Re: [PATCH] riscv: mm: init: Consolidate vars, functions

From: Alex Ghiti
Date: Sun Jun 06 2021 - 03:28:42 EST


Hi Jisheng,

Le 4/06/2021 à 19:09, Jisheng Zhang a écrit :
On Thu, 3 Jun 2021 19:54:18 +0800
Jisheng Zhang <jszhang3@xxxxxxxxxxxxxxxx> wrote:

On Thu, 3 Jun 2021 08:27:14 +0800
Jisheng Zhang <jszhang3@xxxxxxxxxxxxxxxx> wrote:

On Wed, 2 Jun 2021 12:54:40 -0700
Nathan Chancellor <nathan@xxxxxxxxxx> wrote:
On Wed, Jun 02, 2021 at 11:58:51PM +0800, Jisheng Zhang wrote:
On Wed, 2 Jun 2021 23:12:26 +0800
Jisheng Zhang <jszhang3@xxxxxxxxxxxxxxxx> wrote:
On Tue, 1 Jun 2021 13:21:05 -0700
Nathan Chancellor <nathan@xxxxxxxxxx> wrote:
Hi Jisheng,

Hi Nathan,

On Sun, May 16, 2021 at 09:15:56PM +0800, Jisheng Zhang wrote:
From: Jisheng Zhang <jszhang@xxxxxxxxxx>

Consolidate the following items in init.c

Staticize global vars as much as possible;
Add __initdata mark if the global var isn't needed after init
Add __init mark if the func isn't needed after init
Add __ro_after_init if the global var is read only after init

Signed-off-by: Jisheng Zhang <jszhang@xxxxxxxxxx>
---
arch/riscv/include/asm/set_memory.h | 2 +-
arch/riscv/mm/init.c | 36 +++++++++++++++--------------
2 files changed, 20 insertions(+), 18 deletions(-)

diff --git a/arch/riscv/include/asm/set_memory.h b/arch/riscv/include/asm/set_memory.h
index 086f757e8ba3..9d4d455726d4 100644
--- a/arch/riscv/include/asm/set_memory.h
+++ b/arch/riscv/include/asm/set_memory.h
@@ -27,7 +27,7 @@ static inline int set_memory_rw_nx(unsigned long addr, int numpages) { return 0;
#endif
#if defined(CONFIG_64BIT) && defined(CONFIG_STRICT_KERNEL_RWX)
-void protect_kernel_linear_mapping_text_rodata(void);
+void __init protect_kernel_linear_mapping_text_rodata(void);
#else
static inline void protect_kernel_linear_mapping_text_rodata(void) {}
#endif
diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
index 4c4c92ce0bb8..eac2d5c27b3e 100644
--- a/arch/riscv/mm/init.c
+++ b/arch/riscv/mm/init.c
@@ -53,7 +53,7 @@ struct pt_alloc_ops {
#endif
};
-static phys_addr_t dma32_phys_limit __ro_after_init;
+static phys_addr_t dma32_phys_limit __initdata;
static void __init zone_sizes_init(void)
{
@@ -184,7 +184,7 @@ extern char _sdata[], _edata[];
#endif /* CONFIG_XIP_KERNEL */
#ifdef CONFIG_MMU
-static struct pt_alloc_ops _pt_ops __ro_after_init;
+static struct pt_alloc_ops _pt_ops __initdata;
#ifdef CONFIG_XIP_KERNEL
#define pt_ops (*(struct pt_alloc_ops *)XIP_FIXUP(&_pt_ops))
@@ -200,13 +200,13 @@ EXPORT_SYMBOL(va_pa_offset);
#endif
/* Offset between kernel mapping virtual address and kernel load address */
#ifdef CONFIG_64BIT
-unsigned long va_kernel_pa_offset;
+unsigned long va_kernel_pa_offset __ro_after_init;
EXPORT_SYMBOL(va_kernel_pa_offset);
#endif
#ifdef CONFIG_XIP_KERNEL
#define va_kernel_pa_offset (*((unsigned long *)XIP_FIXUP(&va_kernel_pa_offset)))
#endif
-unsigned long va_kernel_xip_pa_offset;
+unsigned long va_kernel_xip_pa_offset __ro_after_init;
EXPORT_SYMBOL(va_kernel_xip_pa_offset);
#ifdef CONFIG_XIP_KERNEL
#define va_kernel_xip_pa_offset (*((unsigned long *)XIP_FIXUP(&va_kernel_xip_pa_offset)))
@@ -216,7 +216,7 @@ EXPORT_SYMBOL(pfn_base);
pgd_t swapper_pg_dir[PTRS_PER_PGD] __page_aligned_bss;
pgd_t trampoline_pg_dir[PTRS_PER_PGD] __page_aligned_bss;
-pte_t fixmap_pte[PTRS_PER_PTE] __page_aligned_bss;
+static pte_t fixmap_pte[PTRS_PER_PTE] __page_aligned_bss;
pgd_t early_pg_dir[PTRS_PER_PGD] __initdata __aligned(PAGE_SIZE);
@@ -253,7 +253,7 @@ static inline pte_t *__init get_pte_virt_fixmap(phys_addr_t pa)
return (pte_t *)set_fixmap_offset(FIX_PTE, pa);
}
-static inline pte_t *get_pte_virt_late(phys_addr_t pa)
+static inline pte_t *__init get_pte_virt_late(phys_addr_t pa)
{
return (pte_t *) __va(pa);
}
@@ -272,7 +272,7 @@ static inline phys_addr_t __init alloc_pte_fixmap(uintptr_t va)
return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
}
-static phys_addr_t alloc_pte_late(uintptr_t va)
+static phys_addr_t __init alloc_pte_late(uintptr_t va)
{
unsigned long vaddr;
@@ -296,10 +296,10 @@ static void __init create_pte_mapping(pte_t *ptep,
#ifndef __PAGETABLE_PMD_FOLDED
-pmd_t trampoline_pmd[PTRS_PER_PMD] __page_aligned_bss;
-pmd_t fixmap_pmd[PTRS_PER_PMD] __page_aligned_bss;
-pmd_t early_pmd[PTRS_PER_PMD] __initdata __aligned(PAGE_SIZE);
-pmd_t early_dtb_pmd[PTRS_PER_PMD] __initdata __aligned(PAGE_SIZE);
+static pmd_t trampoline_pmd[PTRS_PER_PMD] __page_aligned_bss;
+static pmd_t fixmap_pmd[PTRS_PER_PMD] __page_aligned_bss;
+static pmd_t early_pmd[PTRS_PER_PMD] __initdata __aligned(PAGE_SIZE);
+static pmd_t early_dtb_pmd[PTRS_PER_PMD] __initdata __aligned(PAGE_SIZE);
#ifdef CONFIG_XIP_KERNEL
#define trampoline_pmd ((pmd_t *)XIP_FIXUP(trampoline_pmd))
@@ -319,7 +319,7 @@ static pmd_t *__init get_pmd_virt_fixmap(phys_addr_t pa)
return (pmd_t *)set_fixmap_offset(FIX_PMD, pa);
}
-static pmd_t *get_pmd_virt_late(phys_addr_t pa)
+static pmd_t *__init get_pmd_virt_late(phys_addr_t pa)
{
return (pmd_t *) __va(pa);
}
@@ -336,7 +336,7 @@ static phys_addr_t __init alloc_pmd_fixmap(uintptr_t va)
return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
}
-static phys_addr_t alloc_pmd_late(uintptr_t va)
+static phys_addr_t __init alloc_pmd_late(uintptr_t va)
{
unsigned long vaddr;
@@ -454,14 +454,16 @@ asmlinkage void __init __copy_data(void)
#error "setup_vm() is called from head.S before relocate so it should not use absolute addressing."
#endif
-uintptr_t load_pa, load_sz;
+static uintptr_t load_pa __initdata;

Making load_pa static causing clang built kernels to no longer boot,
hanging after just a few lines of output in the console:

https://github.com/ClangBuiltLinux/continuous-integration2/runs/2717606254?check_suite_focus=true

I am not sure why that would make a difference nor why GCC is okay with
it. If it is a clang bug, it appears to be there for a while, given that
it reproduces back to clang-11.

I can reproduce the issue. Here are my findindings:

* gcc + binutils can't reproduce it
* clang + llvm-utils + ias can reproduce it
* clang + binutils can reproduce it

Yes, this seems like something strictly related to clang so that is what
I have been testing with.
All below tests are done with clang + binutils.

Then I applied below modification:

-static uintptr_t load_pa __initdata;
+uintptr_t load_pa __initdata;

I got below panic log:
[ 0.015418] Unable to handle kernel paging request at virtual address fffffffffffffff9
[ 0.016432] Oops [#1]
[ 0.016679] Modules linked in:
[ 0.017103] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.13.0-rc1+ #5
[ 0.017711] Hardware name: riscv-virtio,qemu (DT)
[ 0.018201] epc : trace_hardirqs_on+0x60/0xb2
[ 0.018582] ra : restore_all+0xe/0x66
[ 0.018879] epc : ffffffff800cb09a ra : ffffffff800027b8 sp : ffffffff80c03dd0
[ 0.019376] gp : ffffffff80d001c8 tp : ffffffff80c0c180 t0 : 0000000000000020
[ 0.019870] t1 : ffffffff80006e40 t2 : ffffffff800d2e0a s0 : ffffffff80c03e00
[ 0.020346] s1 : 0000000000000001 a0 : 0000000000000001 a1 : 0000000000000001
[ 0.020800] a2 : 0000000000000001 a3 : 0000000000000000 a4 : 0000000000000000
[ 0.021243] a5 : 0000000000000000 a6 : 0000000000000000 a7 : 0000000054494d45
[ 0.021717] s2 : ffffffff800027b8 s3 : ffffffff80d35968 s4 : ffffffff8061e1d8
[ 0.022179] s5 : ffffffff80c0c180 s6 : ffffffe000e34b50 s7 : 00000000800130f0
[ 0.022674] s8 : 000000000000007f s9 : 0000000080012010 s10: 0000000000000000
[ 0.023176] s11: 0000000000000000 t3 : ffffffff80d00108 t4 : ffffffff80006e40
[ 0.023693] t5 : ffffffff80006e40 t6 : ffffffff800d2e0a
[ 0.024153] status: 0000000000000100 badaddr: fffffffffffffff9 cause: 000000000000000d
[ 0.025367] Call Trace:
[ 0.025749] [<ffffffff800cb09a>] trace_hardirqs_on+0x60/0xb2
[ 0.026402] [<ffffffff800027b8>] restore_all+0xe/0x66
[ 0.027261] Unable to handle kernel paging request at virtual address fffffffffffffffa
[ 0.027827] Oops [#2]
[ 0.028013] Modules linked in:
[ 0.028321] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G D 5.13.0-rc1+ #5
[ 0.028839] Hardware name: riscv-virtio,qemu (DT)
[ 0.029166] epc : trace_hardirqs_on+0x60/0xb2
[ 0.029505] ra : restore_all+0xe/0x66
[ 0.029785] epc : ffffffff800cb09a ra : ffffffff800027b8 sp : ffffffff80c03a80
[ 0.030266] gp : ffffffff80d001c8 tp : ffffffff80c0c180 t0 : 0000000000000020
[ 0.030748] t1 : ffffffff80006e40 t2 : ffffffff800d2e0a s0 : ffffffff80c03ab0
[ 0.031227] s1 : 0000000000000001 a0 : 0000000000000002 a1 : 0000000000000002
[ 0.031717] a2 : 0000000000000001 a3 : 0000000000000000 a4 : 0000000000000000
[ 0.032199] a5 : 0000000000000000 a6 : 0000000000000000 a7 : 0000000054494d45
[ 0.032680] s2 : ffffffff800027b8 s3 : ffffffff80d35968 s4 : ffffffff8061e1d8
[ 0.033160] s5 : ffffffff80c0c180 s6 : ffffffe000e34b50 s7 : 00000000800130f0
[ 0.033642] s8 : 000000000000007f s9 : 0000000080012010 s10: 0000000000000000
[ 0.034123] s11: 0000000000000000 t3 : ffffffff80d00108 t4 : ffffffff80006e40
[ 0.034601] t5 : ffffffff80006e40 t6 : ffffffff800d2e0a
[ 0.034965] status: 0000000000000100 badaddr: fffffffffffffffa cause: 000000000000000d
[ 0.035492] Call Trace:
[ 0.035682] [<ffffffff800cb09a>] trace_hardirqs_on+0x60/0xb2
[ 0.036077] [<ffffffff800027b8>] restore_all+0xe/0x66
[ 0.036545] ---[ end trace 7f4fbff09d927668 ]---
[ 0.037188] Kernel panic - not syncing: Attempted to kill the idle task!
[ 0.038107] ---[ end Kernel panic - not syncing: Attempted to kill the idle task! ]---

Then I checked 5.13-rc1, above panic log can be reproduced too. So the issue
should exist there for a while. I never tried clang with riscv, did you remember
which last commit or version clang works, I may try to bisect.
V> >

More findings:

*The above panic issue can also be seen from 5.12-rc2. If disable FTRACE, then
the panic disappears, kernel can boot

*so I retested riscv next tree w/ FTRACE disabled, kernel can boot w/ below
modification:

Yeah, I do not enable CONFIG_FTRACE because it is not enabled in
ARCH=riscv defconfig by default.
-static uintptr_t load_pa __initdata;
+uintptr_t load_pa __initdata;

This is a weird issue. Any clue is appreciated.

Unfortunately, this is outside of my realm of expertise, as I am
unfamiliar with RISC-V at this level. Maybe Palmer has some ideas. I
would think that changing this variable to static would be fine given
that the symbol is only used in this translation unit but clearly not. I
have attempted to debug this in gdb but that does not really get me
anywhere: I cannot break on start_kernel() for whatever reason and the
kernel never gets to my breakpoint in setup_vm().

I did decide to through a BUG() around arch/riscv/mm/init.c to see
exactly which statement causes everything to hang. I landed on:

csr_write(CSR_SATP, PFN_DOWN(__pa_symbol(swapper_pg_dir)) | SATP_MODE);

in setup_vm_final(). No idea how that is relevant to this.

Some people in the ClangBuiltLinux bi-weekly meeting today pointed out
that since load_pa is assigned to the linker defined symbol _start,
there could be some optimization that goes awry here. I have keyed that

Thanks for the inspiration. Below patch fixes the hang issue, but I didn't
go through all necessary WRITE_ONCE convertions.

diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
index ae32f78207f0..fa9336a2583f 100644
--- a/arch/riscv/mm/init.c
+++ b/arch/riscv/mm/init.c
@@ -504,7 +504,7 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
va_kernel_xip_pa_offset = kernel_virt_addr - xiprom;
#else
- load_pa = (uintptr_t)(&_start);
+ WRITE_ONCE(load_pa, (uintptr_t)(&_start));
load_sz = (uintptr_t)(&_end) - load_pa;
#endif

I think I found the root cause, but I dunno whether this is clang bug or
we need WRITE_ONCE patch.
W/O WRITE_ONCE, the setup_vm() is compiled to

ffffffff8040472a <setup_vm>:
ffffffff8040472a: 1101 addi sp,sp,-32
ffffffff8040472c: ec06 sd ra,24(sp)
ffffffff8040472e: e822 sd s0,16(sp)
ffffffff80404730: e426 sd s1,8(sp)
ffffffff80404732: e04a sd s2,0(sp)
ffffffff80404734: 1000 addi s0,sp,32
ffffffff80404736: 892a mv s2,a0
ffffffff80404738: 001fd517 auipc a0,0x1fd
ffffffff8040473c: 8c850513 addi a0,a0,-1848 # ffffffff80601000 <load_pa>
ffffffff80404740: 4585 li a1,1
ffffffff80404742: 00b50023 sb a1,0(a0) // BUG!
ffffffff80404746: 001fd517 auipc a0,0x1fd
ffffffff8040474a: 8c250513 addi a0,a0,-1854 # ffffffff80601008 <load_sz>
ffffffff8040474e: 00b50023 sb a1,0(a0) // BUG!
ffffffff80404752: ffbfc517 auipc a0,0xffbfc
ffffffff80404756: 8ae50513 addi a0,a0,-1874 # ffffffff80000000 <_start>
ffffffff8040475a: 55fd li a1,-1
ffffffff8040475c: 02559613 slli a2,a1,0x25
ffffffff80404760: 83018593 addi a1,gp,-2000 # ffffffff80ca6428 <kernel_virt_addr>
ffffffff80404764: 618c ld a1,0(a1)
ffffffff80404766: 8e09 sub a2,a2,a0
...

It seems load_pa and load_sz are stored with 1, this is obviously not what the
code expected.


W/ WRITE_ONCE, the setup_vm() is compiled to:

ffffffff8040472a <setup_vm>:
ffffffff8040472a: 1101 addi sp,sp,-32
ffffffff8040472c: ec06 sd ra,24(sp)
ffffffff8040472e: e822 sd s0,16(sp)
ffffffff80404730: e426 sd s1,8(sp)
ffffffff80404732: e04a sd s2,0(sp)
ffffffff80404734: 1000 addi s0,sp,32
ffffffff80404736: 892a mv s2,a0
ffffffff80404738: 001fd597 auipc a1,0x1fd
ffffffff8040473c: 8c858593 addi a1,a1,-1848 # ffffffff80601000 <load_pa>
ffffffff80404740: ffbfc517 auipc a0,0xffbfc
ffffffff80404744: 8c050513 addi a0,a0,-1856 # ffffffff80000000 <_start>
ffffffff80404748: e188 sd a0,0(a1)
ffffffff8040474a: 001fd597 auipc a1,0x1fd
ffffffff8040474e: 8be58593 addi a1,a1,-1858 # ffffffff80601008 <load_sz>
ffffffff80404752: 4605 li a2,1
ffffffff80404754: 00c58023 sb a2,0(a1)
ffffffff80404758: 55fd li a1,-1
ffffffff8040475a: 02559613 slli a2,a1,0x25
ffffffff8040475e: 83018593 addi a1,gp,-2000 # ffffffff80ca6428 <kernel_virt_addr>
ffffffff80404762: 618c ld a1,0(a1)
ffffffff80404764G 8e09 sub a2,a2,a0
...

This is what the code expected.

This issue can also be solved by avoiding global vars: load_pa and load_sz.

Hi Alexandre,

IMHO, the load_pa and load_sz can be removed, I have one patch to remove
them, it works. I'm not sure whether will the two vars be used in your
future patches?


More or less, I want to move all the address conversion macros to inline functions as suggested by Christoph, and there I will need load_sz and load_pa. But I'm not sure it is really necessary.

Anyway, that would "solve" this issue, but I think we all agree the real problem should be solved at clang level.


Thanks in advance,
Jisheng


_______________________________________________
linux-riscv mailing list
linux-riscv@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/linux-riscv