XNU hibernation
===============
Suspending the entire system state to RAM.
Goal
----
This document discusses the design and implementation of XNU hibernation. The
reader is assumed to generally understand how standard suspend to RAM (S2R)
works in XNU; a detailed discussion of S2R is beyond the scope of this
discussion.
Vocabulary
----------
* Polled I/O : a mode of operation supported by I/O drivers (primarily storage
devices) where operations may be conducted from a single-threaded
context with interrupts disabled
* S2R : Suspend to RAM (aka sleep)
* WKdm : Wilson-Kaplan direct mapped compression
Background
----------
In order to prolong battery life, XNU supports suspending/powering off various
devices and preserving the state of those devices in memory. This feature is
referred to as suspend to RAM (S2R). In this mode, IOKit delivers a number of
notifications to IOServices to allow them to participate in S2R.
What is hibernation?
--------------------
Hibernation is a feature built on the foundation of S2R. However, while S2R
preserves state in memory (which must therefore remain powered), hibernation
preserves contents to persistent storage (the disk) and then completely powers
the system off.
Hibernation entry
-----------------
During hibernation, XNU invokes essentially the normal S2R machinery, but with a
few hibernation-specific differences:
* PMRootDomain calls `IOHibernateSystemSleep()` before system sleep (devices
awake, normal execution context).
* `IOHibernateSystemSleep()` opens the hibernation file (or partition) at the
BSD level, grabs its extents and searches for a polling driver willing to work
with that IOMedia.
* The BSD code makes an ioctl to the storage driver to get the partition base
offset to the disk, and other ioctls to get the transfer constraints.
* If successful, the file is written to make sure it's initially not bootable
(in case of later failure) and the `boot-image` nvram variable is set to point
to the first block of the file. (has to be done here because writing to nvram
may block, so we have to do this before preemption is disabled).
* `hibernate_page_list_allocate()` is called to allocate page bitmaps for all
DRAM.
- The hibernation code represents every page of physical memory in page
bitmaps of type `hibernate_bitmap_t`. There is one page bitmap per range of
memory, with a bit to represent each page in that range; these page bitmaps
are in turn stored in a `hibernate_page_list_t`. The page bitmaps are used
to represent, for each page, whether preservation of that page is necessary.
- On ARM64, `secure_hmac_get_io_ranges()` is called to get a list of the I/O
regions that need to be included in the hibernation image (for example, the
GPU UAT handoff region). These I/O regions are typically DRAM regions carved
out by iBoot that exist outside of kernel-managed memory. A page bitmap is
allocated for each one of these ranges (as well as a single bitmap for the
kernel-managed DRAM memory).
* `hibernate_processor_setup()` is called to set up some platform-specific state
needed in the hibernation image header. `hibernate_processor_setup()` also
sets a flag in the boot processor's `cpu_data_t` to indicate that hibernation
is in progress on this CPU.
* At this point, `gIOHibernateState` is set to the value
`kIOHibernateStateHibernating`.
* Regular sleep progresses; some drivers may inspect the root domain property
`kIOHibernateStateKey` to modify behavior. The platform drivers save state to
memory as usual, but any drivers required for hibernation I/O are left in a
state such that polled I/Os can be issued.
* By the time regular sleep has completed, all CPUs but the boot CPU have been
halted, and we are running on the boot CPU's idle thread in the shutdown
context, with preemption disabled.
* Eventually the platform calls `hibernate_write_image()` in the shutdown
context on the last cpu, at which point memory is ready to be saved. This call
is made from `acpi_hibernate()` on Intel and from `ml_arm_sleep()` on ARM64.
* `hibernate_write_image()` runs in the shutdown context, where no blocking is
permitted because preemption is disabled. `hibernate_write_image()` calls
`hibernate_page_list_setall()` to get the page bitmaps of DRAM that need to be
saved.
* All pages are assumed to be saved (as part of the wired image) unless
explicitly subtracted by `hibernate_page_list_setall()`.
`hibernate_page_list_setall()` calls `hibernate_page_list_setall_machine()` to
make platform-specific amendments to the page bitmaps.
* `hibernate_write_image()` writes the image header and extents list. The header
includes the second file extent so that only the header block is needed to
read the file, regardless of the underlying filesystem.
- The extents list describes the file's layout on disk. This block list makes
it possible for the platform booter to read the hibernation file from disk
without having to understand the underlying filesystem.
* Some sections of memory are written directly (and uncompressed) to the image.
These are the portions of XNU itself that are required during hibernation
resume, as well as some other data that is required by the platform booter.
- On Intel, the `__HIB` segment is written to the hibernation image.
- On ARM64, because of ctrr/ktrr, a single `__HIB` segment isn't possible.
Instead, a number of sections of the kernel are written:
`__TEXT_EXEC,__hib_text`, `__DATA,__hib_data`, and
`__DATA_CONST,__hib_const`. The `__PPL` segment is also stored to the image
so that the PPL hmac driver can be used during hibernation resume. Certain
other pieces of memory must also be written unmodified to the hibernation
image for use by iBoot. Those pieces are described in the device tree so
that XNU doesn't need to know the details.
`secure_hmac_fetch_hibseg_and_info()` is used to determine the set of memory
regions to be stored in this phase. This routine also calculates an HMAC
that can be used by the booter to validate this content.
* The portions of XNU (code and data) that are stored directly to the
hibernation image should be entirely self-contained; these are the only
portions of XNU that are available during resume to decompress the image.
* Some additional pages are removed from the page bitmaps; these include various
temporary buffers used for hibernation.
* The page bitmaps are written to the image.
* More areas are removed from the page bitmaps (after they have been written to
the image); these include the pages already stored directly to the image, as
well as the stack that hibernation resume will run on.
`hibernate_page_list_set_volatile()` is invoked to make platform-specific
amendments to the page bitmaps.
* Each wired page is compressed and written and then each non-wired page.
Compression and disk writes can occur in parallel if the polled mode I/O
driver supports this.
- On ARM64, `secure_hmac_update_and_compress_page()` is called for each page
included in the image so that the PPL can compute an HMAC of the hibernation
payload.
* The image header records the values of `mach_absolute_time()` and
`mach_continuous_time()` close to the end of `hibernate_write_image()`. These
values can be used to fix up the offets applied to the hardware clock after
hibernation exit.
* The image header is finalized.
- On ARM64, `secure_hmac_final()` is called to compute the HMAC of the
hibernation payload. There are actually two separate HMACs computed, one for
the wired pages and one for the non-wired pages. These HMACs are stored in
the image header.
- On ARM64, `secure_hmac_fetch_rorgn_sha()` and `secure_hmac_fetch_rorgn_hmac()` are
called to obtain the SHA256 and HMAC of the read-only region. They were
calculated on cold boot. They are stored in the image header.
This is described in more detail in the "Security details" section of this
document.
- On ARM64, `secure_hmac_finalize_image()` is called to compute the HMAC of the
header of the image. This is described in more detail in the "Security
details" section of this document.
* The image header is written to the start of the file and the polling driver
closed.
* The machine powers down.
- On Intel, depending on power settings, the system could sleep instead at
this point. This allows for "safe sleep" where RAM remains powered until the
user wakes the system or the battery dies.
- On ARM64, we do not support this mode because hibernation is intended to
only be invoked on a critical battery event.
Hibernation exit
----------------
* The platform booter sees the `boot-image` nvram variable containing the device
and block number of the image, reads the header, and if the signature is
correct proceeds. The `boot-image` variable is cleared.
- On ARM64, iBoot takes the read-only region SHA256 value from the image
header and calculates an HMAC. It then compares the HMAC against the
value stored in the image header. If they do not match, iBoot panics.
* The platform booter reads the portion of the image used for wired pages, to
memory. Its assumed this will fit in memory in its entirety. The image is
decrypted (either transparently by ANS or in software, depending on platform
support). The platform booter is not expected to decompress any of the
payload; that is the kernel's responsibility.
* The platform booter copies the portions of XNU that were previously saved to
the image back to their original physical addresses in memory.
* The platform booter invokes `hibernate_machine_entrypoint()`, passing in the
location of the image in memory. Translation is off. Only code and data that
was mapped by the booter is safe to call, since all the other wired pages are
still compressed in the image.
- On Intel, `hibernate_machine_entrypoint()` sets up a simple temporary page
table; this page table will later be modified as necessary while pages are
being restored.
- On ARM64, `hibernate_machine_entrypoint()` sets up a temporary page table
such that all of the required XNU code pages are executable, all data pages
are readable/writable as necessary, and all of the rest of memory is mapped
such that it can be written to during restore. Some device registers also
have to be mapped to support serial logging and using the hmac block.
* Any pages occupied by the raw image are removed from from the page bitmaps.
- On Intel, this is done in `hibernate_kernel_entrypoint()`.
- On ARM64, we have to do this from `hibernate_machine_entrypoint()` because
we borrow free pages (as indicated by the page bitmaps) to store the
temporary page table.
* `hibernate_machine_entrypoint()` calls `hibernate_kernel_entrypoint()`.
* `hibernate_kernel_entrypoint()` uses the page bitmaps to determine which pages
can be uncompressed from the wired image directly to their final location. Any
pages that conflict with the image itself are copied to interim scratch space.
* After all of the image has been parsed, the pages that were temporarily copied
to scratch are uncompressed to their final location, overwriting pages in the
wired image.
- `hibernate_restore_phys_page()` is used to actually copy pages to their
final location.
* At this point, `gIOHibernateState` is set to
`kIOHibernateStateWakingFromHibernate`.
* `pal_hib_patchup()` is called to perform platform-specific post-resume fixups
- On Intel, `pal_hib_patchup()` is a no-op.
- On ARM64, `pal_hib_patchup()` is responsible for validating the HMAC of the
wired pages. `pal_hib_patchup()` also fixes up other state (such as some
PPL-related context).
* After all of the wired pages have been restored, a wake from sleep is
simulated.
- On Intel, `hibernate_kernel_entrypoint()` calls `acpi_wake_prot_entry()`.
- On ARM64, `hibernate_kernel_entrypoint()` returns to
`hibernate_machine_entrypoint()`, which then jumps to `reset_vector`.
* The kernel proceeds on essentially a normal S2R wake, with some
hibernation-specific changes.
- On ARM64, an important difference is that a normal S2R wake on some
platforms will run through the reconfig engine, whereas a hibernate wake
cannot invoke the reconfig engine and must emulate some of the reconfig
sequence on the AP.
- On ARM64, some further fixup is done in `arm_init_cpu()`.
+ `wake_abstime` needs to be restored to the last absolute time captured
during hibernation entry. This is necessary because during normal S2R,
`wake_abstime` is captured too early; later calls to
`mach_absolute_time()` in the hibernation entry path cause the
`s_last_absolute_time` test to fail if we don't do this fixup.
+ `hwclock_conttime_offset` is set to the `hwClockOffset` value that iBoot
computed. This is necessary since `ml_get_hwclock()` does not tick across
hibernation but `mach_continuous_time()` is expected to.
+ The boot CPU's idle thread preemption_count also has to be fixed up. This
is necessary because the page containing preemption_count is captured
when the count is set to 1 (since the page is captured from within the
PPL).
* After the platform CPU init code is called, `hibernate_machine_init()` is
called to restore the rest of memory, using the polled mode driver, before
other threads can run or any devices are turned on. This split of wired vs.
non-wired pages reduces the memory usage for the platform booter, and allows
decompression in parallel with disk reads for the non-wired pages.
* The polling driver is closed down and regular wake proceeds.
* When the kernel calls IOKit to wake (normal execution context)
`hibernate_teardown()` is called to release any memory.
* The hibernation file is closed via BSD.
Hibernation file management
---------------------------
powerd in userspace is responsible for managing the lifecycle of the hibernation
file. The details of this lifecycle are beyond the scope of this document, but
essentially, it gets created and its space is preallocated by powerd the first
time the system hibernates. powerd can also grow the file as necessary.
Security details
----------------
### Intel:
* The hibernation image is encrypted with a key obtained from the APFS
`APFSMEDIA_GETHIBERKEY` platform function.
### ARM64:
* The hibernation image is encrypted with a key obtained from the SEP. The
details for how this key is derived and used are beyond the scope of this
document, but are documented in detail in the AppleSEPOS project
(doc/SecureHibernation).
* Various portions of the hibernation image have HMACs calculated over them. All
HMACs are calculated by the PPL. The exact scheme for computing these HMACs is
documented in more detail in ppl_hib.c, but the HMACs that are calculated are:
- `imageHeaderHMAC` is an HMAC of the header of the image, up to
`imageHeaderHMACSize`. However, because of the order that data is written
(the header is the last thing actually written), the HMAC is actually
calculated as `HMAC(SHA([data after header up to imageHeaderHMACSize],
[header]))`.
- `handoffHMAC` is an HMAC of the `IOHibernateHandoff` data passed from iBoot
to XNU
- `image1PagesHMAC` is an HMAC of the wired pages that were stored to the
hibernation image
- `image2PagesHMAC` is an HMAC of the non-wired pages that were stored to the
hibernation image
* The PPL hibernation driver also keeps track of every PPL-owned page being
hashed (both kernel-managed memory and I/O memory owned by the PPL). This will
be double-checked in `secure_hmac_finalize_image()` to ensure that all PPL-owned
memory is included in the hibernation image. Any missing pages will panic the
system as the absence of PPL pages in the image could be a security risk (and
surely a bug).
* During early boot, `secure_hmac_compute_rorgn_hmac()` is used to measure the
entirety of the rorgn. On hibernation resume, the same function is invoked to
verify that the rorgn matches its original contents.
- Only the SHA256 of the rorgn is compared on resume. The SIO HMAC key1, used
to compute this HMAC, is invalidated by iBoot on the resume path after it
verifies the HMAC. See rdar://75750348 (xnu should store the SHA of the
read-only region along with the hash in memory for iBoot to validate on
hibernate resume).