1 .. SPDX-License-Identifier: GPL-2.0
7 Kernel stacks on x86-64 bit
8 ===========================
10 Most of the text from Keith Owens, hacked by AK
12 x86_64 page size (PAGE_SIZE) is 4K.
14 Like all other architectures, x86_64 has a kernel stack for every
15 active thread. These thread stacks are THREAD_SIZE (2*PAGE_SIZE) big.
16 These stacks contain useful data as long as a thread is alive or a
17 zombie. While the thread is in user space the kernel stack is empty
18 except for the thread_info structure at the bottom.
20 In addition to the per thread stacks, there are specialized stacks
21 associated with each CPU. These stacks are only used while the kernel
22 is in control on that CPU; when a CPU returns to user space the
23 specialized stacks contain no useful data. The main CPU stacks are:
25 * Interrupt stack. IRQ_STACK_SIZE
27 Used for external hardware interrupts. If this is the first external
28 hardware interrupt (i.e. not a nested hardware interrupt) then the
29 kernel switches from the current task to the interrupt stack. Like
30 the split thread and interrupt stacks on i386, this gives more room
31 for kernel interrupt processing without having to increase the size
32 of every per thread stack.
34 The interrupt stack is also used when processing a softirq.
36 Switching to the kernel interrupt stack is done by software based on a
37 per CPU interrupt nest counter. This is needed because x86-64 "IST"
38 hardware stacks cannot nest without races.
40 x86_64 also has a feature which is not available on i386, the ability
41 to automatically switch to a new stack for designated events such as
42 double fault or NMI, which makes it easier to handle these unusual
43 events on x86_64. This feature is called the Interrupt Stack Table
44 (IST). There can be up to 7 IST entries per CPU. The IST code is an
45 index into the Task State Segment (TSS). The IST entries in the TSS
46 point to dedicated stacks; each stack can be a different size.
48 An IST is selected by a non-zero value in the IST field of an
49 interrupt-gate descriptor. When an interrupt occurs and the hardware
50 loads such a descriptor, the hardware automatically sets the new stack
51 pointer based on the IST value, then invokes the interrupt handler. If
52 the interrupt came from user mode, then the interrupt handler prologue
53 will switch back to the per-thread stack. If software wants to allow
54 nested IST interrupts then the handler must adjust the IST values on
55 entry to and exit from the interrupt handler. (This is occasionally
56 done, e.g. for debug exceptions.)
58 Events with different IST codes (i.e. with different stacks) can be
59 nested. For example, a debug interrupt can safely be interrupted by an
60 NMI. arch/x86_64/kernel/entry.S::paranoidentry adjusts the stack
61 pointers on entry to and exit from all IST events, in theory allowing
62 IST events with the same code to be nested. However in most cases, the
63 stack size allocated to an IST assumes no nesting for the same code.
64 If that assumption is ever broken then the stacks will become corrupt.
66 The currently assigned IST stacks are:
68 * ESTACK_DF. EXCEPTION_STKSZ (PAGE_SIZE).
70 Used for interrupt 8 - Double Fault Exception (#DF).
72 Invoked when handling one exception causes another exception. Happens
73 when the kernel is very confused (e.g. kernel stack pointer corrupt).
74 Using a separate stack allows the kernel to recover from it well enough
75 in many cases to still output an oops.
77 * ESTACK_NMI. EXCEPTION_STKSZ (PAGE_SIZE).
79 Used for non-maskable interrupts (NMI).
81 NMI can be delivered at any time, including when the kernel is in the
82 middle of switching stacks. Using IST for NMI events avoids making
83 assumptions about the previous state of the kernel stack.
85 * ESTACK_DB. EXCEPTION_STKSZ (PAGE_SIZE).
87 Used for hardware debug interrupts (interrupt 1) and for software
88 debug interrupts (INT3).
90 When debugging a kernel, debug interrupts (both hardware and
91 software) can occur at any time. Using IST for these interrupts
92 avoids making assumptions about the previous state of the kernel
95 To handle nested #DB correctly there exist two instances of DB stacks. On
96 #DB entry the IST stackpointer for #DB is switched to the second instance
97 so a nested #DB starts from a clean stack. The nested #DB switches
98 the IST stackpointer to a guard hole to catch triple nesting.
100 * ESTACK_MCE. EXCEPTION_STKSZ (PAGE_SIZE).
102 Used for interrupt 18 - Machine Check Exception (#MC).
104 MCE can be delivered at any time, including when the kernel is in the
105 middle of switching stacks. Using IST for MCE events avoids making
106 assumptions about the previous state of the kernel stack.
108 For more details see the Intel IA32 or AMD AMD64 architecture manuals.
111 Printing backtraces on x86
112 ==========================
114 The question about the '?' preceding function names in an x86 stacktrace
115 keeps popping up, here's an indepth explanation. It helps if the reader
116 stares at print_context_stack() and the whole machinery in and around
117 arch/x86/kernel/dumpstack.c.
119 Adapted from Ingo's mail, Message-ID: <20150521101614.GA10889@gmail.com>:
121 We always scan the full kernel stack for return addresses stored on
122 the kernel stack(s) [1]_, from stack top to stack bottom, and print out
123 anything that 'looks like' a kernel text address.
125 If it fits into the frame pointer chain, we print it without a question
126 mark, knowing that it's part of the real backtrace.
128 If the address does not fit into our expected frame pointer chain we
129 still print it, but we print a '?'. It can mean two things:
131 - either the address is not part of the call chain: it's just stale
132 values on the kernel stack, from earlier function calls. This is
135 - or it is part of the call chain, but the frame pointer was not set
136 up properly within the function, so we don't recognize it.
138 This way we will always print out the real call chain (plus a few more
139 entries), regardless of whether the frame pointer was set up correctly
140 or not - but in most cases we'll get the call chain right as well. The
141 entries printed are strictly in stack order, so you can deduce more
142 information from that as well.
144 The most important property of this method is that we _never_ lose
145 information: we always strive to print _all_ addresses on the stack(s)
146 that look like kernel text addresses, so if debug information is wrong,
147 we still print out the real call chain as well - just with more question
150 .. [1] For things like IRQ and IST stacks, we also scan those stacks, in
151 the right order, and try to cross from one stack into another
152 reconstructing the call chain. This works most of the time.