powerpc/mm/64s/hash: Reallocate context ids on fork

commit ca72d883 upstream.

When using the Hash Page Table (HPT) MMU, userspace memory mappings
are managed at two levels. Firstly in the Linux page tables, much like
other architectures, and secondly in the SLB (Segment Lookaside
Buffer) and HPT. It's the SLB and HPT that are actually used by the
hardware to do translations.

As part of the series adding support for 4PB user virtual address
space using the hash MMU, we added support for allocating multiple
"context ids" per process, one for each 512TB chunk of address space.
These are tracked in an array called extended_id in the mm_context_t
of a process that has done a mapping above 512TB.

If such a process forks (ie. clone(2) without CLONE_VM set) it's mm is
copied, including the mm_context_t, and then init_new_context() is
called to reinitialise parts of the mm_context_t as appropriate to
separate the address spaces of the two processes.

The key step in ensuring the two processes have separate address
spaces is to allocate a new context id for the process, this is done
at the beginning of hash__init_new_context(). If we didn't allocate a
new context id then the two processes would share mappings as far as
the SLB and HPT are concerned, even though their Linux page tables
would be separate.

For mappings above 512TB, which use the extended_id array, we
neglected to allocate new context ids on fork, meaning the parent and
child use the same ids and therefore share those mappings even though
they're supposed to be separate. This can lead to the parent seeing
writes done by the child, which is essentially memory corruption.

There is an additional exposure which is that if the child process
exits, all its context ids are freed, including the context ids that
are still in use by the parent for mappings above 512TB. One or more
of those ids can then be reallocated to a third process, that process
can then read/write to the parent's mappings above 512TB. Additionally
if the freed id is used for the third process's primary context id,
then the parent is able to read/write to the third process's mappings
*below* 512TB.

All of these are fundamental failures to enforce separation between
processes. The only mitigating factor is that the bug only occurs if a
process creates mappings above 512TB, and most applications still do
not create such mappings.

Only machines using the hash page table MMU are affected, eg. PowerPC
970 (G5), PA6T, Power5/6/7/8/9. By default Power9 bare metal machines
(powernv) use the Radix MMU and are not affected, unless the machine
has been explicitly booted in HPT mode (using disable_radix on the
kernel command line). KVM guests on Power9 may be affected if the host
or guest is configured to use the HPT MMU. LPARs under PowerVM on
Power9 are affected as they always use the HPT MMU. Kernels built with
PAGE_SIZE=4K are not affected.

The fix is relatively simple, we need to reallocate context ids for
all extended mappings on fork.

Fixes: f384796c ("powerpc/mm: Add support for handling > 512TB address in SLB miss")
Cc: stable@vger.kernel.org # v4.17+
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

arch/powerpc/mm/mmu_context_book3s64.c

@@ -55,14 +55,48 @@ EXPORT_SYMBOL_GPL(hash__alloc_context_id);
 
 void slb_setup_new_exec(void);
 
+static int realloc_context_ids(mm_context_t *ctx)
+{
+	int i, id;
+
+	/*
+	 * id 0 (aka. ctx->id) is special, we always allocate a new one, even if
+	 * there wasn't one allocated previously (which happens in the exec
+	 * case where ctx is newly allocated).
+	 *
+	 * We have to be a bit careful here. We must keep the existing ids in
+	 * the array, so that we can test if they're non-zero to decide if we
+	 * need to allocate a new one. However in case of error we must free the
+	 * ids we've allocated but *not* any of the existing ones (or risk a
+	 * UAF). That's why we decrement i at the start of the error handling
+	 * loop, to skip the id that we just tested but couldn't reallocate.
+	 */
+	for (i = 0; i < ARRAY_SIZE(ctx->extended_id); i++) {
+		if (i == 0 || ctx->extended_id[i]) {
+			id = hash__alloc_context_id();
+			if (id < 0)
+				goto error;
+
+			ctx->extended_id[i] = id;
+		}
+	}
+
+	/* The caller expects us to return id */
+	return ctx->id;
+
+error:
+	for (i--; i >= 0; i--) {
+		if (ctx->extended_id[i])
+			ida_free(&mmu_context_ida, ctx->extended_id[i]);
+	}
+
+	return id;
+}
+
 static int hash__init_new_context(struct mm_struct *mm)
 {
 	int index;
 
-	index = hash__alloc_context_id();
-	if (index < 0)
-		return index;
-
 	/*
 	 * The old code would re-promote on fork, we don't do that when using
 	 * slices as it could cause problem promoting slices that have been
@@ -80,6 +114,10 @@ static int hash__init_new_context(struct mm_struct *mm)
 	if (mm->context.id == 0)
 		slice_init_new_context_exec(mm);
 
+	index = realloc_context_ids(&mm->context);
+	if (index < 0)
+		return index;
+
 	subpage_prot_init_new_context(mm);
 
 	pkey_mm_init(mm);