diff options
| author | Anup Patel <Anup.Patel@wdc.com> | 2019-06-28 13:36:21 -0700 |
|---|---|---|
| committer | Paul Walmsley <paul.walmsley@sifive.com> | 2019-07-09 09:08:04 -0700 |
| commit | 671f9a3e2e24cdeb2d2856abee7422f093e23e29 (patch) | |
| tree | 12c0271582f801f9dc2b0a2be59ee891e6c5f99f /arch/riscv/kernel/head.S | |
| parent | 2ebca1cbb4a5a31f8d89f22d9d410b432a97f7a7 (diff) | |
| download | lwn-671f9a3e2e24cdeb2d2856abee7422f093e23e29.tar.gz lwn-671f9a3e2e24cdeb2d2856abee7422f093e23e29.zip | |
RISC-V: Setup initial page tables in two stages
Currently, the setup_vm() does initial page table setup in one-shot
very early before enabling MMU. Due to this, the setup_vm() has to map
all possible kernel virtual addresses since it does not know size and
location of RAM. This means we have kernel mappings for non-existent
RAM and any buggy driver (or kernel) code doing out-of-bound access
to RAM will not fault and cause underterministic behaviour.
Further, the setup_vm() creates PMD mappings (i.e. 2M mappings) for
RV64 systems. This means for PAGE_OFFSET=0xffffffe000000000 (i.e.
MAXPHYSMEM_128GB=y), the setup_vm() will require 129 pages (i.e.
516 KB) of memory for initial page tables which is never freed. The
memory required for initial page tables will further increase if
we chose a lower value of PAGE_OFFSET (e.g. 0xffffff0000000000)
This patch implements two-staged initial page table setup, as follows:
1. Early (i.e. setup_vm()): This stage maps kernel image and DTB in
a early page table (i.e. early_pg_dir). The early_pg_dir will be used
only by boot HART so it can be freed as-part of init memory free-up.
2. Final (i.e. setup_vm_final()): This stage maps all possible RAM
banks in the final page table (i.e. swapper_pg_dir). The boot HART
will start using swapper_pg_dir at the end of setup_vm_final(). All
non-boot HARTs directly use the swapper_pg_dir created by boot HART.
We have following advantages with this new approach:
1. Kernel mappings for non-existent RAM don't exists anymore.
2. Memory consumed by initial page tables is now indpendent of the
chosen PAGE_OFFSET.
3. Memory consumed by initial page tables on RV64 system is 2 pages
(i.e. 8 KB) which has significantly reduced and these pages will be
freed as-part of the init memory free-up.
The patch also provides a foundation for implementing strict kernel
mappings where we protect kernel text and rodata using PTE permissions.
Suggested-by: Mike Rapoport <rppt@linux.ibm.com>
Signed-off-by: Anup Patel <anup.patel@wdc.com>
[paul.walmsley@sifive.com: updated to apply; fixed a checkpatch warning]
Signed-off-by: Paul Walmsley <paul.walmsley@sifive.com>
Diffstat (limited to 'arch/riscv/kernel/head.S')
| -rw-r--r-- | arch/riscv/kernel/head.S | 17 |
1 files changed, 9 insertions, 8 deletions
diff --git a/arch/riscv/kernel/head.S b/arch/riscv/kernel/head.S index 4e46f31072da..e368106f2228 100644 --- a/arch/riscv/kernel/head.S +++ b/arch/riscv/kernel/head.S @@ -55,7 +55,9 @@ clear_bss_done: /* Initialize page tables and relocate to virtual addresses */ la sp, init_thread_union + THREAD_SIZE + mv a0, s1 call setup_vm + la a0, early_pg_dir call relocate /* Restore C environment */ @@ -64,25 +66,23 @@ clear_bss_done: la sp, init_thread_union + THREAD_SIZE /* Start the kernel */ - mv a0, s1 call parse_dtb tail start_kernel relocate: /* Relocate return address */ li a1, PAGE_OFFSET - la a0, _start - sub a1, a1, a0 + la a2, _start + sub a1, a1, a2 add ra, ra, a1 /* Point stvec to virtual address of intruction after satp write */ - la a0, 1f - add a0, a0, a1 - csrw CSR_STVEC, a0 + la a2, 1f + add a2, a2, a1 + csrw CSR_STVEC, a2 /* Compute satp for kernel page tables, but don't load it yet */ - la a2, swapper_pg_dir - srl a2, a2, PAGE_SHIFT + srl a2, a0, PAGE_SHIFT li a1, SATP_MODE or a2, a2, a1 @@ -148,6 +148,7 @@ relocate: fence /* Enable virtual memory and relocate to virtual address */ + la a0, swapper_pg_dir call relocate tail smp_callin |