arm64: module: rework module VA range selection

Currently, the modules region is 128M in size, which is a problem for
some large modules. Shanker reports [1] that the NVIDIA GPU driver alone
can consume 110M of module space in some configurations. We'd like to
make the modules region a full 2G such that we can always make use of a
2G range.

It's possible to build kernel images which are larger than 128M in some
configurations, such as when many debug options are selected and many
drivers are built in. In these configurations, we can't legitimately
select a base for a 128M module region, though we currently select a
value for which allocation will fail. It would be nicer to have a
diagnostic message in this case.

Similarly, in theory it's possible to build a kernel image which is
larger than 2G and which cannot support modules. While this isn't likely
to be the case for any realistic kernel deplyed in the field, it would
be nice if we could print a diagnostic in this case.

This patch reworks the module VA range selection to use a 2G range, and
improves handling of cases where we cannot select legitimate module
regions. We now attempt to select a 128M region and a 2G region:

* The 128M region is selected such that modules can use direct branches
  (with JUMP26/CALL26 relocations) to branch to kernel code and other
  modules, and so that modules can reference data and text (using PREL32
  relocations) anywhere in the kernel image and other modules.

  This region covers the entire kernel image (rather than just the text)
  to ensure that all PREL32 relocations are in range even when the
  kernel data section is absurdly large. Where we cannot allocate from
  this region, we'll fall back to the full 2G region.

* The 2G region is selected such that modules can use direct branches
  with PLTs to branch to kernel code and other modules, and so that
  modules can use reference data and text (with PREL32 relocations) in
  the kernel image and other modules.

  This region covers the entire kernel image, and the 128M region (if
  one is selected).

The two module regions are randomized independently while ensuring the
constraints described above.

[1] https://lore.kernel.org/linux-arm-kernel/159ceeab-09af-3174-5058-445bc8dcf85b@nvidia.com/Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Cc: Shanker Donthineni <sdonthineni@nvidia.com>
Cc: Will Deacon <will@kernel.org>
Tested-by: Shanker Donthineni <sdonthineni@nvidia.com>
Link: https://lore.kernel.org/r/20230530110328.2213762-7-mark.rutland@arm.comSigned-off-by: Catalin Marinas <catalin.marinas@arm.com>

Documentation/arm64/memory.rst

@@ -33,8 +33,8 @@ AArch64 Linux memory layout with 4KB pages + 4 levels (48-bit)::
   0000000000000000	0000ffffffffffff	 256TB		user
   ffff000000000000	ffff7fffffffffff	 128TB		kernel logical memory map
  [ffff600000000000	ffff7fffffffffff]	  32TB		[kasan shadow region]
-  ffff800000000000	ffff800007ffffff	 128MB		modules
-  ffff800008000000	fffffbffefffffff	 124TB		vmalloc
+  ffff800000000000	ffff80007fffffff	   2GB		modules
+  ffff800080000000	fffffbffefffffff	 124TB		vmalloc
   fffffbfff0000000	fffffbfffdffffff	 224MB		fixed mappings (top down)
   fffffbfffe000000	fffffbfffe7fffff	   8MB		[guard region]
   fffffbfffe800000	fffffbffff7fffff	  16MB		PCI I/O space
@@ -50,8 +50,8 @@ AArch64 Linux memory layout with 64KB pages + 3 levels (52-bit with HW support):
   0000000000000000	000fffffffffffff	   4PB		user
   fff0000000000000	ffff7fffffffffff	  ~4PB		kernel logical memory map
  [fffd800000000000	ffff7fffffffffff]	 512TB		[kasan shadow region]
-  ffff800000000000	ffff800007ffffff	 128MB		modules
-  ffff800008000000	fffffbffefffffff	 124TB		vmalloc
+  ffff800000000000	ffff80007fffffff	   2GB		modules
+  ffff800080000000	fffffbffefffffff	 124TB		vmalloc
   fffffbfff0000000	fffffbfffdffffff	 224MB		fixed mappings (top down)
   fffffbfffe000000	fffffbfffe7fffff	   8MB		[guard region]
   fffffbfffe800000	fffffbffff7fffff	  16MB		PCI I/O space

arch/arm64/include/asm/memory.h

@@ -46,7 +46,7 @@
 #define KIMAGE_VADDR		(MODULES_END)
 #define MODULES_END		(MODULES_VADDR + MODULES_VSIZE)
 #define MODULES_VADDR		(_PAGE_END(VA_BITS_MIN))
-#define MODULES_VSIZE		(SZ_128M)
+#define MODULES_VSIZE		(SZ_2G)
 #define VMEMMAP_START		(-(UL(1) << (VA_BITS - VMEMMAP_SHIFT)))
 #define VMEMMAP_END		(VMEMMAP_START + VMEMMAP_SIZE)
 #define PCI_IO_END		(VMEMMAP_START - SZ_8M)

arch/arm64/kernel/module.c

@@ -7,6 +7,8 @@
  * Author: Will Deacon <will.deacon@arm.com>
  */
 
+#define pr_fmt(fmt) "Modules: " fmt
+
 #include <linux/bitops.h>
 #include <linux/elf.h>
 #include <linux/ftrace.h>
@@ -24,72 +26,119 @@
 #include <asm/scs.h>
 #include <asm/sections.h>
 
-static u64 __ro_after_init module_alloc_base = (u64)_etext - MODULES_VSIZE;
+static u64 module_direct_base __ro_after_init = 0;
+static u64 module_plt_base __ro_after_init = 0;
 
-#ifdef CONFIG_RANDOMIZE_BASE
-static int __init kaslr_module_init(void)
+/*
+ * Choose a random page-aligned base address for a window of 'size' bytes which
+ * entirely contains the interval [start, end - 1].
+ */
+static u64 __init random_bounding_box(u64 size, u64 start, u64 end)
 {
-	u64 module_range;
-	u32 seed;
+	u64 max_pgoff, pgoff;
 
-	if (!kaslr_enabled())
+	if ((end - start) >= size)
 		return 0;
 
-	seed = get_random_u32();
+	max_pgoff = (size - (end - start)) / PAGE_SIZE;
+	pgoff = get_random_u32_inclusive(0, max_pgoff);
 
-	if (IS_ENABLED(CONFIG_RANDOMIZE_MODULE_REGION_FULL)) {
-		/*
-		 * Randomize the module region over a 2 GB window covering the
-		 * kernel. This reduces the risk of modules leaking information
-		 * about the address of the kernel itself, but results in
-		 * branches between modules and the core kernel that are
-		 * resolved via PLTs. (Branches between modules will be
-		 * resolved normally.)
-		 */
-		module_range = SZ_2G - (u64)(_end - _stext);
-		module_alloc_base = max((u64)_end - SZ_2G, (u64)MODULES_VADDR);
+	return start - pgoff * PAGE_SIZE;
+}
+
+/*
+ * Modules may directly reference data and text anywhere within the kernel
+ * image and other modules. References using PREL32 relocations have a +/-2G
+ * range, and so we need to ensure that the entire kernel image and all modules
+ * fall within a 2G window such that these are always within range.
+ *
+ * Modules may directly branch to functions and code within the kernel text,
+ * and to functions and code within other modules. These branches will use
+ * CALL26/JUMP26 relocations with a +/-128M range. Without PLTs, we must ensure
+ * that the entire kernel text and all module text falls within a 128M window
+ * such that these are always within range. With PLTs, we can expand this to a
+ * 2G window.
+ *
+ * We chose the 128M region to surround the entire kernel image (rather than
+ * just the text) as using the same bounds for the 128M and 2G regions ensures
+ * by construction that we never select a 128M region that is not a subset of
+ * the 2G region. For very large and unusual kernel configurations this means
+ * we may fall back to PLTs where they could have been avoided, but this keeps
+ * the logic significantly simpler.
+ */
+static int __init module_init_limits(void)
+{
+	u64 kernel_end = (u64)_end;
+	u64 kernel_start = (u64)_text;
+	u64 kernel_size = kernel_end - kernel_start;
+
+	/*
+	 * The default modules region is placed immediately below the kernel
+	 * image, and is large enough to use the full 2G relocation range.
+	 */
+	BUILD_BUG_ON(KIMAGE_VADDR != MODULES_END);
+	BUILD_BUG_ON(MODULES_VSIZE < SZ_2G);
+
+	if (!kaslr_enabled()) {
+		if (kernel_size < SZ_128M)
+			module_direct_base = kernel_end - SZ_128M;
+		if (kernel_size < SZ_2G)
+			module_plt_base = kernel_end - SZ_2G;
 	} else {
-		/*
-		 * Randomize the module region by setting module_alloc_base to
-		 * a PAGE_SIZE multiple in the range [_etext - MODULES_VSIZE,
-		 * _stext) . This guarantees that the resulting region still
-		 * covers [_stext, _etext], and that all relative branches can
-		 * be resolved without veneers unless this region is exhausted
-		 * and we fall back to a larger 2GB window in module_alloc()
-		 * when ARM64_MODULE_PLTS is enabled.
-		 */
-		module_range = MODULES_VSIZE - (u64)(_etext - _stext);
+		u64 min = kernel_start;
+		u64 max = kernel_end;
+
+		if (IS_ENABLED(CONFIG_RANDOMIZE_MODULE_REGION_FULL)) {
+			pr_info("2G module region forced by RANDOMIZE_MODULE_REGION_FULL\n");
+		} else {
+			module_direct_base = random_bounding_box(SZ_128M, min, max);
+			if (module_direct_base) {
+				min = module_direct_base;
+				max = module_direct_base + SZ_128M;
+			}
+		}
+
+		module_plt_base = random_bounding_box(SZ_2G, min, max);
 	}
 
-	/* use the lower 21 bits to randomize the base of the module region */
-	module_alloc_base += (module_range * (seed & ((1 << 21) - 1))) >> 21;
-	module_alloc_base &= PAGE_MASK;
+	pr_info("%llu pages in range for non-PLT usage",
+		module_direct_base ? (SZ_128M - kernel_size) / PAGE_SIZE : 0);
+	pr_info("%llu pages in range for PLT usage",
+		module_plt_base ? (SZ_2G - kernel_size) / PAGE_SIZE : 0);
 
 	return 0;
 }
-subsys_initcall(kaslr_module_init)
-#endif
+subsys_initcall(module_init_limits);
 
 void *module_alloc(unsigned long size)
 {
-	u64 module_alloc_end = module_alloc_base + MODULES_VSIZE;
-	void *p;
+	void *p = NULL;
 
 	/*
 	 * Where possible, prefer to allocate within direct branch range of the
-	 * kernel such that no PLTs are necessary. This may fail, so we pass
-	 * __GFP_NOWARN to silence the resulting warning.
+	 * kernel such that no PLTs are necessary.
 	 */
-	p = __vmalloc_node_range(size, MODULE_ALIGN, module_alloc_base,
-				module_alloc_end, GFP_KERNEL | __GFP_NOWARN,
-				PAGE_KERNEL, 0, NUMA_NO_NODE,
-				__builtin_return_address(0));
+	if (module_direct_base) {
+		p = __vmalloc_node_range(size, MODULE_ALIGN,
+					 module_direct_base,
+					 module_direct_base + SZ_128M,
+					 GFP_KERNEL | __GFP_NOWARN,
+					 PAGE_KERNEL, 0, NUMA_NO_NODE,
+					 __builtin_return_address(0));
+	}
+
+	if (!p && module_plt_base) {
+		p = __vmalloc_node_range(size, MODULE_ALIGN,
+					 module_plt_base,
+					 module_plt_base + SZ_2G,
+					 GFP_KERNEL | __GFP_NOWARN,
+					 PAGE_KERNEL, 0, NUMA_NO_NODE,
+					 __builtin_return_address(0));
+	}
 
 	if (!p) {
-		p = __vmalloc_node_range(size, MODULE_ALIGN, module_alloc_base,
-				module_alloc_base + SZ_2G, GFP_KERNEL,
-				PAGE_KERNEL, 0, NUMA_NO_NODE,
-				__builtin_return_address(0));
+		pr_warn_ratelimited("%s: unable to allocate memory\n",
+				    __func__);
 	}
 
 	if (p && (kasan_alloc_module_shadow(p, size, GFP_KERNEL) < 0)) {