf807e720000-0xffffaf807e800000 0x00000000fe720000 896K
PTE D A G . . W R V
0xffffaf807e800000-0xffffaf807fe00000 0x00000000fe800000 22M
PMD D A G . . W R V
0xffffaf807fe00000-0xffffaf807ff54000 0x00000000ffe00000 1360K
PTE D A G . . W R V
0xffffaf807ff55000-0xffffaf8080000000 0x00000000fff55000 684K
PTE D A G . . W R V
0xffffaf8080000000-0xffffaf83c0000000 0x0000000100000000 13G
PUD D A G . . W R V
0xffffaf83c0000000-0xffffaf83ffe00000 0x0000000440000000 1022M
PMD D A G . . W R V
0xffffaf83ffe00000-0xffffaf8400000000 0x000000047fe00000 2M
PTE D A G . . W R V

I found that it was easier to align phys_ram_base on the lower 2MB
boundary. Aligning on the upper boundary is more complicated to me
since there may be "something" between phys_ram_base and the upper 2MB
boundary that needs to be accessed using the linear mapping (DT is
accessed using fixmap so not a problem, initrd? ACPI tables? I don't
know actually).


And would there be possible influence from the diff between phys_ram_base and memblock_start_of_DRAM()?

Weirdly simple though, I may be missing something, so any comment/test
is welcome, it is currently running our internal CI.

Thanks,

Alex


Thanks,

Alex



Using PUD/P4D/PGD pages for the linear mapping to improve the performance is marginal from a recent talk [4]
from Mike Rapoport. OpenSbi had marked all the PMP-protected regions as "no-map" [5] to practice this talk.


I noticed best_map_size() still creates some huge-paged mapping (like above 2G PUD) with this patch.
How about to revert best_map_size() to disable huge-paged mapping to practice the Mike's talk.

For all those reasons, let's revert these related commits:

- commit 3335068f8721 ("riscv: Use PUD/P4D/PGD pages for the linear mapping")
- commit 49a0a3731596 ("riscv: Check the virtual alignment before choosing a map size")
- commit ed309ce52218 ("RISC-V: mark hibernation as nonportable")

[1]: https://lore.kernel.org/linux-riscv/CAAYs2=gQvkhTeioMmqRDVGjdtNF_vhB+vm_1dHJxPNi75YDQ_Q@xxxxxxxxxxxxxx/
[2]: https://lore.kernel.org/linux-kernel/20230530080425.18612-1-alexghiti@xxxxxxxxxxxx/
[3]: https://lore.kernel.org/linux-riscv/tencent_7C3B580B47C1B17C16488EC1@xxxxxx/
[4]: https://lwn.net/Articles/931406/
[5]: https://github.com/riscv-software-src/opensbi/commit/8153b2622b08802cc542f30a1fcba407a5667ab9

Song Shuai (3):
Revert "RISC-V: mark hibernation as nonportable"
Revert "riscv: Check the virtual alignment before choosing a map size"
Revert "riscv: Use PUD/P4D/PGD pages for the linear mapping"

arch/riscv/Kconfig | 5 +---
arch/riscv/include/asm/page.h | 16 -------------
arch/riscv/mm/init.c | 43 +++++++----------------------------
arch/riscv/mm/physaddr.c | 16 -------------
drivers/of/fdt.c | 11 ++++-----
5 files changed, 14 insertions(+), 77 deletions(-)

--
2.20.1



--
Thanks
Song Shuai