1.. SPDX-License-Identifier: GPL-2.0
2
3=====================================
4Virtually Mapped Kernel Stack Support
5=====================================
6
7:Author: Shuah Khan <skhan@linuxfoundation.org>
8
9.. contents:: :local:
10
11Overview
12--------
13
14This is a compilation of information from the code and original patch
15series that introduced the `Virtually Mapped Kernel Stacks feature
16<https://lwn.net/Articles/694348/>`
17
18Introduction
19------------
20
21Kernel stack overflows are often hard to debug and make the kernel
22susceptible to exploits. Problems could show up at a later time making
23it difficult to isolate and root-cause.
24
25Virtually-mapped kernel stacks with guard pages causes kernel stack
26overflows to be caught immediately rather than causing difficult to
27diagnose corruptions.
28
29HAVE_ARCH_VMAP_STACK and VMAP_STACK configuration options enable
30support for virtually mapped stacks with guard pages. This feature
31causes reliable faults when the stack overflows. The usability of
32the stack trace after overflow and response to the overflow itself
33is architecture dependent.
34
35.. note::
36        As of this writing, arm64, powerpc, riscv, s390, um, and x86 have
37        support for VMAP_STACK.
38
39HAVE_ARCH_VMAP_STACK
40--------------------
41
42Architectures that can support Virtually Mapped Kernel Stacks should
43enable this bool configuration option. The requirements are:
44
45- vmalloc space must be large enough to hold many kernel stacks. This
46  may rule out many 32-bit architectures.
47- Stacks in vmalloc space need to work reliably.  For example, if
48  vmap page tables are created on demand, either this mechanism
49  needs to work while the stack points to a virtual address with
50  unpopulated page tables or arch code (switch_to() and switch_mm(),
51  most likely) needs to ensure that the stack's page table entries
52  are populated before running on a possibly unpopulated stack.
53- If the stack overflows into a guard page, something reasonable
54  should happen. The definition of "reasonable" is flexible, but
55  instantly rebooting without logging anything would be unfriendly.
56
57VMAP_STACK
58----------
59
60VMAP_STACK bool configuration option when enabled allocates virtually
61mapped task stacks. This option depends on HAVE_ARCH_VMAP_STACK.
62
63- Enable this if you want the use virtually-mapped kernel stacks
64  with guard pages. This causes kernel stack overflows to be caught
65  immediately rather than causing difficult-to-diagnose corruption.
66
67.. note::
68
69        Using this feature with KASAN requires architecture support
70        for backing virtual mappings with real shadow memory, and
71        KASAN_VMALLOC must be enabled.
72
73.. note::
74
75        VMAP_STACK is enabled, it is not possible to run DMA on stack
76        allocated data.
77
78Kernel configuration options and dependencies keep changing. Refer to
79the latest code base:
80
81`Kconfig <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/Kconfig>`
82
83Allocation
84-----------
85
86When a new kernel thread is created, thread stack is allocated from
87virtually contiguous memory pages from the page level allocator. These
88pages are mapped into contiguous kernel virtual space with PAGE_KERNEL
89protections.
90
91alloc_thread_stack_node() calls __vmalloc_node_range() to allocate stack
92with PAGE_KERNEL protections.
93
94- Allocated stacks are cached and later reused by new threads, so memcg
95  accounting is performed manually on assigning/releasing stacks to tasks.
96  Hence, __vmalloc_node_range is called without __GFP_ACCOUNT.
97- vm_struct is cached to be able to find when thread free is initiated
98  in interrupt context. free_thread_stack() can be called in interrupt
99  context.
100- On arm64, all VMAP's stacks need to have the same alignment to ensure
101  that VMAP'd stack overflow detection works correctly. Arch specific
102  vmap stack allocator takes care of this detail.
103- This does not address interrupt stacks - according to the original patch
104
105Thread stack allocation is initiated from clone(), fork(), vfork(),
106kernel_thread() via kernel_clone(). Leaving a few hints for searching
107the code base to understand when and how thread stack is allocated.
108
109Bulk of the code is in:
110`kernel/fork.c <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/fork.c>`.
111
112stack_vm_area pointer in task_struct keeps track of the virtually allocated
113stack and a non-null stack_vm_area pointer serves as a indication that the
114virtually mapped kernel stacks are enabled.
115
116::
117
118        struct vm_struct *stack_vm_area;
119
120Stack overflow handling
121-----------------------
122
123Leading and trailing guard pages help detect stack overflows. When stack
124overflows into the guard pages, handlers have to be careful not overflow
125the stack again. When handlers are called, it is likely that very little
126stack space is left.
127
128On x86, this is done by handling the page fault indicating the kernel
129stack overflow on the double-fault stack.
130
131Testing VMAP allocation with guard pages
132----------------------------------------
133
134How do we ensure that VMAP_STACK is actually allocating with a leading
135and trailing guard page? The following lkdtm tests can help detect any
136regressions.
137
138::
139
140        void lkdtm_STACK_GUARD_PAGE_LEADING()
141        void lkdtm_STACK_GUARD_PAGE_TRAILING()
142
143Conclusions
144-----------
145
146- A percpu cache of vmalloced stacks appears to be a bit faster than a
147  high-order stack allocation, at least when the cache hits.
148- THREAD_INFO_IN_TASK gets rid of arch-specific thread_info entirely and
149  simply embed the thread_info (containing only flags) and 'int cpu' into
150  task_struct.
151- The thread stack can be free'ed as soon as the task is dead (without
152  waiting for RCU) and then, if vmapped stacks are in use, cache the
153  entire stack for reuse on the same cpu.
154