RISC-V: Don't allow write+exec only page mapping request in mmap
authorYash Shah <yash.shah@sifive.com>
Tue, 16 Jun 2020 14:03:06 +0000 (19:33 +0530)
committerPalmer Dabbelt <palmerdabbelt@google.com>
Fri, 19 Jun 2020 00:28:53 +0000 (17:28 -0700)
As per the table 4.4 of version "20190608-Priv-MSU-Ratified" of the
RISC-V instruction set manual[0], the PTE permission bit combination of
"write+exec only" is reserved for future use. Hence, don't allow such
mapping request in mmap call.

An issue is been reported by David Abdurachmanov, that while running
stress-ng with "sysbadaddr" argument, RCU stalls are observed on RISC-V
specific kernel.

This issue arises when the stress-sysbadaddr request for pages with
"write+exec only" permission bits and then passes the address obtain
from this mmap call to various system call. For the riscv kernel, the
mmap call should fail for this particular combination of permission bits
since it's not valid.

[0]: http://dabbelt.com/~palmer/keep/riscv-isa-manual/riscv-privileged-20190608-1.pdf

Signed-off-by: Yash Shah <yash.shah@sifive.com>
Reported-by: David Abdurachmanov <david.abdurachmanov@gmail.com>
[Palmer: Refer to the latest ISA specification at the only link I could
find, and update the terminology.]
Signed-off-by: Palmer Dabbelt <palmerdabbelt@google.com>
arch/riscv/kernel/sys_riscv.c

index f3619f5..12f8a7f 100644 (file)
@@ -8,6 +8,7 @@
 #include <linux/syscalls.h>
 #include <asm/unistd.h>
 #include <asm/cacheflush.h>
+#include <asm-generic/mman-common.h>
 
 static long riscv_sys_mmap(unsigned long addr, unsigned long len,
                           unsigned long prot, unsigned long flags,
@@ -16,6 +17,11 @@ static long riscv_sys_mmap(unsigned long addr, unsigned long len,
 {
        if (unlikely(offset & (~PAGE_MASK >> page_shift_offset)))
                return -EINVAL;
+
+       if ((prot & PROT_WRITE) && (prot & PROT_EXEC))
+               if (unlikely(!(prot & PROT_READ)))
+                       return -EINVAL;
+
        return ksys_mmap_pgoff(addr, len, prot, flags, fd,
                               offset >> (PAGE_SHIFT - page_shift_offset));
 }