Linux: Implement interfaces for memory protection keys

This adds system call wrappers for pkey_alloc, pkey_free, pkey_mprotect,
and x86-64 implementations of pkey_get and pkey_set, which abstract over
the PKRU CPU register and hide the actual number of memory protection
keys supported by the CPU.  pkey_mprotect with a -1 key is implemented
using mprotect, so it will work even if the kernel does not support the
pkey_mprotect system call.

The system call wrapers use unsigned int instead of unsigned long for
parameters, so that no special treatment for x32 is needed.  The flags
argument is currently unused, and the access rights bit mask is limited
to two bits by the current PKRU register layout anyway.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>

1 files changed, 232 insertions, 0 deletions

diff --git a/manual/memory.texi b/manual/memory.texi
index 1b431bf5da..b95f6aa1b9 100644
--- a/manual/memory.texi
+++ b/manual/memory.texi
@@ -3171,6 +3171,238 @@ process memory, no matter how it was allocated.  However, portable use
 of the function requires that it is only used with memory regions
 returned by @code{mmap} or @code{mmap64}.
 
+@subsection Memory Protection Keys
+
+@cindex memory protection key
+@cindex protection key
+@cindex MPK
+On some systems, further restrictions can be added to specific pages
+using @dfn{memory protection keys}.  These restrictions work as follows:
+
+@itemize @bullet
+@item
+All memory pages are associated with a protection key.  The default
+protection key does not cause any additional protections to be applied
+during memory accesses.  New keys can be allocated with the
+@code{pkey_alloc} function, and applied to pages using
+@code{pkey_mprotect}.
+
+@item
+Each thread has a set of separate access right restriction for each
+protection key.  These access rights can be manipulated using the
+@code{pkey_set} and @code{pkey_get} functions.
+
+@item
+During a memory access, the system obtains the protection key for the
+accessed page and uses that to determine the applicable access rights,
+as configured for the current thread.  If the access is restricted, a
+segmentation fault is the result ((@pxref{Program Error Signals}).
+These checks happen in addition to the @code{PROT_}* protection flags
+set by @code{mprotect} or @code{pkey_mprotect}.
+@end itemize
+
+New threads and subprocesses inherit the access rights of the current
+thread.  If a protection key is allocated subsequently, existing threads
+(except the current) will use an unspecified system default for the
+access rights associated with newly allocated keys.
+
+Upon entering a signal handler, the system resets the access rights of
+the current thread so that pages with the default key can be accessed,
+but the access rights for other protection keys are unspecified.
+
+Applications are expected to allocate a key once using
+@code{pkey_alloc}, and apply the key to memory regions which need
+special protection with @code{pkey_mprotect}:
+
+@smallexample
+  int key = pkey_alloc (0, PKEY_DISABLE_ACCESS);
+  if (key < 0)
+    /* Perform error checking, including fallback for lack of support.  */
+    ...;
+
+  /* Apply the key to a special memory region used to store critical
+     data.  */
+  if (pkey_mprotect (region, region_length,
+                     PROT_READ | PROT_WRITE, key) < 0)
+    ...; /* Perform error checking (generally fatal).  */
+@end smallexample
+
+If the key allocation fails due to lack of support for memory protection
+keys, the @code{pkey_mprotect} call can usually be skipped.  In this
+case, the region will not be protected by default.  It is also possible
+to call @code{pkey_mprotect} with a key value of @math{-1}, in which
+case it will behave in the same way as @code{mprotect}.
+
+After key allocation assignment to memory pages, @code{pkey_set} can be
+used to temporarily acquire access to the memory region and relinquish
+it again:
+
+@smallexample
+  if (key >= 0 && pkey_set (key, 0) < 0)
+    ...; /* Perform error checking (generally fatal).  */
+  /* At this point, the current thread has read-write access to the
+     memory region.  */
+  ...
+  /* Revoke access again.  */
+  if (key >= 0 && pkey_set (key, PKEY_DISABLE_ACCESS) < 0)
+    ...; /* Perform error checking (generally fatal).  */
+@end smallexample
+
+In this example, a negative key value indicates that no key had been
+allocated, which means that the system lacks support for memory
+protection keys and it is not necessary to change the the access rights
+of the current thread (because it always has access).
+
+Compared to using @code{mprotect} to change the page protection flags,
+this approach has two advantages: It is thread-safe in the sense that
+the access rights are only changed for the current thread, so another
+thread which changes its own access rights concurrently to gain access
+to the mapping will not suddenly see its access rights revoked.  And
+@code{pkey_set} typically does not involve a call into the kernel and a
+context switch, so it is more efficient.
+
+@deftypefun int pkey_alloc (unsigned int @var{flags}, unsigned int @var{restrictions})
+@standards{Linux, sys/mman.h}
+@safety{@prelim{}@mtsafe{}@assafe{}@acunsafe{@acucorrupt{}}}
+Allocate a new protection key.  The @var{flags} argument is reserved and
+must be zero.  The @var{restrictions} argument specifies access rights
+which are applied to the current thread (as if with @code{pkey_set}
+below).  Access rights of other threads are not changed.
+
+The function returns the new protection key, a non-negative number, or
+@math{-1} on error.
+
+The following @code{errno} error conditions are defined for this
+function:
+
+@table @code
+@item ENOSYS
+The system does not implement memory protection keys.
+
+@item EINVAL
+The @var{flags} argument is not zero.
+
+The @var{restrictions} argument is invalid.
+
+The system does not implement memory protection keys or runs in a mode
+in which memory protection keys are disabled.
+
+@item ENOSPC
+All available protection keys already have been allocated.
+@end table
+@end deftypefun
+
+@deftypefun int pkey_free (int @var{key})
+@standards{Linux, sys/mman.h}
+@safety{@prelim{}@mtsafe{}@assafe{}@acsafe{}}
+Deallocate the protection key, so that it can be reused by
+@code{pkey_alloc}.
+
+Calling this function does not change the access rights of the freed
+protection key.  The calling thread and other threads may retain access
+to it, even if it is subsequently allocated again.  For this reason, it
+is not recommended to call the @code{pkey_free} function.
+
+@table @code
+@item ENOSYS
+The system does not implement memory protection keys.
+
+@item EINVAL
+The @var{key} argument is not a valid protection key.
+@end table
+@end deftypefun
+
+@deftypefun int pkey_mprotect (void *@var{address}, size_t @var{length}, int @var{protection}, int @var{key})
+@standards{Linux, sys/mman.h}
+@safety{@prelim{}@mtsafe{}@assafe{}@acsafe{}}
+Similar to @code{mprotect}, but also set the memory protection key for
+the memory region to @code{key}.
+
+Some systems use memory protection keys to emulate certain combinations
+of @var{protection} flags.  Under such circumstances, specifying an
+explicit protection key may behave as if additional flags have been
+specified in @var{protection}, even though this does not happen with the
+default protection key.  For example, some systems can support
+@code{PROT_EXEC}-only mappings only with a default protection key, and
+memory with a key which was allocated using @code{pkey_alloc} will still
+be readable if @code{PROT_EXEC} is specified without @code{PROT_READ}.
+
+If @var{key} is @math{-1}, the default protection key is applied to the
+mapping, just as if @code{mprotect} had been called.
+
+The @code{pkey_mprotect} function returns @math{0} on success and
+@math{-1} on failure.  The same @code{errno} error conditions as for
+@code{mprotect} are defined for this function, with the following
+addition:
+
+@table @code
+@item EINVAL
+The @var{key} argument is not @math{-1} or a valid memory protection
+key allocated using @code{pkey_alloc}.
+
+@item ENOSYS
+The system does not implement memory protection keys, and @var{key} is
+not @math{-1}.
+@end table
+@end deftypefun
+
+@deftypefun int pkey_set (int @var{key}, unsigned int @var{rights})
+@standards{Linux, sys/mman.h}
+@safety{@prelim{}@mtsafe{}@assafe{}@acsafe{}}
+Change the access rights of the current thread for memory pages with the
+protection key @var{key} to @var{rights}.  If @var{rights} is zero, no
+additional access restrictions on top of the page protection flags are
+applied.  Otherwise, @var{rights} is a combination of the following
+flags:
+
+@vtable @code
+@item PKEY_DISABLE_WRITE
+@standards{Linux, sys/mman.h}
+Subsequent attempts to write to memory with the specified protection
+key will fault.
+
+@item PKEY_DISABLE_ACCESS
+@standards{Linux, sys/mman.h}
+Subsequent attempts to write to or read from memory with the specified
+protection key will fault.
+@end vtable
+
+Operations not specified as flags are not restricted.  In particular,
+this means that the memory region will remain executable if it was
+mapped with the @code{PROT_EXEC} protection flag and
+@code{PKEY_DISABLE_ACCESS} has been specified.
+
+Calling the @code{pkey_set} function with a protection key which was not
+allocated by @code{pkey_alloc} results in undefined behavior.  This
+means that calling this function on systems which do not support memory
+protection keys is undefined.
+
+The @code{pkey_set} function returns @math{0} on success and @math{-1}
+on failure.
+
+The following @code{errno} error conditions are defined for this
+function:
+
+@table @code
+@item EINVAL
+The system does not support the access rights restrictions expressed in
+the @var{rights} argument.
+@end table
+@end deftypefun
+
+@deftypefun int pkey_get (int @var{key})
+@standards{Linux, sys/mman.h}
+@safety{@prelim{}@mtsafe{}@assafe{}@acsafe{}}
+Return the access rights of the current thread for memory pages with
+protection key @var{key}.  The return value is zero or a combination of
+the @code{PKEY_DISABLE_}* flags; see the @code{pkey_set} function.
+
+Calling the @code{pkey_get} function with a protection key which was not
+allocated by @code{pkey_alloc} results in undefined behavior.  This
+means that calling this function on systems which do not support memory
+protection keys is undefined.
+@end deftypefun
+
 @node Locking Pages
 @section Locking Pages
 @cindex locking pages