riscv: introduce interfaces to patch kernel code
authorZong Li <zong.li@sifive.com>
Mon, 9 Mar 2020 16:55:43 +0000 (00:55 +0800)
committerPalmer Dabbelt <palmerdabbelt@google.com>
Thu, 26 Mar 2020 16:24:52 +0000 (09:24 -0700)
On strict kernel memory permission, we couldn't patch code without
writable permission. Preserve two holes in fixmap area, so we can map
the kernel code temporarily to fixmap area, then patch the instructions.

We need two pages here because we support the compressed instruction, so
the instruction might be align to 2 bytes. When patching the 32-bit
length instruction which is 2 bytes alignment, it will across two pages.

Introduce two interfaces to patch kernel code:
riscv_patch_text_nosync:
 - patch code without synchronization, it's caller's responsibility to
   synchronize all CPUs if needed.
riscv_patch_text:
 - patch code and always synchronize with stop_machine()

Signed-off-by: Zong Li <zong.li@sifive.com>
Signed-off-by: Palmer Dabbelt <palmerdabbelt@google.com>
arch/riscv/include/asm/fixmap.h
arch/riscv/include/asm/patch.h [new file with mode: 0644]
arch/riscv/kernel/Makefile
arch/riscv/kernel/patch.c [new file with mode: 0644]

index 42d2c42f3cc946d2b3b9eabe6758eb9fcd64e4ea..2368d49eb4efebe50bbceb0c2b165b685cbd1134 100644 (file)
@@ -27,6 +27,8 @@ enum fixed_addresses {
        FIX_FDT = FIX_FDT_END + FIX_FDT_SIZE / PAGE_SIZE - 1,
        FIX_PTE,
        FIX_PMD,
+       FIX_TEXT_POKE1,
+       FIX_TEXT_POKE0,
        FIX_EARLYCON_MEM_BASE,
        __end_of_fixed_addresses
 };
diff --git a/arch/riscv/include/asm/patch.h b/arch/riscv/include/asm/patch.h
new file mode 100644 (file)
index 0000000..b5918a6
--- /dev/null
@@ -0,0 +1,12 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2020 SiFive
+ */
+
+#ifndef _ASM_RISCV_PATCH_H
+#define _ASM_RISCV_PATCH_H
+
+int riscv_patch_text_nosync(void *addr, const void *insns, size_t len);
+int riscv_patch_text(void *addr, u32 insn);
+
+#endif /* _ASM_RISCV_PATCH_H */
index f40205cb9a2277c1583adfe10bd973c6cec261d9..d189bd3d8501201f75899118076e1d8a361e7647 100644 (file)
@@ -4,7 +4,8 @@
 #
 
 ifdef CONFIG_FTRACE
-CFLAGS_REMOVE_ftrace.o = -pg
+CFLAGS_REMOVE_ftrace.o = -pg
+CFLAGS_REMOVE_patch.o  = -pg
 endif
 
 extra-y += head.o
@@ -26,6 +27,7 @@ obj-y += traps.o
 obj-y  += riscv_ksyms.o
 obj-y  += stacktrace.o
 obj-y  += cacheinfo.o
+obj-y  += patch.o
 obj-$(CONFIG_MMU) += vdso.o vdso/
 
 obj-$(CONFIG_RISCV_M_MODE)     += clint.o
diff --git a/arch/riscv/kernel/patch.c b/arch/riscv/kernel/patch.c
new file mode 100644 (file)
index 0000000..8a4fc65
--- /dev/null
@@ -0,0 +1,120 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (C) 2020 SiFive
+ */
+
+#include <linux/spinlock.h>
+#include <linux/mm.h>
+#include <linux/uaccess.h>
+#include <linux/stop_machine.h>
+#include <asm/kprobes.h>
+#include <asm/cacheflush.h>
+#include <asm/fixmap.h>
+
+struct riscv_insn_patch {
+       void *addr;
+       u32 insn;
+       atomic_t cpu_count;
+};
+
+#ifdef CONFIG_MMU
+static DEFINE_RAW_SPINLOCK(patch_lock);
+
+static void __kprobes *patch_map(void *addr, int fixmap)
+{
+       uintptr_t uintaddr = (uintptr_t) addr;
+       struct page *page;
+
+       if (core_kernel_text(uintaddr))
+               page = phys_to_page(__pa_symbol(addr));
+       else if (IS_ENABLED(CONFIG_STRICT_MODULE_RWX))
+               page = vmalloc_to_page(addr);
+       else
+               return addr;
+
+       BUG_ON(!page);
+
+       return (void *)set_fixmap_offset(fixmap, page_to_phys(page) +
+                                        (uintaddr & ~PAGE_MASK));
+}
+
+static void __kprobes patch_unmap(int fixmap)
+{
+       clear_fixmap(fixmap);
+}
+
+static int __kprobes riscv_insn_write(void *addr, const void *insn, size_t len)
+{
+       void *waddr = addr;
+       bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
+       unsigned long flags = 0;
+       int ret;
+
+       raw_spin_lock_irqsave(&patch_lock, flags);
+
+       if (across_pages)
+               patch_map(addr + len, FIX_TEXT_POKE1);
+
+       waddr = patch_map(addr, FIX_TEXT_POKE0);
+
+       ret = probe_kernel_write(waddr, insn, len);
+
+       patch_unmap(FIX_TEXT_POKE0);
+
+       if (across_pages)
+               patch_unmap(FIX_TEXT_POKE1);
+
+       raw_spin_unlock_irqrestore(&patch_lock, flags);
+
+       return ret;
+}
+#else
+static int __kprobes riscv_insn_write(void *addr, const void *insn, size_t len)
+{
+       return probe_kernel_write(addr, insn, len);
+}
+#endif /* CONFIG_MMU */
+
+int __kprobes riscv_patch_text_nosync(void *addr, const void *insns, size_t len)
+{
+       u32 *tp = addr;
+       int ret;
+
+       ret = riscv_insn_write(tp, insns, len);
+
+       if (!ret)
+               flush_icache_range((uintptr_t) tp, (uintptr_t) tp + len);
+
+       return ret;
+}
+
+static int __kprobes riscv_patch_text_cb(void *data)
+{
+       struct riscv_insn_patch *patch = data;
+       int ret = 0;
+
+       if (atomic_inc_return(&patch->cpu_count) == 1) {
+               ret =
+                   riscv_patch_text_nosync(patch->addr, &patch->insn,
+                                           GET_INSN_LENGTH(patch->insn));
+               atomic_inc(&patch->cpu_count);
+       } else {
+               while (atomic_read(&patch->cpu_count) <= num_online_cpus())
+                       cpu_relax();
+               smp_mb();
+       }
+
+       return ret;
+}
+
+int __kprobes riscv_patch_text(void *addr, u32 insn)
+{
+       struct riscv_insn_patch patch = {
+               .addr = addr,
+               .insn = insn,
+               .cpu_count = ATOMIC_INIT(0),
+       };
+
+       return stop_machine_cpuslocked(riscv_patch_text_cb,
+                                      &patch, cpu_online_mask);
+}