1.. SPDX-License-Identifier: GPL-2.0
2
3==========================
4Page Table Isolation (PTI)
5==========================
6
7Overview
8========
9
10Page Table Isolation (pti, previously known as KAISER [1]_) is a
11countermeasure against attacks on the shared user/kernel address
12space such as the "Meltdown" approach [2]_.
13
14To mitigate this class of attacks, we create an independent set of
15page tables for use only when running userspace applications.  When
16the kernel is entered via syscalls, interrupts or exceptions, the
17page tables are switched to the full "kernel" copy.  When the system
18switches back to user mode, the user copy is used again.
19
20The userspace page tables contain only a minimal amount of kernel
21data: only what is needed to enter/exit the kernel such as the
22entry/exit functions themselves and the interrupt descriptor table
23(IDT).  There are a few strictly unnecessary things that get mapped
24such as the first C function when entering an interrupt (see
25comments in pti.c).
26
27This approach helps to ensure that side-channel attacks leveraging
28the paging structures do not function when PTI is enabled.  It can be
29enabled by setting CONFIG_PAGE_TABLE_ISOLATION=y at compile time.
30Once enabled at compile-time, it can be disabled at boot with the
31'nopti' or 'pti=' kernel parameters (see kernel-parameters.txt).
32
33Page Table Management
34=====================
35
36When PTI is enabled, the kernel manages two sets of page tables.
37The first set is very similar to the single set which is present in
38kernels without PTI.  This includes a complete mapping of userspace
39that the kernel can use for things like copy_to_user().
40
41Although _complete_, the user portion of the kernel page tables is
42crippled by setting the NX bit in the top level.  This ensures
43that any missed kernel->user CR3 switch will immediately crash
44userspace upon executing its first instruction.
45
46The userspace page tables map only the kernel data needed to enter
47and exit the kernel.  This data is entirely contained in the 'struct
48cpu_entry_area' structure which is placed in the fixmap which gives
49each CPU's copy of the area a compile-time-fixed virtual address.
50
51For new userspace mappings, the kernel makes the entries in its
52page tables like normal.  The only difference is when the kernel
53makes entries in the top (PGD) level.  In addition to setting the
54entry in the main kernel PGD, a copy of the entry is made in the
55userspace page tables' PGD.
56
57This sharing at the PGD level also inherently shares all the lower
58layers of the page tables.  This leaves a single, shared set of
59userspace page tables to manage.  One PTE to lock, one set of
60accessed bits, dirty bits, etc...
61
62Overhead
63========
64
65Protection against side-channel attacks is important.  But,
66this protection comes at a cost:
67
681. Increased Memory Use
69
70  a. Each process now needs an order-1 PGD instead of order-0.
71     (Consumes an additional 4k per process).
72  b. The 'cpu_entry_area' structure must be 2MB in size and 2MB
73     aligned so that it can be mapped by setting a single PMD
74     entry.  This consumes nearly 2MB of RAM once the kernel
75     is decompressed, but no space in the kernel image itself.
76
772. Runtime Cost
78
79  a. CR3 manipulation to switch between the page table copies
80     must be done at interrupt, syscall, and exception entry
81     and exit (it can be skipped when the kernel is interrupted,
82     though.)  Moves to CR3 are on the order of a hundred
83     cycles, and are required at every entry and exit.
84  b. A "trampoline" must be used for SYSCALL entry.  This
85     trampoline depends on a smaller set of resources than the
86     non-PTI SYSCALL entry code, so requires mapping fewer
87     things into the userspace page tables.  The downside is
88     that stacks must be switched at entry time.
89  c. Global pages are disabled for all kernel structures not
90     mapped into both kernel and userspace page tables.  This
91     feature of the MMU allows different processes to share TLB
92     entries mapping the kernel.  Losing the feature means more
93     TLB misses after a context switch.  The actual loss of
94     performance is very small, however, never exceeding 1%.
95  d. Process Context IDentifiers (PCID) is a CPU feature that
96     allows us to skip flushing the entire TLB when switching page
97     tables by setting a special bit in CR3 when the page tables
98     are changed.  This makes switching the page tables (at context
99     switch, or kernel entry/exit) cheaper.  But, on systems with
100     PCID support, the context switch code must flush both the user
101     and kernel entries out of the TLB.  The user PCID TLB flush is
102     deferred until the exit to userspace, minimizing the cost.
103     See intel.com/sdm for the gory PCID/INVPCID details.
104  e. The userspace page tables must be populated for each new
105     process.  Even without PTI, the shared kernel mappings
106     are created by copying top-level (PGD) entries into each
107     new process.  But, with PTI, there are now *two* kernel
108     mappings: one in the kernel page tables that maps everything
109     and one for the entry/exit structures.  At fork(), we need to
110     copy both.
111  f. In addition to the fork()-time copying, there must also
112     be an update to the userspace PGD any time a set_pgd() is done
113     on a PGD used to map userspace.  This ensures that the kernel
114     and userspace copies always map the same userspace
115     memory.
116  g. On systems without PCID support, each CR3 write flushes
117     the entire TLB.  That means that each syscall, interrupt
118     or exception flushes the TLB.
119  h. INVPCID is a TLB-flushing instruction which allows flushing
120     of TLB entries for non-current PCIDs.  Some systems support
121     PCIDs, but do not support INVPCID.  On these systems, addresses
122     can only be flushed from the TLB for the current PCID.  When
123     flushing a kernel address, we need to flush all PCIDs, so a
124     single kernel address flush will require a TLB-flushing CR3
125     write upon the next use of every PCID.
126
127Possible Future Work
128====================
1291. We can be more careful about not actually writing to CR3
130   unless its value is actually changed.
1312. Allow PTI to be enabled/disabled at runtime in addition to the
132   boot-time switching.
133
134Testing
135========
136
137To test stability of PTI, the following test procedure is recommended,
138ideally doing all of these in parallel:
139
1401. Set CONFIG_DEBUG_ENTRY=y
1412. Run several copies of all of the tools/testing/selftests/x86/ tests
142   (excluding MPX and protection_keys) in a loop on multiple CPUs for
143   several minutes.  These tests frequently uncover corner cases in the
144   kernel entry code.  In general, old kernels might cause these tests
145   themselves to crash, but they should never crash the kernel.
1463. Run the 'perf' tool in a mode (top or record) that generates many
147   frequent performance monitoring non-maskable interrupts (see "NMI"
148   in /proc/interrupts).  This exercises the NMI entry/exit code which
149   is known to trigger bugs in code paths that did not expect to be
150   interrupted, including nested NMIs.  Using "-c" boosts the rate of
151   NMIs, and using two -c with separate counters encourages nested NMIs
152   and less deterministic behavior.
153   ::
154
155	while true; do perf record -c 10000 -e instructions,cycles -a sleep 10; done
156
1574. Launch a KVM virtual machine.
1585. Run 32-bit binaries on systems supporting the SYSCALL instruction.
159   This has been a lightly-tested code path and needs extra scrutiny.
160
161Debugging
162=========
163
164Bugs in PTI cause a few different signatures of crashes
165that are worth noting here.
166
167 * Failures of the selftests/x86 code.  Usually a bug in one of the
168   more obscure corners of entry_64.S
169 * Crashes in early boot, especially around CPU bringup.  Bugs
170   in the trampoline code or mappings cause these.
171 * Crashes at the first interrupt.  Caused by bugs in entry_64.S,
172   like screwing up a page table switch.  Also caused by
173   incorrectly mapping the IRQ handler entry code.
174 * Crashes at the first NMI.  The NMI code is separate from main
175   interrupt handlers and can have bugs that do not affect
176   normal interrupts.  Also caused by incorrectly mapping NMI
177   code.  NMIs that interrupt the entry code must be very
178   careful and can be the cause of crashes that show up when
179   running perf.
180 * Kernel crashes at the first exit to userspace.  entry_64.S
181   bugs, or failing to map some of the exit code.
182 * Crashes at first interrupt that interrupts userspace. The paths
183   in entry_64.S that return to userspace are sometimes separate
184   from the ones that return to the kernel.
185 * Double faults: overflowing the kernel stack because of page
186   faults upon page faults.  Caused by touching non-pti-mapped
187   data in the entry code, or forgetting to switch to kernel
188   CR3 before calling into C functions which are not pti-mapped.
189 * Userspace segfaults early in boot, sometimes manifesting
190   as mount(8) failing to mount the rootfs.  These have
191   tended to be TLB invalidation issues.  Usually invalidating
192   the wrong PCID, or otherwise missing an invalidation.
193
194.. [1] https://gruss.cc/files/kaiser.pdf
195.. [2] https://meltdownattack.com/meltdown.pdf
196