summaryrefslogtreecommitdiffstats
path: root/kernel/Documentation/virtual/kvm/mmu.txt
blob: c59bd9bc41efa984cfd5f0e17ac5fd6156acfca6 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
The x86 kvm shadow mmu
======================

The mmu (in arch/x86/kvm, files mmu.[ch] and paging_tmpl.h) is responsible
for presenting a standard x86 mmu to the guest, while translating guest
physical addresses to host physical addresses.

The mmu code attempts to satisfy the following requirements:

- correctness: the guest should not be able to determine that it is running
               on an emulated mmu except for timing (we attempt to comply
               with the specification, not emulate the characteristics of
               a particular implementation such as tlb size)
- security:    the guest must not be able to touch host memory not assigned
               to it
- performance: minimize the performance penalty imposed by the mmu
- scaling:     need to scale to large memory and large vcpu guests
- hardware:    support the full range of x86 virtualization hardware
- integration: Linux memory management code must be in control of guest memory
               so that swapping, page migration, page merging, transparent
               hugepages, and similar features work without change
- dirty tracking: report writes to guest memory to enable live migration
               and framebuffer-based displays
- footprint:   keep the amount of pinned kernel memory low (most memory
               should be shrinkable)
- reliability:  avoid multipage or GFP_ATOMIC allocations

Acronyms
========

pfn   host page frame number
hpa   host physical address
hva   host virtual address
gfn   guest frame number
gpa   guest physical address
gva   guest virtual address
ngpa  nested guest physical address
ngva  nested guest virtual address
pte   page table entry (used also to refer generically to paging structure
      entries)
gpte  guest pte (referring to gfns)
spte  shadow pte (referring to pfns)
tdp   two dimensional paging (vendor neutral term for NPT and EPT)

Virtual and real hardware supported
===================================

The mmu supports first-generation mmu hardware, which allows an atomic switch
of the current paging mode and cr3 during guest entry, as well as
two-dimensional paging (AMD's NPT and Intel's EPT).  The emulated hardware
it exposes is the traditional 2/3/4 level x86 mmu, with support for global
pages, pae, pse, pse36, cr0.wp, and 1GB pages.  Work is in progress to support
exposing NPT capable hardware on NPT capable hosts.

Translation
===========

The primary job of the mmu is to program the processor's mmu to translate
addresses for the guest.  Different translations are required at different
times:

- when guest paging is disabled, we translate guest physical addresses to
  host physical addresses (gpa->hpa)
- when guest paging is enabled, we translate guest virtual addresses, to
  guest physical addresses, to host physical addresses (gva->gpa->hpa)
- when the guest launches a guest of its own, we translate nested guest
  virtual addresses, to nested guest physical addresses, to guest physical
  addresses, to host physical addresses (ngva->ngpa->gpa->hpa)

The primary challenge is to encode between 1 and 3 translations into hardware
that support only 1 (traditional) and 2 (tdp) translations.  When the
number of required translations matches the hardware, the mmu operates in
direct mode; otherwise it operates in shadow mode (see below).

Memory
======

Guest memory (gpa) is part of the user address space of the process that is
using kvm.  Userspace defines the translation between guest addresses and user
addresses (gpa->hva); note that two gpas may alias to the same hva, but not
vice versa.

These hvas may be backed using any method available to the host: anonymous
memory, file backed memory, and device memory.  Memory might be paged by the
host at any time.

Events
======

The mmu is driven by events, some from the guest, some from the host.

Guest generated events:
- writes to control registers (especially cr3)
- invlpg/invlpga instruction execution
- access to missing or protected translations

Host generated events:
- changes in the gpa->hpa translation (either through gpa->hva changes or
  through hva->hpa changes)
- memory pressure (the shrinker)

Shadow pages
============

The principal data structure is the shadow page, 'struct kvm_mmu_page'.  A
shadow page contains 512 sptes, which can be either leaf or nonleaf sptes.  A
shadow page may contain a mix of leaf and nonleaf sptes.

A nonleaf spte allows the hardware mmu to reach the leaf pages and
is not related to a translation directly.  It points to other shadow pages.

A leaf spte corresponds to either one or two translations encoded into
one paging structure entry.  These are always the lowest level of the
translation stack, with optional higher level translations left to NPT/EPT.
Leaf ptes point at guest pages.

The following table shows translations encoded by leaf ptes, with higher-level
translations in parentheses:

 Non-nested guests:
  nonpaging:     gpa->hpa
  paging:        gva->gpa->hpa
  paging, tdp:   (gva->)gpa->hpa
 Nested guests:
  non-tdp:       ngva->gpa->hpa  (*)
  tdp:           (ngva->)ngpa->gpa->hpa

(*) the guest hypervisor will encode the ngva->gpa translation into its page
    tables if npt is not present

Shadow pages contain the following information:
  role.level:
    The level in the shadow paging hierarchy that this shadow page belongs to.
    1=4k sptes, 2=2M sptes, 3=1G sptes, etc.
  role.direct:
    If set, leaf sptes reachable from this page are for a linear range.
    Examples include real mode translation, large guest pages backed by small
    host pages, and gpa->hpa translations when NPT or EPT is active.
    The linear range starts at (gfn << PAGE_SHIFT) and its size is determined
    by role.level (2MB for first level, 1GB for second level, 0.5TB for third
    level, 256TB for fourth level)
    If clear, this page corresponds to a guest page table denoted by the gfn
    field.
  role.quadrant:
    When role.cr4_pae=0, the guest uses 32-bit gptes while the host uses 64-bit
    sptes.  That means a guest page table contains more ptes than the host,
    so multiple shadow pages are needed to shadow one guest page.
    For first-level shadow pages, role.quadrant can be 0 or 1 and denotes the
    first or second 512-gpte block in the guest page table.  For second-level
    page tables, each 32-bit gpte is converted to two 64-bit sptes
    (since each first-level guest page is shadowed by two first-level
    shadow pages) so role.quadrant takes values in the range 0..3.  Each
    quadrant maps 1GB virtual address space.
  role.access:
    Inherited guest access permissions in the form uwx.  Note execute
    permission is positive, not negative.
  role.invalid:
    The page is invalid and should not be used.  It is a root page that is
    currently pinned (by a cpu hardware register pointing to it); once it is
    unpinned it will be destroyed.
  role.cr4_pae:
    Contains the value of cr4.pae for which the page is valid (e.g. whether
    32-bit or 64-bit gptes are in use).
  role.nxe:
    Contains the value of efer.nxe for which the page is valid.
  role.cr0_wp:
    Contains the value of cr0.wp for which the page is valid.
  role.smep_andnot_wp:
    Contains the value of cr4.smep && !cr0.wp for which the page is valid
    (pages for which this is true are different from other pages; see the
    treatment of cr0.wp=0 below).
  role.smap_andnot_wp:
    Contains the value of cr4.smap && !cr0.wp for which the page is valid
    (pages for which this is true are different from other pages; see the
    treatment of cr0.wp=0 below).
  gfn:
    Either the guest page table containing the translations shadowed by this
    page, or the base page frame for linear translations.  See role.direct.
  spt:
    A pageful of 64-bit sptes containing the translations for this page.
    Accessed by both kvm and hardware.
    The page pointed to by spt will have its page->private pointing back
    at the shadow page structure.
    sptes in spt point either at guest pages, or at lower-level shadow pages.
    Specifically, if sp1 and sp2 are shadow pages, then sp1->spt[n] may point
    at __pa(sp2->spt).  sp2 will point back at sp1 through parent_pte.
    The spt array forms a DAG structure with the shadow page as a node, and
    guest pages as leaves.
  gfns:
    An array of 512 guest frame numbers, one for each present pte.  Used to
    perform a reverse map from a pte to a gfn. When role.direct is set, any
    element of this array can be calculated from the gfn field when used, in
    this case, the array of gfns is not allocated. See role.direct and gfn.
  root_count:
    A counter keeping track of how many hardware registers (guest cr3 or
    pdptrs) are now pointing at the page.  While this counter is nonzero, the
    page cannot be destroyed.  See role.invalid.
  parent_ptes:
    The reverse mapping for the pte/ptes pointing at this page's spt. If
    parent_ptes bit 0 is zero, only one spte points at this pages and
    parent_ptes points at this single spte, otherwise, there exists multiple
    sptes pointing at this page and (parent_ptes & ~0x1) points at a data
    structure with a list of parent_ptes.
  unsync:
    If true, then the translations in this page may not match the guest's
    translation.  This is equivalent to the state of the tlb when a pte is
    changed but before the tlb entry is flushed.  Accordingly, unsync ptes
    are synchronized when the guest executes invlpg or flushes its tlb by
    other means.  Valid for leaf pages.
  unsync_children:
    How many sptes in the page point at pages that are unsync (or have
    unsynchronized children).
  unsync_child_bitmap:
    A bitmap indicating which sptes in spt point (directly or indirectly) at
    pages that may be unsynchronized.  Used to quickly locate all unsychronized
    pages reachable from a given page.
  mmu_valid_gen:
    Generation number of the page.  It is compared with kvm->arch.mmu_valid_gen
    during hash table lookup, and used to skip invalidated shadow pages (see
    "Zapping all pages" below.)
  clear_spte_count:
    Only present on 32-bit hosts, where a 64-bit spte cannot be written
    atomically.  The reader uses this while running out of the MMU lock
    to detect in-progress updates and retry them until the writer has
    finished the write.
  write_flooding_count:
    A guest may write to a page table many times, causing a lot of
    emulations if the page needs to be write-protected (see "Synchronized
    and unsynchronized pages" below).  Leaf pages can be unsynchronized
    so that they do not trigger frequent emulation, but this is not
    possible for non-leafs.  This field counts the number of emulations
    since the last time the page table was actually used; if emulation
    is triggered too frequently on this page, KVM will unmap the page
    to avoid emulation in the future.

Reverse map
===========

The mmu maintains a reverse mapping whereby all ptes mapping a page can be
reached given its gfn.  This is used, for example, when swapping out a page.

Synchronized and unsynchronized pages
=====================================

The guest uses two events to synchronize its tlb and page tables: tlb flushes
and page invalidations (invlpg).

A tlb flush means that we need to synchronize all sptes reachable from the
guest's cr3.  This is expensive, so we keep all guest page tables write
protected, and synchronize sptes to gptes when a gpte is written.

A special case is when a guest page table is reachable from the current
guest cr3.  In this case, the guest is obliged to issue an invlpg instruction
before using the translation.  We take advantage of that by removing write
protection from the guest page, and allowing the guest to modify it freely.
We synchronize modified gptes when the guest invokes invlpg.  This reduces
the amount of emulation we have to do when the guest modifies multiple gptes,
or when the a guest page is no longer used as a page table and is used for
random guest data.

As a side effect we have to resynchronize all reachable unsynchronized shadow
pages on a tlb flush.


Reaction to events
==================

- guest page fault (or npt page fault, or ept violation)

This is the most complicated event.  The cause of a page fault can be:

  - a true guest fault (the guest translation won't allow the access) (*)
  - access to a missing translation
  - access to a protected translation
    - when logging dirty pages, memory is write protected
    - synchronized shadow pages are write protected (*)
  - access to untranslatable memory (mmio)

  (*) not applicable in direct mode

Handling a page fault is performed as follows:

 - if the RSV bit of the error code is set, the page fault is caused by guest
   accessing MMIO and cached MMIO information is available.
   - walk shadow page table
   - check for valid generation number in the spte (see "Fast invalidation of
     MMIO sptes" below)
   - cache the information to vcpu->arch.mmio_gva, vcpu->arch.access and
     vcpu->arch.mmio_gfn, and call the emulator
 - If both P bit and R/W bit of error code are set, this could possibly
   be handled as a "fast page fault" (fixed without taking the MMU lock).  See
   the description in Documentation/virtual/kvm/locking.txt.
 - if needed, walk the guest page tables to determine the guest translation
   (gva->gpa or ngpa->gpa)
   - if permissions are insufficient, reflect the fault back to the guest
 - determine the host page
   - if this is an mmio request, there is no host page; cache the info to
     vcpu->arch.mmio_gva, vcpu->arch.access and vcpu->arch.mmio_gfn
 - walk the shadow page table to find the spte for the translation,
   instantiating missing intermediate page tables as necessary
   - If this is an mmio request, cache the mmio info to the spte and set some
     reserved bit on the spte (see callers of kvm_mmu_set_mmio_spte_mask)
 - try to unsynchronize the page
   - if successful, we can let the guest continue and modify the gpte
 - emulate the instruction
   - if failed, unshadow the page and let the guest continue
 - update any translations that were modified by the instruction

invlpg handling:

  - walk the shadow page hierarchy and drop affected translations
  - try to reinstantiate the indicated translation in the hope that the
    guest will use it in the near future

Guest control register updates:

- mov to cr3
  - look up new shadow roots
  - synchronize newly reachable shadow pages

- mov to cr0/cr4/efer
  - set up mmu context for new paging mode
  - look up new shadow roots
  - synchronize newly reachable shadow pages

Host translation updates:

  - mmu notifier called with updated hva
  - look up affected sptes through reverse map
  - drop (or update) translations

Emulating cr0.wp
================

If tdp is not enabled, the host must keep cr0.wp=1 so page write protection
works for the guest kernel, not guest guest userspace.  When the guest
cr0.wp=1, this does not present a problem.  However when the guest cr0.wp=0,
we cannot map the permissions for gpte.u=1, gpte.w=0 to any spte (the
semantics require allowing any guest kernel access plus user read access).

We handle this by mapping the permissions to two possible sptes, depending
on fault type:

- kernel write fault: spte.u=0, spte.w=1 (allows full kernel access,
  disallows user access)
- read fault: spte.u=1, spte.w=0 (allows full read access, disallows kernel
  write access)

(user write faults generate a #PF)

In the first case there are two additional complications:
- if CR4.SMEP is enabled: since we've turned the page into a kernel page,
  the kernel may now execute it.  We handle this by also setting spte.nx.
  If we get a user fetch or read fault, we'll change spte.u=1 and
  spte.nx=gpte.nx back.
- if CR4.SMAP is disabled: since the page has been changed to a kernel
  page, it can not be reused when CR4.SMAP is enabled. We set
  CR4.SMAP && !CR0.WP into shadow page's role to avoid this case. Note,
  here we do not care the case that CR4.SMAP is enabled since KVM will
  directly inject #PF to guest due to failed permission check.

To prevent an spte that was converted into a kernel page with cr0.wp=0
from being written by the kernel after cr0.wp has changed to 1, we make
the value of cr0.wp part of the page role.  This means that an spte created
with one value of cr0.wp cannot be used when cr0.wp has a different value -
it will simply be missed by the shadow page lookup code.  A similar issue
exists when an spte created with cr0.wp=0 and cr4.smep=0 is used after
changing cr4.smep to 1.  To avoid this, the value of !cr0.wp && cr4.smep
is also made a part of the page role.

Large pages
===========

The mmu supports all combinations of large and small guest and host pages.
Supported page sizes include 4k, 2M, 4M, and 1G.  4M pages are treated as
two separate 2M pages, on both guest and host, since the mmu always uses PAE
paging.

To instantiate a large spte, four constraints must be satisfied:

- the spte must point to a large host page
- the guest pte must be a large pte of at least equivalent size (if tdp is
  enabled, there is no guest pte and this condition is satisfied)
- if the spte will be writeable, the large page frame may not overlap any
  write-protected pages
- the guest page must be wholly contained by a single memory slot

To check the last two conditions, the mmu maintains a ->write_count set of
arrays for each memory slot and large page size.  Every write protected page
causes its write_count to be incremented, thus preventing instantiation of
a large spte.  The frames at the end of an unaligned memory slot have
artificially inflated ->write_counts so they can never be instantiated.

Zapping all pages (page generation count)
=========================================

For the large memory guests, walking and zapping all pages is really slow
(because there are a lot of pages), and also blocks memory accesses of
all VCPUs because it needs to hold the MMU lock.

To make it be more scalable, kvm maintains a global generation number
which is stored in kvm->arch.mmu_valid_gen.  Every shadow page stores
the current global generation-number into sp->mmu_valid_gen when it
is created.  Pages with a mismatching generation number are "obsolete".

When KVM need zap all shadow pages sptes, it just simply increases the global
generation-number then reload root shadow pages on all vcpus.  As the VCPUs
create new shadow page tables, the old pages are not used because of the
mismatching generation number.

KVM then walks through all pages and zaps obsolete pages.  While the zap
operation needs to take the MMU lock, the lock can be released periodically
so that the VCPUs can make progress.

Fast invalidation of MMIO sptes
===============================

As mentioned in "Reaction to events" above, kvm will cache MMIO
information in leaf sptes.  When a new memslot is added or an existing
memslot is changed, this information may become stale and needs to be
invalidated.  This also needs to hold the MMU lock while walking all
shadow pages, and is made more scalable with a similar technique.

MMIO sptes have a few spare bits, which are used to store a
generation number.  The global generation number is stored in
kvm_memslots(kvm)->generation, and increased whenever guest memory info
changes.  This generation number is distinct from the one described in
the previous section.

When KVM finds an MMIO spte, it checks the generation number of the spte.
If the generation number of the spte does not equal the global generation
number, it will ignore the cached MMIO information and handle the page
fault through the slow path.

Since only 19 bits are used to store generation-number on mmio spte, all
pages are zapped when there is an overflow.

Unfortunately, a single memory access might access kvm_memslots(kvm) multiple
times, the last one happening when the generation number is retrieved and
stored into the MMIO spte.  Thus, the MMIO spte might be created based on
out-of-date information, but with an up-to-date generation number.

To avoid this, the generation number is incremented again after synchronize_srcu
returns; thus, the low bit of kvm_memslots(kvm)->generation is only 1 during a
memslot update, while some SRCU readers might be using the old copy.  We do not
want to use an MMIO sptes created with an odd generation number, and we can do
this without losing a bit in the MMIO spte.  The low bit of the generation
is not stored in MMIO spte, and presumed zero when it is extracted out of the
spte.  If KVM is unlucky and creates an MMIO spte while the low bit is 1,
the next access to the spte will always be a cache miss.


Further reading
===============

- NPT presentation from KVM Forum 2008
  http://www.linux-kvm.org/wiki/images/c/c8/KvmForum2008%24kdf2008_21.pdf