platform/kernel/linux-rpi.git
2 years agopercpu_ref_init(): clean ->percpu_count_ref on failure
Al Viro [Wed, 18 May 2022 06:13:40 +0000 (02:13 -0400)]
percpu_ref_init(): clean ->percpu_count_ref on failure

[ Upstream commit a91714312eb16f9ecd1f7f8b3efe1380075f28d4 ]

That way percpu_ref_exit() is safe after failing percpu_ref_init().
At least one user (cgroup_create()) had a double-free that way;
there might be other similar bugs.  Easier to fix in percpu_ref_init(),
rather than playing whack-a-mole in sloppy users...

Usual symptoms look like a messed refcounting in one of subsystems
that use percpu allocations (might be percpu-refcount, might be
something else).  Having refcounts for two different objects share
memory is Not Nice(tm)...

Reported-by: syzbot+5b1e53987f858500ec00@syzkaller.appspotmail.com
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Sasha Levin <sashal@kernel.org>
2 years agoKVM: arm64: Don't hypercall before EL2 init
Quentin Perret [Fri, 13 May 2022 09:26:07 +0000 (09:26 +0000)]
KVM: arm64: Don't hypercall before EL2 init

[ Upstream commit 2e40316753ee552fb598e8da8ca0d20a04e67453 ]

Will reported the following splat when running with Protected KVM
enabled:

[    2.427181] ------------[ cut here ]------------
[    2.427668] WARNING: CPU: 3 PID: 1 at arch/arm64/kvm/mmu.c:489 __create_hyp_private_mapping+0x118/0x1ac
[    2.428424] Modules linked in:
[    2.429040] CPU: 3 PID: 1 Comm: swapper/0 Not tainted 5.18.0-rc2-00084-g8635adc4efc7 #1
[    2.429589] Hardware name: QEMU QEMU Virtual Machine, BIOS 0.0.0 02/06/2015
[    2.430286] pstate: 80000005 (Nzcv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[    2.430734] pc : __create_hyp_private_mapping+0x118/0x1ac
[    2.431091] lr : create_hyp_exec_mappings+0x40/0x80
[    2.431377] sp : ffff80000803baf0
[    2.431597] x29: ffff80000803bb00 x28: 0000000000000000 x27: 0000000000000000
[    2.432156] x26: 0000000000000000 x25: 0000000000000000 x24: 0000000000000000
[    2.432561] x23: ffffcd96c343b000 x22: 0000000000000000 x21: ffff80000803bb40
[    2.433004] x20: 0000000000000004 x19: 0000000000001800 x18: 0000000000000000
[    2.433343] x17: 0003e68cf7efdd70 x16: 0000000000000004 x15: fffffc81f602a2c8
[    2.434053] x14: ffffdf8380000000 x13: ffffcd9573200000 x12: ffffcd96c343b000
[    2.434401] x11: 0000000000000004 x10: ffffcd96c1738000 x9 : 0000000000000004
[    2.434812] x8 : ffff80000803bb40 x7 : 7f7f7f7f7f7f7f7f x6 : 544f422effff306b
[    2.435136] x5 : 000000008020001e x4 : ffff207d80a88c00 x3 : 0000000000000005
[    2.435480] x2 : 0000000000001800 x1 : 000000014f4ab800 x0 : 000000000badca11
[    2.436149] Call trace:
[    2.436600]  __create_hyp_private_mapping+0x118/0x1ac
[    2.437576]  create_hyp_exec_mappings+0x40/0x80
[    2.438180]  kvm_init_vector_slots+0x180/0x194
[    2.458941]  kvm_arch_init+0x80/0x274
[    2.459220]  kvm_init+0x48/0x354
[    2.459416]  arm_init+0x20/0x2c
[    2.459601]  do_one_initcall+0xbc/0x238
[    2.459809]  do_initcall_level+0x94/0xb4
[    2.460043]  do_initcalls+0x54/0x94
[    2.460228]  do_basic_setup+0x1c/0x28
[    2.460407]  kernel_init_freeable+0x110/0x178
[    2.460610]  kernel_init+0x20/0x1a0
[    2.460817]  ret_from_fork+0x10/0x20
[    2.461274] ---[ end trace 0000000000000000 ]---

Indeed, the Protected KVM mode promotes __create_hyp_private_mapping()
to a hypercall as EL1 no longer has access to the hypervisor's stage-1
page-table. However, the call from kvm_init_vector_slots() happens after
pKVM has been initialized on the primary CPU, but before it has been
initialized on secondaries. As such, if the KVM initcall procedure is
migrated from one CPU to another in this window, the hypercall may end up
running on a CPU for which EL2 has not been initialized.

Fortunately, the pKVM hypervisor doesn't rely on the host to re-map the
vectors in the private range, so the hypercall in question is in fact
superfluous. Skip it when pKVM is enabled.

Reported-by: Will Deacon <will@kernel.org>
Signed-off-by: Quentin Perret <qperret@google.com>
[maz: simplified the checks slightly]
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20220513092607.35233-1-qperret@google.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
2 years agopinctrl: sunxi: fix f1c100s uart2 function
IotaHydrae [Wed, 4 May 2022 11:59:04 +0000 (19:59 +0800)]
pinctrl: sunxi: fix f1c100s uart2 function

[ Upstream commit fa8785e5931367e2b43f2c507f26bcf3e281c0ca ]

Change suniv f1c100s pinctrl,PD14 multiplexing function lvds1 to uart2

When the pin PD13 and PD14 is setting up to uart2 function in dts,
there's an error occurred:
1c20800.pinctrl: unsupported function uart2 on pin PD14

Because 'uart2' is not any one multiplexing option of PD14,
and pinctrl don't know how to configure it.

So change the pin PD14 lvds1 function to uart2.

Signed-off-by: IotaHydrae <writeforever@foxmail.com>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
Link: https://lore.kernel.org/r/tencent_70C1308DDA794C81CAEF389049055BACEC09@qq.com
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
2 years agoALSA: usb-audio: Don't get sample rate for MCT Trigger 5 USB-to-HDMI
Forest Crossman [Wed, 4 May 2022 00:24:44 +0000 (19:24 -0500)]
ALSA: usb-audio: Don't get sample rate for MCT Trigger 5 USB-to-HDMI

[ Upstream commit d7be213849232a2accb219d537edf056d29186b4 ]

This device doesn't support reading the sample rate, so we need to apply
this quirk to avoid a 15-second delay waiting for three timeouts.

Signed-off-by: Forest Crossman <cyrozap@gmail.com>
Link: https://lore.kernel.org/r/20220504002444.114011-2-cyrozap@gmail.com
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Signed-off-by: Sasha Levin <sashal@kernel.org>
2 years agoLinux 5.15.44
Greg Kroah-Hartman [Mon, 30 May 2022 07:29:18 +0000 (09:29 +0200)]
Linux 5.15.44

Link: https://lore.kernel.org/r/20220527084850.364560116@linuxfoundation.org
Tested-by: Guenter Roeck <linux@roeck-us.net>
Tested-by: Bagas Sanjaya <bagasdotme@gmail.com>
Tested-by: Linux Kernel Functional Testing <lkft@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoALSA: ctxfi: Add SB046x PCI ID
Edward Matijevic [Sat, 21 May 2022 04:45:15 +0000 (23:45 -0500)]
ALSA: ctxfi: Add SB046x PCI ID

commit 1b073ebb174d0c7109b438e0a5eb4495137803ec upstream.

Adds the PCI ID for X-Fi cards sold under the Platnum and XtremeMusic names

Before: snd_ctxfi 0000:05:05.0: chip 20K1 model Unknown (1102:0021) is found
After: snd_ctxfi 0000:05:05.0: chip 20K1 model SB046x (1102:0021) is found

[ This is only about defining the model name string, and the rest is
  handled just like before, as a default unknown device.
  Edward confirmed that the stuff has been working fine -- tiwai ]

Signed-off-by: Edward Matijevic <motolav@gmail.com>
Cc: <stable@vger.kernel.org>
Link: https://lore.kernel.org/r/cae7d1a4-8bd9-7dfe-7427-db7e766f7272@gmail.com
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: check for signals after page of pool writes
Jason A. Donenfeld [Sun, 22 May 2022 20:25:41 +0000 (22:25 +0200)]
random: check for signals after page of pool writes

commit 1ce6c8d68f8ac587f54d0a271ac594d3d51f3efb upstream.

get_random_bytes_user() checks for signals after producing a PAGE_SIZE
worth of output, just like /dev/zero does. write_pool() is doing
basically the same work (actually, slightly more expensive), and so
should stop to check for signals in the same way. Let's also name it
write_pool_user() to match get_random_bytes_user(), so this won't be
misused in the future.

Before this patch, massive writes to /dev/urandom would tie up the
process for an extremely long time and make it unterminatable. After, it
can be successfully interrupted. The following test program can be used
to see this works as intended:

  #include <unistd.h>
  #include <fcntl.h>
  #include <signal.h>
  #include <stdio.h>

  static unsigned char x[~0U];

  static void handle(int) { }

  int main(int argc, char *argv[])
  {
    pid_t pid = getpid(), child;
    int fd;
    signal(SIGUSR1, handle);
    if (!(child = fork())) {
      for (;;)
        kill(pid, SIGUSR1);
    }
    fd = open("/dev/urandom", O_WRONLY);
    pause();
    printf("interrupted after writing %zd bytes\n", write(fd, x, sizeof(x)));
    close(fd);
    kill(child, SIGTERM);
    return 0;
  }

Result before: "interrupted after writing 2147479552 bytes"
Result after: "interrupted after writing 4096 bytes"

Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: wire up fops->splice_{read,write}_iter()
Jens Axboe [Thu, 19 May 2022 23:31:37 +0000 (17:31 -0600)]
random: wire up fops->splice_{read,write}_iter()

commit 79025e727a846be6fd215ae9cdb654368ac3f9a6 upstream.

Now that random/urandom is using {read,write}_iter, we can wire it up to
using the generic splice handlers.

Fixes: 36e2c7421f02 ("fs: don't allow splice read/write without explicit ops")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
[Jason: added the splice_write path. Note that sendfile() and such still
 does not work for read, though it does for write, because of a file
 type restriction in splice_direct_to_actor(), which I'll address
 separately.]
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: convert to using fops->write_iter()
Jens Axboe [Thu, 19 May 2022 23:43:15 +0000 (17:43 -0600)]
random: convert to using fops->write_iter()

commit 22b0a222af4df8ee9bb8e07013ab44da9511b047 upstream.

Now that the read side has been converted to fix a regression with
splice, convert the write side as well to have some symmetry in the
interface used (and help deprecate ->write()).

Signed-off-by: Jens Axboe <axboe@kernel.dk>
[Jason: cleaned up random_ioctl a bit, require full writes in
 RNDADDENTROPY since it's crediting entropy, simplify control flow of
 write_pool(), and incorporate suggestions from Al.]
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: convert to using fops->read_iter()
Jens Axboe [Thu, 19 May 2022 23:31:36 +0000 (17:31 -0600)]
random: convert to using fops->read_iter()

commit 1b388e7765f2eaa137cf5d92b47ef5925ad83ced upstream.

This is a pre-requisite to wiring up splice() again for the random
and urandom drivers. It also allows us to remove the INT_MAX check in
getrandom(), because import_single_range() applies capping internally.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
[Jason: rewrote get_random_bytes_user() to simplify and also incorporate
 additional suggestions from Al.]
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: unify batched entropy implementations
Jason A. Donenfeld [Sat, 14 May 2022 22:22:05 +0000 (00:22 +0200)]
random: unify batched entropy implementations

commit 3092adcef3ffd2ef59634998297ca8358461ebce upstream.

There are currently two separate batched entropy implementations, for
u32 and u64, with nearly identical code, with the goal of avoiding
unaligned memory accesses and letting the buffers be used more
efficiently. Having to maintain these two functions independently is a
bit of a hassle though, considering that they always need to be kept in
sync.

This commit factors them out into a type-generic macro, so that the
expansion produces the same code as before, such that diffing the
assembly shows no differences. This will also make it easier in the
future to add u16 and u8 batches.

This was initially tested using an always_inline function and letting
gcc constant fold the type size in, but the code gen was less efficient,
and in general it was more verbose and harder to follow. So this patch
goes with the boring macro solution, similar to what's already done for
the _wait functions in random.h.

Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: move randomize_page() into mm where it belongs
Jason A. Donenfeld [Sat, 14 May 2022 11:59:30 +0000 (13:59 +0200)]
random: move randomize_page() into mm where it belongs

commit 5ad7dd882e45d7fe432c32e896e2aaa0b21746ea upstream.

randomize_page is an mm function. It is documented like one. It contains
the history of one. It has the naming convention of one. It looks
just like another very similar function in mm, randomize_stack_top().
And it has always been maintained and updated by mm people. There is no
need for it to be in random.c. In the "which shape does not look like
the other ones" test, pointing to randomize_page() is correct.

So move randomize_page() into mm/util.c, right next to the similar
randomize_stack_top() function.

This commit contains no actual code changes.

Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: move initialization functions out of hot pages
Jason A. Donenfeld [Fri, 13 May 2022 14:17:12 +0000 (16:17 +0200)]
random: move initialization functions out of hot pages

commit 560181c27b582557d633ecb608110075433383af upstream.

Much of random.c is devoted to initializing the rng and accounting for
when a sufficient amount of entropy has been added. In a perfect world,
this would all happen during init, and so we could mark these functions
as __init. But in reality, this isn't the case: sometimes the rng only
finishes initializing some seconds after system init is finished.

For this reason, at the moment, a whole host of functions that are only
used relatively close to system init and then never again are intermixed
with functions that are used in hot code all the time. This creates more
cache misses than necessary.

In order to pack the hot code closer together, this commit moves the
initialization functions that can't be marked as __init into
.text.unlikely by way of the __cold attribute.

Of particular note is moving credit_init_bits() into a macro wrapper
that inlines the crng_ready() static branch check. This avoids a
function call to a nop+ret, and most notably prevents extra entropy
arithmetic from being computed in mix_interrupt_randomness().

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: make consistent use of buf and len
Jason A. Donenfeld [Fri, 13 May 2022 11:18:46 +0000 (13:18 +0200)]
random: make consistent use of buf and len

commit a19402634c435a4eae226df53c141cdbb9922e7b upstream.

The current code was a mix of "nbytes", "count", "size", "buffer", "in",
and so forth. Instead, let's clean this up by naming input parameters
"buf" (or "ubuf") and "len", so that you always understand that you're
reading this variety of function argument.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: use proper return types on get_random_{int,long}_wait()
Jason A. Donenfeld [Fri, 13 May 2022 10:32:23 +0000 (12:32 +0200)]
random: use proper return types on get_random_{int,long}_wait()

commit 7c3a8a1db5e03d02cc0abb3357a84b8b326dfac3 upstream.

Before these were returning signed values, but the API is intended to be
used with unsigned values.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: remove extern from functions in header
Jason A. Donenfeld [Fri, 13 May 2022 10:29:38 +0000 (12:29 +0200)]
random: remove extern from functions in header

commit 7782cfeca7d420e8bb707613d4cfb0f7ff29bb3a upstream.

Accoriding to the kernel style guide, having `extern` on functions in
headers is old school and deprecated, and doesn't add anything. So remove
them from random.h, and tidy up the file a little bit too.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: use static branch for crng_ready()
Jason A. Donenfeld [Tue, 3 May 2022 13:30:45 +0000 (15:30 +0200)]
random: use static branch for crng_ready()

commit f5bda35fba615ace70a656d4700423fa6c9bebee upstream.

Since crng_ready() is only false briefly during initialization and then
forever after becomes true, we don't need to evaluate it after, making
it a prime candidate for a static branch.

One complication, however, is that it changes state in a particular call
to credit_init_bits(), which might be made from atomic context, which
means we must kick off a workqueue to change the static key. Further
complicating things, credit_init_bits() may be called sufficiently early
on in system initialization such that system_wq is NULL.

Fortunately, there exists the nice function execute_in_process_context(),
which will immediately execute the function if !in_interrupt(), and
otherwise defer it to a workqueue. During early init, before workqueues
are available, in_interrupt() is always false, because interrupts
haven't even been enabled yet, which means the function in that case
executes immediately. Later on, after workqueues are available,
in_interrupt() might be true, but in that case, the work is queued in
system_wq and all goes well.

Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Sultan Alsawaf <sultan@kerneltoast.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: credit architectural init the exact amount
Jason A. Donenfeld [Thu, 12 May 2022 13:32:26 +0000 (15:32 +0200)]
random: credit architectural init the exact amount

commit 12e45a2a6308105469968951e6d563e8f4fea187 upstream.

RDRAND and RDSEED can fail sometimes, which is fine. We currently
initialize the RNG with 512 bits of RDRAND/RDSEED. We only need 256 bits
of those to succeed in order to initialize the RNG. Instead of the
current "all or nothing" approach, actually credit these contributions
the amount that is actually contributed.

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: handle latent entropy and command line from random_init()
Jason A. Donenfeld [Thu, 5 May 2022 00:20:22 +0000 (02:20 +0200)]
random: handle latent entropy and command line from random_init()

commit 2f14062bb14b0fcfcc21e6dc7d5b5c0d25966164 upstream.

Currently, start_kernel() adds latent entropy and the command line to
the entropy bool *after* the RNG has been initialized, deferring when
it's actually used by things like stack canaries until the next time
the pool is seeded. This surely is not intended.

Rather than splitting up which entropy gets added where and when between
start_kernel() and random_init(), just do everything in random_init(),
which should eliminate these kinds of bugs in the future.

While we're at it, rename the awkwardly titled "rand_initialize()" to
the more standard "random_init()" nomenclature.

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: use proper jiffies comparison macro
Jason A. Donenfeld [Tue, 10 May 2022 13:20:42 +0000 (15:20 +0200)]
random: use proper jiffies comparison macro

commit 8a5b8a4a4ceb353b4dd5bafd09e2b15751bcdb51 upstream.

This expands to exactly the same code that it replaces, but makes things
consistent by using the same macro for jiffy comparisons throughout.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: remove ratelimiting for in-kernel unseeded randomness
Jason A. Donenfeld [Mon, 9 May 2022 14:13:18 +0000 (16:13 +0200)]
random: remove ratelimiting for in-kernel unseeded randomness

commit cc1e127bfa95b5fb2f9307e7168bf8b2b45b4c5e upstream.

The CONFIG_WARN_ALL_UNSEEDED_RANDOM debug option controls whether the
kernel warns about all unseeded randomness or just the first instance.
There's some complicated rate limiting and comparison to the previous
caller, such that even with CONFIG_WARN_ALL_UNSEEDED_RANDOM enabled,
developers still don't see all the messages or even an accurate count of
how many were missed. This is the result of basically parallel
mechanisms aimed at accomplishing more or less the same thing, added at
different points in random.c history, which sort of compete with the
first-instance-only limiting we have now.

It turns out, however, that nobody cares about the first unseeded
randomness instance of in-kernel users. The same first user has been
there for ages now, and nobody is doing anything about it. It isn't even
clear that anybody _can_ do anything about it. Most places that can do
something about it have switched over to using get_random_bytes_wait()
or wait_for_random_bytes(), which is the right thing to do, but there is
still much code that needs randomness sometimes during init, and as a
geeneral rule, if you're not using one of the _wait functions or the
readiness notifier callback, you're bound to be doing it wrong just
based on that fact alone.

So warning about this same first user that can't easily change is simply
not an effective mechanism for anything at all. Users can't do anything
about it, as the Kconfig text points out -- the problem isn't in
userspace code -- and kernel developers don't or more often can't react
to it.

Instead, show the warning for all instances when CONFIG_WARN_ALL_UNSEEDED_RANDOM
is set, so that developers can debug things need be, or if it isn't set,
don't show a warning at all.

At the same time, CONFIG_WARN_ALL_UNSEEDED_RANDOM now implies setting
random.ratelimit_disable=1 on by default, since if you care about one
you probably care about the other too. And we can clean up usage around
the related urandom_warning ratelimiter as well (whose behavior isn't
changing), so that it properly counts missed messages after the 10
message threshold is reached.

Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: move initialization out of reseeding hot path
Jason A. Donenfeld [Mon, 9 May 2022 11:53:24 +0000 (13:53 +0200)]
random: move initialization out of reseeding hot path

commit 68c9c8b192c6dae9be6278e98ee44029d5da2d31 upstream.

Initialization happens once -- by way of credit_init_bits() -- and then
it never happens again. Therefore, it doesn't need to be in
crng_reseed(), which is a hot path that is called multiple times. It
also doesn't make sense to have there, as initialization activity is
better associated with initialization routines.

After the prior commit, crng_reseed() now won't be called by multiple
concurrent callers, which means that we can safely move the
"finialize_init" logic into crng_init_bits() unconditionally.

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: avoid initializing twice in credit race
Jason A. Donenfeld [Mon, 9 May 2022 11:40:55 +0000 (13:40 +0200)]
random: avoid initializing twice in credit race

commit fed7ef061686cc813b1f3d8d0edc6c35b4d3537b upstream.

Since all changes of crng_init now go through credit_init_bits(), we can
fix a long standing race in which two concurrent callers of
credit_init_bits() have the new bit count >= some threshold, but are
doing so with crng_init as a lower threshold, checked outside of a lock,
resulting in crng_reseed() or similar being called twice.

In order to fix this, we can use the original cmpxchg value of the bit
count, and only change crng_init when the bit count transitions from
below a threshold to meeting the threshold.

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: use symbolic constants for crng_init states
Jason A. Donenfeld [Sun, 8 May 2022 11:20:30 +0000 (13:20 +0200)]
random: use symbolic constants for crng_init states

commit e3d2c5e79a999aa4e7d6f0127e16d3da5a4ff70d upstream.

crng_init represents a state machine, with three states, and various
rules for transitions. For the longest time, we've been managing these
with "0", "1", and "2", and expecting people to figure it out. To make
the code more obvious, replace these with proper enum values
representing the transition, and then redocument what each of these
states mean.

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Cc: Joe Perches <joe@perches.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agosiphash: use one source of truth for siphash permutations
Jason A. Donenfeld [Sat, 7 May 2022 12:03:46 +0000 (14:03 +0200)]
siphash: use one source of truth for siphash permutations

commit e73aaae2fa9024832e1f42e30c787c7baf61d014 upstream.

The SipHash family of permutations is currently used in three places:

- siphash.c itself, used in the ordinary way it was intended.
- random32.c, in a construction from an anonymous contributor.
- random.c, as part of its fast_mix function.

Each one of these places reinvents the wheel with the same C code, same
rotation constants, and same symmetry-breaking constants.

This commit tidies things up a bit by placing macros for the
permutations and constants into siphash.h, where each of the three .c
users can access them. It also leaves a note dissuading more users of
them from emerging.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: help compiler out with fast_mix() by using simpler arguments
Jason A. Donenfeld [Fri, 6 May 2022 21:19:43 +0000 (23:19 +0200)]
random: help compiler out with fast_mix() by using simpler arguments

commit 791332b3cbb080510954a4c152ce02af8832eac9 upstream.

Now that fast_mix() has more than one caller, gcc no longer inlines it.
That's fine. But it also doesn't handle the compound literal argument we
pass it very efficiently, nor does it handle the loop as well as it
could. So just expand the code to spell out this function so that it
generates the same code as it did before. Performance-wise, this now
behaves as it did before the last commit. The difference in actual code
size on x86 is 45 bytes, which is less than a cache line.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: do not use input pool from hard IRQs
Jason A. Donenfeld [Fri, 6 May 2022 16:30:51 +0000 (18:30 +0200)]
random: do not use input pool from hard IRQs

commit e3e33fc2ea7fcefd0d761db9d6219f83b4248f5c upstream.

Years ago, a separate fast pool was added for interrupts, so that the
cost associated with taking the input pool spinlocks and mixing into it
would be avoided in places where latency is critical. However, one
oversight was that add_input_randomness() and add_disk_randomness()
still sometimes are called directly from the interrupt handler, rather
than being deferred to a thread. This means that some unlucky interrupts
will be caught doing a blake2s_compress() call and potentially spinning
on input_pool.lock, which can also be taken by unprivileged users by
writing into /dev/urandom.

In order to fix this, add_timer_randomness() now checks whether it is
being called from a hard IRQ and if so, just mixes into the per-cpu IRQ
fast pool using fast_mix(), which is much faster and can be done
lock-free. A nice consequence of this, as well, is that it means hard
IRQ context FPU support is likely no longer useful.

The entropy estimation algorithm used by add_timer_randomness() is also
somewhat different than the one used for add_interrupt_randomness(). The
former looks at deltas of deltas of deltas, while the latter just waits
for 64 interrupts for one bit or for one second since the last bit. In
order to bridge these, and since add_interrupt_randomness() runs after
an add_timer_randomness() that's called from hard IRQ, we add to the
fast pool credit the related amount, and then subtract one to account
for add_interrupt_randomness()'s contribution.

A downside of this, however, is that the num argument is potentially
attacker controlled, which puts a bit more pressure on the fast_mix()
sponge to do more than it's really intended to do. As a mitigating
factor, the first 96 bits of input aren't attacker controlled (a cycle
counter followed by zeros), which means it's essentially two rounds of
siphash rather than one, which is somewhat better. It's also not that
much different from add_interrupt_randomness()'s use of the irq stack
instruction pointer register.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Filipe Manana <fdmanana@suse.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: order timer entropy functions below interrupt functions
Jason A. Donenfeld [Fri, 6 May 2022 16:27:38 +0000 (18:27 +0200)]
random: order timer entropy functions below interrupt functions

commit a4b5c26b79ffdfcfb816c198f2fc2b1e7b5b580f upstream.

There are no code changes here; this is just a reordering of functions,
so that in subsequent commits, the timer entropy functions can call into
the interrupt ones.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: do not pretend to handle premature next security model
Jason A. Donenfeld [Sat, 30 Apr 2022 20:03:29 +0000 (22:03 +0200)]
random: do not pretend to handle premature next security model

commit e85c0fc1d94c52483a603651748d4c76d6aa1c6b upstream.

Per the thread linked below, "premature next" is not considered to be a
realistic threat model, and leads to more serious security problems.

"Premature next" is the scenario in which:

- Attacker compromises the current state of a fully initialized RNG via
  some kind of infoleak.
- New bits of entropy are added directly to the key used to generate the
  /dev/urandom stream, without any buffering or pooling.
- Attacker then, somehow having read access to /dev/urandom, samples RNG
  output and brute forces the individual new bits that were added.
- Result: the RNG never "recovers" from the initial compromise, a
  so-called violation of what academics term "post-compromise security".

The usual solutions to this involve some form of delaying when entropy
gets mixed into the crng. With Fortuna, this involves multiple input
buckets. With what the Linux RNG was trying to do prior, this involves
entropy estimation.

However, by delaying when entropy gets mixed in, it also means that RNG
compromises are extremely dangerous during the window of time before
the RNG has gathered enough entropy, during which time nonces may become
predictable (or repeated), ephemeral keys may not be secret, and so
forth. Moreover, it's unclear how realistic "premature next" is from an
attack perspective, if these attacks even make sense in practice.

Put together -- and discussed in more detail in the thread below --
these constitute grounds for just doing away with the current code that
pretends to handle premature next. I say "pretends" because it wasn't
doing an especially great job at it either; should we change our mind
about this direction, we would probably implement Fortuna to "fix" the
"problem", in which case, removing the pretend solution still makes
sense.

This also reduces the crng reseed period from 5 minutes down to 1
minute. The rationale from the thread might lead us toward reducing that
even further in the future (or even eliminating it), but that remains a
topic of a future commit.

At a high level, this patch changes semantics from:

    Before: Seed for the first time after 256 "bits" of estimated
    entropy have been accumulated since the system booted. Thereafter,
    reseed once every five minutes, but only if 256 new "bits" have been
    accumulated since the last reseeding.

    After: Seed for the first time after 256 "bits" of estimated entropy
    have been accumulated since the system booted. Thereafter, reseed
    once every minute.

Most of this patch is renaming and removing: POOL_MIN_BITS becomes
POOL_INIT_BITS, credit_entropy_bits() becomes credit_init_bits(),
crng_reseed() loses its "force" parameter since it's now always true,
the drain_entropy() function no longer has any use so it's removed,
entropy estimation is skipped if we've already init'd, the various
notifiers for "low on entropy" are now only active prior to init, and
finally, some documentation comments are cleaned up here and there.

Link: https://lore.kernel.org/lkml/YmlMGx6+uigkGiZ0@zx2c4.com/
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Nadia Heninger <nadiah@cs.ucsd.edu>
Cc: Tom Ristenpart <ristenpart@cornell.edu>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: use first 128 bits of input as fast init
Jason A. Donenfeld [Sat, 30 Apr 2022 13:08:20 +0000 (15:08 +0200)]
random: use first 128 bits of input as fast init

commit 5c3b747ef54fa2a7318776777f6044540d99f721 upstream.

Before, the first 64 bytes of input, regardless of how entropic it was,
would be used to mutate the crng base key directly, and none of those
bytes would be credited as having entropy. Then 256 bits of credited
input would be accumulated, and only then would the rng transition from
the earlier "fast init" phase into being actually initialized.

The thinking was that by mixing and matching fast init and real init, an
attacker who compromised the fast init state, considered easy to do
given how little entropy might be in those first 64 bytes, would then be
able to bruteforce bits from the actual initialization. By keeping these
separate, bruteforcing became impossible.

However, by not crediting potentially creditable bits from those first 64
bytes of input, we delay initialization, and actually make the problem
worse, because it means the user is drawing worse random numbers for a
longer period of time.

Instead, we can take the first 128 bits as fast init, and allow them to
be credited, and then hold off on the next 128 bits until they've
accumulated. This is still a wide enough margin to prevent bruteforcing
the rng state, while still initializing much faster.

Then, rather than trying to piecemeal inject into the base crng key at
various points, instead just extract from the pool when we need it, for
the crng_init==0 phase. Performance may even be better for the various
inputs here, since there are likely more calls to mix_pool_bytes() then
there are to get_random_bytes() during this phase of system execution.

Since the preinit injection code is gone, bootloader randomness can then
do something significantly more straight forward, removing the weird
system_wq hack in hwgenerator randomness.

Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: do not use batches when !crng_ready()
Jason A. Donenfeld [Tue, 3 May 2022 12:14:32 +0000 (14:14 +0200)]
random: do not use batches when !crng_ready()

commit cbe89e5a375a51bbb952929b93fa973416fea74e upstream.

It's too hard to keep the batches synchronized, and pointless anyway,
since in !crng_ready(), we're updating the base_crng key really often,
where batching only hurts. So instead, if the crng isn't ready, just
call into get_random_bytes(). At this stage nothing is performance
critical anyhow.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: insist on random_get_entropy() existing in order to simplify
Jason A. Donenfeld [Tue, 12 Apr 2022 17:59:57 +0000 (19:59 +0200)]
random: insist on random_get_entropy() existing in order to simplify

commit 4b758eda851eb9336ca86a0041a4d3da55f66511 upstream.

All platforms are now guaranteed to provide some value for
random_get_entropy(). In case some bug leads to this not being so, we
print a warning, because that indicates that something is really very
wrong (and likely other things are impacted too). This should never be
hit, but it's a good and cheap way of finding out if something ever is
problematic.

Since we now have viable fallback code for random_get_entropy() on all
platforms, which is, in the worst case, not worse than jiffies, we can
count on getting the best possible value out of it. That means there's
no longer a use for using jiffies as entropy input. It also means we no
longer have a reason for doing the round-robin register flow in the IRQ
handler, which was always of fairly dubious value.

Instead we can greatly simplify the IRQ handler inputs and also unify
the construction between 64-bits and 32-bits. We now collect the cycle
counter and the return address, since those are the two things that
matter. Because the return address and the irq number are likely
related, to the extent we mix in the irq number, we can just xor it into
the top unchanging bytes of the return address, rather than the bottom
changing bytes of the cycle counter as before. Then, we can do a fixed 2
rounds of SipHash/HSipHash. Finally, we use the same construction of
hashing only half of the [H]SipHash state on 32-bit and 64-bit. We're
not actually discarding any entropy, since that entropy is carried
through until the next time. And more importantly, it lets us do the
same sponge-like construction everywhere.

Cc: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoxtensa: use fallback for random_get_entropy() instead of zero
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
xtensa: use fallback for random_get_entropy() instead of zero

commit e10e2f58030c5c211d49042a8c2a1b93d40b2ffb upstream.

In the event that random_get_entropy() can't access a cycle counter or
similar, falling back to returning 0 is really not the best we can do.
Instead, at least calling random_get_entropy_fallback() would be
preferable, because that always needs to return _something_, even
falling back to jiffies eventually. It's not as though
random_get_entropy_fallback() is super high precision or guaranteed to
be entropic, but basically anything that's not zero all the time is
better than returning zero all the time.

This is accomplished by just including the asm-generic code like on
other architectures, which means we can get rid of the empty stub
function here.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Acked-by: Max Filippov <jcmvbkbc@gmail.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agosparc: use fallback for random_get_entropy() instead of zero
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
sparc: use fallback for random_get_entropy() instead of zero

commit ac9756c79797bb98972736b13cfb239fd2cffb79 upstream.

In the event that random_get_entropy() can't access a cycle counter or
similar, falling back to returning 0 is really not the best we can do.
Instead, at least calling random_get_entropy_fallback() would be
preferable, because that always needs to return _something_, even
falling back to jiffies eventually. It's not as though
random_get_entropy_fallback() is super high precision or guaranteed to
be entropic, but basically anything that's not zero all the time is
better than returning zero all the time.

This is accomplished by just including the asm-generic code like on
other architectures, which means we can get rid of the empty stub
function here.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoum: use fallback for random_get_entropy() instead of zero
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
um: use fallback for random_get_entropy() instead of zero

commit 9f13fb0cd11ed2327abff69f6501a2c124c88b5a upstream.

In the event that random_get_entropy() can't access a cycle counter or
similar, falling back to returning 0 is really not the best we can do.
Instead, at least calling random_get_entropy_fallback() would be
preferable, because that always needs to return _something_, even
falling back to jiffies eventually. It's not as though
random_get_entropy_fallback() is super high precision or guaranteed to
be entropic, but basically anything that's not zero all the time is
better than returning zero all the time.

This is accomplished by just including the asm-generic code like on
other architectures, which means we can get rid of the empty stub
function here.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Richard Weinberger <richard@nod.at>
Cc: Anton Ivanov <anton.ivanov@cambridgegreys.com>
Acked-by: Johannes Berg <johannes@sipsolutions.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agox86/tsc: Use fallback for random_get_entropy() instead of zero
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
x86/tsc: Use fallback for random_get_entropy() instead of zero

commit 3bd4abc07a267e6a8b33d7f8717136e18f921c53 upstream.

In the event that random_get_entropy() can't access a cycle counter or
similar, falling back to returning 0 is suboptimal. Instead, fallback
to calling random_get_entropy_fallback(), which isn't extremely high
precision or guaranteed to be entropic, but is certainly better than
returning zero all the time.

If CONFIG_X86_TSC=n, then it's possible for the kernel to run on systems
without RDTSC, such as 486 and certain 586, so the fallback code is only
required for that case.

As well, fix up both the new function and the get_cycles() function from
which it was derived to use cpu_feature_enabled() rather than
boot_cpu_has(), and use !IS_ENABLED() instead of #ifndef.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: x86@kernel.org
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agonios2: use fallback for random_get_entropy() instead of zero
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
nios2: use fallback for random_get_entropy() instead of zero

commit c04e72700f2293013dab40208e809369378f224c upstream.

In the event that random_get_entropy() can't access a cycle counter or
similar, falling back to returning 0 is really not the best we can do.
Instead, at least calling random_get_entropy_fallback() would be
preferable, because that always needs to return _something_, even
falling back to jiffies eventually. It's not as though
random_get_entropy_fallback() is super high precision or guaranteed to
be entropic, but basically anything that's not zero all the time is
better than returning zero all the time.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Acked-by: Dinh Nguyen <dinguyen@kernel.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoarm: use fallback for random_get_entropy() instead of zero
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
arm: use fallback for random_get_entropy() instead of zero

commit ff8a8f59c99f6a7c656387addc4d9f2247d75077 upstream.

In the event that random_get_entropy() can't access a cycle counter or
similar, falling back to returning 0 is really not the best we can do.
Instead, at least calling random_get_entropy_fallback() would be
preferable, because that always needs to return _something_, even
falling back to jiffies eventually. It's not as though
random_get_entropy_fallback() is super high precision or guaranteed to
be entropic, but basically anything that's not zero all the time is
better than returning zero all the time.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agomips: use fallback for random_get_entropy() instead of just c0 random
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
mips: use fallback for random_get_entropy() instead of just c0 random

commit 1c99c6a7c3c599a68321b01b9ec243215ede5a68 upstream.

For situations in which we don't have a c0 counter register available,
we've been falling back to reading the c0 "random" register, which is
usually bounded by the amount of TLB entries and changes every other
cycle or so. This means it wraps extremely often. We can do better by
combining this fast-changing counter with a potentially slower-changing
counter from random_get_entropy_fallback() in the more significant bits.
This commit combines the two, taking into account that the changing bits
are in a different bit position depending on the CPU model. In addition,
we previously were falling back to 0 for ancient CPUs that Linux does
not support anyway; remove that dead path entirely.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Tested-by: Maciej W. Rozycki <macro@orcam.me.uk>
Acked-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoriscv: use fallback for random_get_entropy() instead of zero
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
riscv: use fallback for random_get_entropy() instead of zero

commit 6d01238623faa9425f820353d2066baf6c9dc872 upstream.

In the event that random_get_entropy() can't access a cycle counter or
similar, falling back to returning 0 is really not the best we can do.
Instead, at least calling random_get_entropy_fallback() would be
preferable, because that always needs to return _something_, even
falling back to jiffies eventually. It's not as though
random_get_entropy_fallback() is super high precision or guaranteed to
be entropic, but basically anything that's not zero all the time is
better than returning zero all the time.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Acked-by: Palmer Dabbelt <palmer@rivosinc.com>
Reviewed-by: Palmer Dabbelt <palmer@rivosinc.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agom68k: use fallback for random_get_entropy() instead of zero
Jason A. Donenfeld [Fri, 8 Apr 2022 16:03:13 +0000 (18:03 +0200)]
m68k: use fallback for random_get_entropy() instead of zero

commit 0f392c95391f2d708b12971a07edaa7973f9eece upstream.

In the event that random_get_entropy() can't access a cycle counter or
similar, falling back to returning 0 is really not the best we can do.
Instead, at least calling random_get_entropy_fallback() would be
preferable, because that always needs to return _something_, even
falling back to jiffies eventually. It's not as though
random_get_entropy_fallback() is super high precision or guaranteed to
be entropic, but basically anything that's not zero all the time is
better than returning zero all the time.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Acked-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agotimekeeping: Add raw clock fallback for random_get_entropy()
Jason A. Donenfeld [Sun, 10 Apr 2022 14:49:50 +0000 (16:49 +0200)]
timekeeping: Add raw clock fallback for random_get_entropy()

commit 1366992e16bddd5e2d9a561687f367f9f802e2e4 upstream.

The addition of random_get_entropy_fallback() provides access to
whichever time source has the highest frequency, which is useful for
gathering entropy on platforms without available cycle counters. It's
not necessarily as good as being able to quickly access a cycle counter
that the CPU has, but it's still something, even when it falls back to
being jiffies-based.

In the event that a given arch does not define get_cycles(), falling
back to the get_cycles() default implementation that returns 0 is really
not the best we can do. Instead, at least calling
random_get_entropy_fallback() would be preferable, because that always
needs to return _something_, even falling back to jiffies eventually.
It's not as though random_get_entropy_fallback() is super high precision
or guaranteed to be entropic, but basically anything that's not zero all
the time is better than returning zero all the time.

Finally, since random_get_entropy_fallback() is used during extremely
early boot when randomizing freelists in mm_init(), it can be called
before timekeeping has been initialized. In that case there really is
nothing we can do; jiffies hasn't even started ticking yet. So just give
up and return 0.

Suggested-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agopowerpc: define get_cycles macro for arch-override
Jason A. Donenfeld [Sat, 23 Apr 2022 19:11:41 +0000 (21:11 +0200)]
powerpc: define get_cycles macro for arch-override

commit 408835832158df0357e18e96da7f2d1ed6b80e7f upstream.

PowerPC defines a get_cycles() function, but it does not do the usual
`#define get_cycles get_cycles` dance, making it impossible for generic
code to see if an arch-specific function was defined. While the
get_cycles() ifdef is not currently used, the following timekeeping
patch in this series will depend on the macro existing (or not existing)
when defining random_get_entropy().

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@ozlabs.org>
Cc: Paul Mackerras <paulus@samba.org>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoalpha: define get_cycles macro for arch-override
Jason A. Donenfeld [Sat, 23 Apr 2022 19:11:41 +0000 (21:11 +0200)]
alpha: define get_cycles macro for arch-override

commit 1097710bc9660e1e588cf2186a35db3d95c4d258 upstream.

Alpha defines a get_cycles() function, but it does not do the usual
`#define get_cycles get_cycles` dance, making it impossible for generic
code to see if an arch-specific function was defined. While the
get_cycles() ifdef is not currently used, the following timekeeping
patch in this series will depend on the macro existing (or not existing)
when defining random_get_entropy().

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Acked-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoparisc: define get_cycles macro for arch-override
Jason A. Donenfeld [Sat, 23 Apr 2022 19:11:41 +0000 (21:11 +0200)]
parisc: define get_cycles macro for arch-override

commit 8865bbe6ba1120e67f72201b7003a16202cd42be upstream.

PA-RISC defines a get_cycles() function, but it does not do the usual
`#define get_cycles get_cycles` dance, making it impossible for generic
code to see if an arch-specific function was defined. While the
get_cycles() ifdef is not currently used, the following timekeeping
patch in this series will depend on the macro existing (or not existing)
when defining random_get_entropy().

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Acked-by: Helge Deller <deller@gmx.de>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agos390: define get_cycles macro for arch-override
Jason A. Donenfeld [Sat, 23 Apr 2022 19:11:41 +0000 (21:11 +0200)]
s390: define get_cycles macro for arch-override

commit 2e3df523256cb9836de8441e9c791a796759bb3c upstream.

S390x defines a get_cycles() function, but it does not do the usual
`#define get_cycles get_cycles` dance, making it impossible for generic
code to see if an arch-specific function was defined. While the
get_cycles() ifdef is not currently used, the following timekeeping
patch in this series will depend on the macro existing (or not existing)
when defining random_get_entropy().

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Christian Borntraeger <borntraeger@linux.ibm.com>
Cc: Sven Schnelle <svens@linux.ibm.com>
Acked-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoia64: define get_cycles macro for arch-override
Jason A. Donenfeld [Sat, 23 Apr 2022 19:11:41 +0000 (21:11 +0200)]
ia64: define get_cycles macro for arch-override

commit 57c0900b91d8891ab43f0e6b464d059fda51d102 upstream.

Itanium defines a get_cycles() function, but it does not do the usual
`#define get_cycles get_cycles` dance, making it impossible for generic
code to see if an arch-specific function was defined. While the
get_cycles() ifdef is not currently used, the following timekeeping
patch in this series will depend on the macro existing (or not existing)
when defining random_get_entropy().

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoinit: call time_init() before rand_initialize()
Jason A. Donenfeld [Thu, 5 May 2022 00:20:22 +0000 (02:20 +0200)]
init: call time_init() before rand_initialize()

commit fe222a6ca2d53c38433cba5d3be62a39099e708e upstream.

Currently time_init() is called after rand_initialize(), but
rand_initialize() makes use of the timer on various platforms, and
sometimes this timer needs to be initialized by time_init() first. In
order for random_get_entropy() to not return zero during early boot when
it's potentially used as an entropy source, reverse the order of these
two calls. The block doing random initialization was right before
time_init() before, so changing the order shouldn't have any complicated
effects.

Cc: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Stafford Horne <shorne@gmail.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: fix sysctl documentation nits
Jason A. Donenfeld [Tue, 3 May 2022 19:43:58 +0000 (21:43 +0200)]
random: fix sysctl documentation nits

commit 069c4ea6871c18bd368f27756e0f91ffb524a788 upstream.

A semicolon was missing, and the almost-alphabetical-but-not ordering
was confusing, so regroup these by category instead.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: document crng_fast_key_erasure() destination possibility
Jason A. Donenfeld [Mon, 18 Apr 2022 18:57:31 +0000 (20:57 +0200)]
random: document crng_fast_key_erasure() destination possibility

commit 8717627d6ac53251ee012c3c7aca392f29f38a42 upstream.

This reverts 35a33ff3807d ("random: use memmove instead of memcpy for
remaining 32 bytes"), which was made on a totally bogus basis. The thing
it was worried about overlapping came from the stack, not from one of
its arguments, as Eric pointed out.

But the fact that this confusion even happened draws attention to the
fact that it's a bit non-obvious that the random_data parameter can
alias chacha_state, and in fact should do so when the caller can't rely
on the stack being cleared in a timely manner. So this commit documents
that.

Reported-by: Eric Biggers <ebiggers@kernel.org>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: make random_get_entropy() return an unsigned long
Jason A. Donenfeld [Fri, 8 Apr 2022 16:14:57 +0000 (18:14 +0200)]
random: make random_get_entropy() return an unsigned long

commit b0c3e796f24b588b862b61ce235d3c9417dc8983 upstream.

Some implementations were returning type `unsigned long`, while others
that fell back to get_cycles() were implicitly returning a `cycles_t` or
an untyped constant int literal. That makes for weird and confusing
code, and basically all code in the kernel already handled it like it
was an `unsigned long`. I recently tried to handle it as the largest
type it could be, a `cycles_t`, but doing so doesn't really help with
much.

Instead let's just make random_get_entropy() return an unsigned long all
the time. This also matches the commonly used `arch_get_random_long()`
function, so now RDRAND and RDTSC return the same sized integer, which
means one can fallback to the other more gracefully.

Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Cc: Theodore Ts'o <tytso@mit.edu>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: allow partial reads if later user copies fail
Jason A. Donenfeld [Thu, 7 Apr 2022 19:23:08 +0000 (21:23 +0200)]
random: allow partial reads if later user copies fail

commit 5209aed5137880fa229746cb521f715e55596460 upstream.

Rather than failing entirely if a copy_to_user() fails at some point,
instead we should return a partial read for the amount that succeeded
prior, unless none succeeded at all, in which case we return -EFAULT as
before.

This makes it consistent with other reader interfaces. For example, the
following snippet for /dev/zero outputs "4" followed by "1":

  int fd;
  void *x = mmap(NULL, 4096, PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
  assert(x != MAP_FAILED);
  fd = open("/dev/zero", O_RDONLY);
  assert(fd >= 0);
  printf("%zd\n", read(fd, x, 4));
  printf("%zd\n", read(fd, x + 4095, 4));
  close(fd);

This brings that same standard behavior to the various RNG reader
interfaces.

While we're at it, we can streamline the loop logic a little bit.

Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Jann Horn <jannh@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: check for signals every PAGE_SIZE chunk of /dev/[u]random
Jason A. Donenfeld [Wed, 6 Apr 2022 00:36:16 +0000 (02:36 +0200)]
random: check for signals every PAGE_SIZE chunk of /dev/[u]random

commit e3c1c4fd9e6d14059ed93ebfe15e1c57793b1a05 upstream.

In 1448769c9cdb ("random: check for signal_pending() outside of
need_resched() check"), Jann pointed out that we previously were only
checking the TIF_NOTIFY_SIGNAL and TIF_SIGPENDING flags if the process
had TIF_NEED_RESCHED set, which meant in practice, super long reads to
/dev/[u]random would delay signal handling by a long time. I tried this
using the below program, and indeed I wasn't able to interrupt a
/dev/urandom read until after several megabytes had been read. The bug
he fixed has always been there, and so code that reads from /dev/urandom
without checking the return value of read() has mostly worked for a long
time, for most sizes, not just for <= 256.

Maybe it makes sense to keep that code working. The reason it was so
small prior, ignoring the fact that it didn't work anyway, was likely
because /dev/random used to block, and that could happen for pretty
large lengths of time while entropy was gathered. But now, it's just a
chacha20 call, which is extremely fast and is just operating on pure
data, without having to wait for some external event. In that sense,
/dev/[u]random is a lot more like /dev/zero.

Taking a page out of /dev/zero's read_zero() function, it always returns
at least one chunk, and then checks for signals after each chunk. Chunk
sizes there are of length PAGE_SIZE. Let's just copy the same thing for
/dev/[u]random, and check for signals and cond_resched() for every
PAGE_SIZE amount of data. This makes the behavior more consistent with
expectations, and should mitigate the impact of Jann's fix for the
age-old signal check bug.

---- test program ----

  #include <unistd.h>
  #include <signal.h>
  #include <stdio.h>
  #include <sys/random.h>

  static unsigned char x[~0U];

  static void handle(int) { }

  int main(int argc, char *argv[])
  {
    pid_t pid = getpid(), child;
    signal(SIGUSR1, handle);
    if (!(child = fork())) {
      for (;;)
        kill(pid, SIGUSR1);
    }
    pause();
    printf("interrupted after reading %zd bytes\n", getrandom(x, sizeof(x), 0));
    kill(child, SIGTERM);
    return 0;
  }

Cc: Jann Horn <jannh@google.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: check for signal_pending() outside of need_resched() check
Jann Horn [Tue, 5 Apr 2022 16:39:31 +0000 (18:39 +0200)]
random: check for signal_pending() outside of need_resched() check

commit 1448769c9cdb69ad65287f4f7ab58bc5f2f5d7ba upstream.

signal_pending() checks TIF_NOTIFY_SIGNAL and TIF_SIGPENDING, which
signal that the task should bail out of the syscall when possible. This
is a separate concept from need_resched(), which checks
TIF_NEED_RESCHED, signaling that the task should preempt.

In particular, with the current code, the signal_pending() bailout
probably won't work reliably.

Change this to look like other functions that read lots of data, such as
read_zero().

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: do not allow user to keep crng key around on stack
Jason A. Donenfeld [Tue, 5 Apr 2022 14:40:51 +0000 (16:40 +0200)]
random: do not allow user to keep crng key around on stack

commit aba120cc101788544aa3e2c30c8da88513892350 upstream.

The fast key erasure RNG design relies on the key that's used to be used
and then discarded. We do this, making judicious use of
memzero_explicit().  However, reads to /dev/urandom and calls to
getrandom() involve a copy_to_user(), and userspace can use FUSE or
userfaultfd, or make a massive call, dynamically remap memory addresses
as it goes, and set the process priority to idle, in order to keep a
kernel stack alive indefinitely. By probing
/proc/sys/kernel/random/entropy_avail to learn when the crng key is
refreshed, a malicious userspace could mount this attack every 5 minutes
thereafter, breaking the crng's forward secrecy.

In order to fix this, we just overwrite the stack's key with the first
32 bytes of the "free" fast key erasure output. If we're returning <= 32
bytes to the user, then we can still return those bytes directly, so
that short reads don't become slower. And for long reads, the difference
is hopefully lost in the amortization, so it doesn't change much, with
that amortization helping variously for medium reads.

We don't need to do this for get_random_bytes() and the various
kernel-space callers, and later, if we ever switch to always batching,
this won't be necessary either, so there's no need to change the API of
these functions.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jann Horn <jannh@google.com>
Fixes: c92e040d575a ("random: add backtracking protection to the CRNG")
Fixes: 186873c549df ("random: use simpler fast key erasure flow on per-cpu keys")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: do not split fast init input in add_hwgenerator_randomness()
Jan Varho [Mon, 4 Apr 2022 16:42:30 +0000 (19:42 +0300)]
random: do not split fast init input in add_hwgenerator_randomness()

commit 527a9867af29ff89f278d037db704e0ed50fb666 upstream.

add_hwgenerator_randomness() tries to only use the required amount of input
for fast init, but credits all the entropy, rather than a fraction of
it. Since it's hard to determine how much entropy is left over out of a
non-unformly random sample, either give it all to fast init or credit
it, but don't attempt to do both. In the process, we can clean up the
injection code to no longer need to return a value.

Signed-off-by: Jan Varho <jan.varho@gmail.com>
[Jason: expanded commit message]
Fixes: 73c7733f122e ("random: do not throw away excess input to crng_fast_load")
Cc: stable@vger.kernel.org # 5.17+, requires af704c856e88
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: mix build-time latent entropy into pool at init
Jason A. Donenfeld [Thu, 31 Mar 2022 15:01:01 +0000 (11:01 -0400)]
random: mix build-time latent entropy into pool at init

commit 1754abb3e7583c570666fa1e1ee5b317e88c89a0 upstream.

Prior, the "input_pool_data" array needed no real initialization, and so
it was easy to mark it with __latent_entropy to populate it during
compile-time. In switching to using a hash function, this required us to
specifically initialize it to some specific state, which means we
dropped the __latent_entropy attribute. An unfortunate side effect was
this meant the pool was no longer seeded using compile-time random data.
In order to bring this back, we declare an array in rand_initialize()
with __latent_entropy and call mix_pool_bytes() on that at init, which
accomplishes the same thing as before. We make this __initconst, so that
it doesn't take up space at runtime after init.

Fixes: 6e8ec2552c7d ("random: use computational hash for entropy extraction")
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: re-add removed comment about get_random_{u32,u64} reseeding
Jason A. Donenfeld [Wed, 23 Mar 2022 04:21:52 +0000 (22:21 -0600)]
random: re-add removed comment about get_random_{u32,u64} reseeding

commit dd7aa36e535797926d8eb311da7151919130139d upstream.

The comment about get_random_{u32,u64}() not invoking reseeding got
added in an unrelated commit, that then was recently reverted by
0313bc278dac ("Revert "random: block in /dev/urandom""). So this adds
that little comment snippet back, and improves the wording a bit too.

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: treat bootloader trust toggle the same way as cpu trust toggle
Jason A. Donenfeld [Wed, 23 Mar 2022 03:43:12 +0000 (21:43 -0600)]
random: treat bootloader trust toggle the same way as cpu trust toggle

commit d97c68d178fbf8aaaf21b69b446f2dfb13909316 upstream.

If CONFIG_RANDOM_TRUST_CPU is set, the RNG initializes using RDRAND.
But, the user can disable (or enable) this behavior by setting
`random.trust_cpu=0/1` on the kernel command line. This allows system
builders to do reasonable things while avoiding howls from tinfoil
hatters. (Or vice versa.)

CONFIG_RANDOM_TRUST_BOOTLOADER is basically the same thing, but regards
the seed passed via EFI or device tree, which might come from RDRAND or
a TPM or somewhere else. In order to allow distros to more easily enable
this while avoiding those same howls (or vice versa), this commit adds
the corresponding `random.trust_bootloader=0/1` toggle.

Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Graham Christensen <graham@grahamc.com>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Link: https://github.com/NixOS/nixpkgs/pull/165355
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: skip fast_init if hwrng provides large chunk of entropy
Jason A. Donenfeld [Tue, 22 Mar 2022 00:48:05 +0000 (18:48 -0600)]
random: skip fast_init if hwrng provides large chunk of entropy

commit af704c856e888fb044b058d731d61b46eeec499d upstream.

At boot time, EFI calls add_bootloader_randomness(), which in turn calls
add_hwgenerator_randomness(). Currently add_hwgenerator_randomness()
feeds the first 64 bytes of randomness to the "fast init"
non-crypto-grade phase. But if add_hwgenerator_randomness() gets called
with more than POOL_MIN_BITS of entropy, there's no point in passing it
off to the "fast init" stage, since that's enough entropy to bootstrap
the real RNG. The "fast init" stage is just there to provide _something_
in the case where we don't have enough entropy to properly bootstrap the
RNG. But if we do have enough entropy to bootstrap the RNG, the current
logic doesn't serve a purpose. So, in the case where we're passed
greater than or equal to POOL_MIN_BITS of entropy, this commit makes us
skip the "fast init" phase.

Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: check for signal and try earlier when generating entropy
Jason A. Donenfeld [Tue, 8 Mar 2022 17:12:16 +0000 (10:12 -0700)]
random: check for signal and try earlier when generating entropy

commit 3e504d2026eb6c8762cd6040ae57db166516824a upstream.

Rather than waiting a full second in an interruptable waiter before
trying to generate entropy, try to generate entropy first and wait
second. While waiting one second might give an extra second for getting
entropy from elsewhere, we're already pretty late in the init process
here, and whatever else is generating entropy will still continue to
contribute. This has implications on signal handling: we call
try_to_generate_entropy() from wait_for_random_bytes(), and
wait_for_random_bytes() always uses wait_event_interruptible_timeout()
when waiting, since it's called by userspace code in restartable
contexts, where signals can pend. Since try_to_generate_entropy() now
runs first, if a signal is pending, it's necessary for
try_to_generate_entropy() to check for signals, since it won't hit the
wait until after try_to_generate_entropy() has returned. And even before
this change, when entering a busy loop in try_to_generate_entropy(), we
should have been checking to see if any signals are pending, so that a
process doesn't get stuck in that loop longer than expected.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: reseed more often immediately after booting
Jason A. Donenfeld [Wed, 9 Mar 2022 06:32:34 +0000 (23:32 -0700)]
random: reseed more often immediately after booting

commit 7a7ff644aeaf071d433caffb3b8ea57354b55bd3 upstream.

In order to chip away at the "premature first" problem, we augment our
existing entropy accounting with more frequent reseedings at boot.

The idea is that at boot, we're getting entropy from various places, and
we're not very sure which of early boot entropy is good and which isn't.
Even when we're crediting the entropy, we're still not totally certain
that it's any good. Since boot is the one time (aside from a compromise)
that we have zero entropy, it's important that we shepherd entropy into
the crng fairly often.

At the same time, we don't want a "premature next" problem, whereby an
attacker can brute force individual bits of added entropy. In lieu of
going full-on Fortuna (for now), we can pick a simpler strategy of just
reseeding more often during the first 5 minutes after boot. This is
still bounded by the 256-bit entropy credit requirement, so we'll skip a
reseeding if we haven't reached that, but in case entropy /is/ coming
in, this ensures that it makes its way into the crng rather rapidly
during these early stages.

Ordinarily we reseed if the previous reseeding is 300 seconds old. This
commit changes things so that for the first 600 seconds of boot time, we
reseed if the previous reseeding is uptime / 2 seconds old. That means
that we'll reseed at the very least double the uptime of the previous
reseeding.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: make consistent usage of crng_ready()
Jason A. Donenfeld [Tue, 8 Mar 2022 18:20:17 +0000 (11:20 -0700)]
random: make consistent usage of crng_ready()

commit a96cfe2d427064325ecbf56df8816c6b871ec285 upstream.

Rather than sometimes checking `crng_init < 2`, we should always use the
crng_ready() macro, so that should we change anything later, it's
consistent. Additionally, that macro already has a likely() around it,
which means we don't need to open code our own likely() and unlikely()
annotations.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: use SipHash as interrupt entropy accumulator
Jason A. Donenfeld [Fri, 11 Feb 2022 13:58:44 +0000 (14:58 +0100)]
random: use SipHash as interrupt entropy accumulator

commit f5eab0e2db4f881fb2b62b3fdad5b9be673dd7ae upstream.

The current fast_mix() function is a piece of classic mailing list
crypto, where it just sort of sprung up by an anonymous author without a
lot of real analysis of what precisely it was accomplishing. As an ARX
permutation alone, there are some easily searchable differential trails
in it, and as a means of preventing malicious interrupts, it completely
fails, since it xors new data into the entire state every time. It can't
really be analyzed as a random permutation, because it clearly isn't,
and it can't be analyzed as an interesting linear algebraic structure
either, because it's also not that. There really is very little one can
say about it in terms of entropy accumulation. It might diffuse bits,
some of the time, maybe, we hope, I guess. But for the most part, it
fails to accomplish anything concrete.

As a reminder, the simple goal of add_interrupt_randomness() is to
simply accumulate entropy until ~64 interrupts have elapsed, and then
dump it into the main input pool, which uses a cryptographic hash.

It would be nice to have something cryptographically strong in the
interrupt handler itself, in case a malicious interrupt compromises a
per-cpu fast pool within the 64 interrupts / 1 second window, and then
inside of that same window somehow can control its return address and
cycle counter, even if that's a bit far fetched. However, with a very
CPU-limited budget, actually doing that remains an active research
project (and perhaps there'll be something useful for Linux to come out
of it). And while the abundance of caution would be nice, this isn't
*currently* the security model, and we don't yet have a fast enough
solution to make it our security model. Plus there's not exactly a
pressing need to do that. (And for the avoidance of doubt, the actual
cluster of 64 accumulated interrupts still gets dumped into our
cryptographically secure input pool.)

So, for now we are going to stick with the existing interrupt security
model, which assumes that each cluster of 64 interrupt data samples is
mostly non-malicious and not colluding with an infoleaker. With this as
our goal, we have a few more choices, simply aiming to accumulate
entropy, while discarding the least amount of it.

We know from <https://eprint.iacr.org/2019/198> that random oracles,
instantiated as computational hash functions, make good entropy
accumulators and extractors, which is the justification for using
BLAKE2s in the main input pool. As mentioned, we don't have that luxury
here, but we also don't have the same security model requirements,
because we're assuming that there aren't malicious inputs. A
pseudorandom function instance can approximately behave like a random
oracle, provided that the key is uniformly random. But since we're not
concerned with malicious inputs, we can pick a fixed key, which is not
secret, knowing that "nature" won't interact with a sufficiently chosen
fixed key by accident. So we pick a PRF with a fixed initial key, and
accumulate into it continuously, dumping the result every 64 interrupts
into our cryptographically secure input pool.

For this, we make use of SipHash-1-x on 64-bit and HalfSipHash-1-x on
32-bit, which are already in use in the kernel's hsiphash family of
functions and achieve the same performance as the function they replace.
It would be nice to do two rounds, but we don't exactly have the CPU
budget handy for that, and one round alone is already sufficient.

As mentioned, we start with a fixed initial key (zeros is fine), and
allow SipHash's symmetry breaking constants to turn that into a useful
starting point. Also, since we're dumping the result (or half of it on
64-bit so as to tax our hash function the same amount on all platforms)
into the cryptographically secure input pool, there's no point in
finalizing SipHash's output, since it'll wind up being finalized by
something much stronger. This means that all we need to do is use the
ordinary round function word-by-word, as normal SipHash does.
Simplified, the flow is as follows:

Initialize:

    siphash_state_t state;
    siphash_init(&state, key={0, 0, 0, 0});

Update (accumulate) on interrupt:

    siphash_update(&state, interrupt_data_and_timing);

Dump into input pool after 64 interrupts:

    blake2s_update(&input_pool, &state, sizeof(state) / 2);

The result of all of this is that the security model is unchanged from
before -- we assume non-malicious inputs -- yet we now implement that
model with a stronger argument. I would like to emphasize, again, that
the purpose of this commit is to improve the existing design, by making
it analyzable, without changing any fundamental assumptions. There may
well be value down the road in changing up the existing design, using
something cryptographically strong, or simply using a ring buffer of
samples rather than having a fast_mix() at all, or changing which and
how much data we collect each interrupt so that we can use something
linear, or a variety of other ideas. This commit does not invalidate the
potential for those in the future.

For example, in the future, if we're able to characterize the data we're
collecting on each interrupt, we may be able to inch toward information
theoretic accumulators. <https://eprint.iacr.org/2021/523> shows that `s
= ror32(s, 7) ^ x` and `s = ror64(s, 19) ^ x` make very good
accumulators for 2-monotone distributions, which would apply to
timestamp counters, like random_get_entropy() or jiffies, but would not
apply to our current combination of the two values, or to the various
function addresses and register values we mix in. Alternatively,
<https://eprint.iacr.org/2021/1002> shows that max-period linear
functions with no non-trivial invariant subspace make good extractors,
used in the form `s = f(s) ^ x`. However, this only works if the input
data is both identical and independent, and obviously a collection of
address values and counters fails; so it goes with theoretical papers.
Future directions here may involve trying to characterize more precisely
what we actually need to collect in the interrupt handler, and building
something specific around that.

However, as mentioned, the morass of data we're gathering at the
interrupt handler presently defies characterization, and so we use
SipHash for now, which works well and performs well.

Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Jean-Philippe Aumasson <jeanphilippe.aumasson@gmail.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: replace custom notifier chain with standard one
Jason A. Donenfeld [Tue, 1 Mar 2022 19:03:49 +0000 (20:03 +0100)]
random: replace custom notifier chain with standard one

commit 5acd35487dc911541672b3ffc322851769c32a56 upstream.

We previously rolled our own randomness readiness notifier, which only
has two users in the whole kernel. Replace this with a more standard
atomic notifier block that serves the same purpose with less code. Also
unexport the symbols, because no modules use it, only unconditional
builtins. The only drawback is that it's possible for a notification
handler returning the "stop" code to prevent further processing, but
given that there are only two users, and that we're unexporting this
anyway, that doesn't seem like a significant drawback for the
simplification we receive here.

Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
[Jason: for stable, also backported to crypto/drbg.c, not unexporting.]
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: don't let 644 read-only sysctls be written to
Jason A. Donenfeld [Mon, 28 Feb 2022 13:00:52 +0000 (14:00 +0100)]
random: don't let 644 read-only sysctls be written to

commit 77553cf8f44863b31da242cf24671d76ddb61597 upstream.

We leave around these old sysctls for compatibility, and we keep them
"writable" for compatibility, but even after writing, we should keep
reporting the same value. This is consistent with how userspaces tend to
use sysctl_random_write_wakeup_bits, writing to it, and then later
reading from it and using the value.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: give sysctl_random_min_urandom_seed a more sensible value
Jason A. Donenfeld [Mon, 28 Feb 2022 12:57:57 +0000 (13:57 +0100)]
random: give sysctl_random_min_urandom_seed a more sensible value

commit d0efdf35a6a71d307a250199af6fce122a7c7e11 upstream.

This isn't used by anything or anywhere, but we can't delete it due to
compatibility. So at least give it the correct value of what it's
supposed to be instead of a garbage one.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: do crng pre-init loading in worker rather than irq
Jason A. Donenfeld [Sun, 13 Feb 2022 17:25:07 +0000 (18:25 +0100)]
random: do crng pre-init loading in worker rather than irq

commit c2a7de4feb6e09f23af7accc0f882a8fa92e7ae5 upstream.

Taking spinlocks from IRQ context is generally problematic for
PREEMPT_RT. That is, in part, why we take trylocks instead. However, a
spin_try_lock() is also problematic since another spin_lock() invocation
can potentially PI-boost the wrong task, as the spin_try_lock() is
invoked from an IRQ-context, so the task on CPU (random task or idle) is
not the actual owner.

Additionally, by deferring the crng pre-init loading to the worker, we
can use the cryptographic hash function rather than xor, which is
perhaps a meaningful difference when considering this data has only been
through the relatively weak fast_mix() function.

The biggest downside of this approach is that the pre-init loading is
now deferred until later, which means things that need random numbers
after interrupts are enabled, but before workqueues are running -- or
before this particular worker manages to run -- are going to get into
trouble. Hopefully in the real world, this window is rather small,
especially since this code won't run until 64 interrupts had occurred.

Cc: Sultan Alsawaf <sultan@kerneltoast.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Theodore Ts'o <tytso@mit.edu>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: unify cycles_t and jiffies usage and types
Jason A. Donenfeld [Thu, 24 Feb 2022 17:30:58 +0000 (18:30 +0100)]
random: unify cycles_t and jiffies usage and types

commit abded93ec1e9692920fe309f07f40bd1035f2940 upstream.

random_get_entropy() returns a cycles_t, not an unsigned long, which is
sometimes 64 bits on various 32-bit platforms, including x86.
Conversely, jiffies is always unsigned long. This commit fixes things to
use cycles_t for fields that use random_get_entropy(), named "cycles",
and unsigned long for fields that use jiffies, named "now". It's also
good to mix in a cycles_t and a jiffies in the same way for both
add_device_randomness and add_timer_randomness, rather than using xor in
one case. Finally, we unify the order of these volatile reads, always
reading the more precise cycles counter, and then jiffies, so that the
cycle counter is as close to the event as possible.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: cleanup UUID handling
Jason A. Donenfeld [Thu, 24 Feb 2022 22:04:56 +0000 (23:04 +0100)]
random: cleanup UUID handling

commit 64276a9939ff414f2f0db38036cf4e1a0a703394 upstream.

Rather than hard coding various lengths, we can use the right constants.
Strings should be `char *` while buffers should be `u8 *`. Rather than
have a nonsensical and unused maxlength, just remove it. Finally, use
snprintf instead of sprintf, just out of good hygiene.

As well, remove the old comment about returning a binary UUID via the
binary sysctl syscall. That syscall was removed from the kernel in 5.5,
and actually, the "uuid_strategy" function and related infrastructure
for even serving it via the binary sysctl syscall was removed with
894d2491153a ("sysctl drivers: Remove dead binary sysctl support") back
in 2.6.33.

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: only wake up writers after zap if threshold was passed
Jason A. Donenfeld [Tue, 22 Feb 2022 13:01:57 +0000 (14:01 +0100)]
random: only wake up writers after zap if threshold was passed

commit a3f9e8910e1584d7725ef7d5ac870920d42d0bb4 upstream.

The only time that we need to wake up /dev/random writers on
RNDCLEARPOOL/RNDZAPPOOL is when we're changing from a value that is
greater than or equal to POOL_MIN_BITS to zero, because if we're
changing from below POOL_MIN_BITS to zero, the writers are already
unblocked.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: round-robin registers as ulong, not u32
Jason A. Donenfeld [Tue, 22 Feb 2022 12:46:10 +0000 (13:46 +0100)]
random: round-robin registers as ulong, not u32

commit da3951ebdcd1cb1d5c750e08cd05aee7b0c04d9a upstream.

When the interrupt handler does not have a valid cycle counter, it calls
get_reg() to read a register from the irq stack, in round-robin.
Currently it does this assuming that registers are 32-bit. This is
_probably_ the case, and probably all platforms without cycle counters
are in fact 32-bit platforms. But maybe not, and either way, it's not
quite correct. This commit fixes that to deal with `unsigned long`
rather than `u32`.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: clear fast pool, crng, and batches in cpuhp bring up
Jason A. Donenfeld [Sun, 13 Feb 2022 21:48:04 +0000 (22:48 +0100)]
random: clear fast pool, crng, and batches in cpuhp bring up

commit 3191dd5a1179ef0fad5a050a1702ae98b6251e8f upstream.

For the irq randomness fast pool, rather than having to use expensive
atomics, which were visibly the most expensive thing in the entire irq
handler, simply take care of the extreme edge case of resetting count to
zero in the cpuhp online handler, just after workqueues have been
reenabled. This simplifies the code a bit and lets us use vanilla
variables rather than atomics, and performance should be improved.

As well, very early on when the CPU comes up, while interrupts are still
disabled, we clear out the per-cpu crng and its batches, so that it
always starts with fresh randomness.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Sultan Alsawaf <sultan@kerneltoast.com>
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: pull add_hwgenerator_randomness() declaration into random.h
Jason A. Donenfeld [Sun, 13 Feb 2022 15:17:01 +0000 (16:17 +0100)]
random: pull add_hwgenerator_randomness() declaration into random.h

commit b777c38239fec5a528e59f55b379e31b1a187524 upstream.

add_hwgenerator_randomness() is a function implemented and documented
inside of random.c. It is the way that hardware RNGs push data into it.
Therefore, it should be declared in random.h. Otherwise sparse complains
with:

random.c:1137:6: warning: symbol 'add_hwgenerator_randomness' was not declared. Should it be static?

The alternative would be to include hw_random.h into random.c, but that
wouldn't really be good for anything except slowing down compile time.

Cc: Matt Mackall <mpm@selenic.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: check for crng_init == 0 in add_device_randomness()
Jason A. Donenfeld [Sat, 12 Feb 2022 22:57:38 +0000 (23:57 +0100)]
random: check for crng_init == 0 in add_device_randomness()

commit 1daf2f387652bf3a7044aea042f5023b3f6b189b upstream.

This has no real functional change, as crng_pre_init_inject() (and
before that, crng_slow_init()) always checks for == 0, not >= 2. So
correct the outer unlocked change to reflect that. Before this used
crng_ready(), which was not correct.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: unify early init crng load accounting
Jason A. Donenfeld [Sat, 12 Feb 2022 22:54:09 +0000 (23:54 +0100)]
random: unify early init crng load accounting

commit da792c6d5f59a76c10a310c5d4c93428fd18f996 upstream.

crng_fast_load() and crng_slow_load() have different semantics:

- crng_fast_load() xors and accounts with crng_init_cnt.
- crng_slow_load() hashes and doesn't account.

However add_hwgenerator_randomness() can afford to hash (it's called
from a kthread), and it should account. Additionally, ones that can
afford to hash don't need to take a trylock but can take a normal lock.
So, we combine these into one function, crng_pre_init_inject(), which
allows us to control these in a uniform way. This will make it simpler
later to simplify this all down when the time comes for that.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: do not take pool spinlock at boot
Jason A. Donenfeld [Sat, 12 Feb 2022 00:26:17 +0000 (01:26 +0100)]
random: do not take pool spinlock at boot

commit afba0b80b977b2a8f16234f2acd982f82710ba33 upstream.

Since rand_initialize() is run while interrupts are still off and
nothing else is running, we don't need to repeatedly take and release
the pool spinlock, especially in the RDSEED loop.

Reviewed-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: defer fast pool mixing to worker
Jason A. Donenfeld [Fri, 4 Feb 2022 15:15:46 +0000 (16:15 +0100)]
random: defer fast pool mixing to worker

commit 58340f8e952b613e0ead0bed58b97b05bf4743c5 upstream.

On PREEMPT_RT, it's problematic to take spinlocks from hard irq
handlers. We can fix this by deferring to a workqueue the dumping of
the fast pool into the input pool.

We accomplish this with some careful rules on fast_pool->count:

  - When it's incremented to >= 64, we schedule the work.
  - If the top bit is set, we never schedule the work, even if >= 64.
  - The worker is responsible for setting it back to 0 when it's done.

There are two small issues around using workqueues for this purpose that
we work around.

The first issue is that mix_interrupt_randomness() might be migrated to
another CPU during CPU hotplug. This issue is rectified by checking that
it hasn't been migrated (after disabling irqs). If it has been migrated,
then we set the count to zero, so that when the CPU comes online again,
it can requeue the work. As part of this, we switch to using an
atomic_t, so that the increment in the irq handler doesn't wipe out the
zeroing if the CPU comes back online while this worker is running.

The second issue is that, though relatively minor in effect, we probably
want to make sure we get a consistent view of the pool onto the stack,
in case it's interrupted by an irq while reading. To do this, we don't
reenable irqs until after the copy. There are only 18 instructions
between the cli and sti, so this is a pretty tiny window.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Jonathan Neuschäfer <j.neuschaefer@gmx.net>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Sultan Alsawaf <sultan@kerneltoast.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: rewrite header introductory comment
Jason A. Donenfeld [Fri, 11 Feb 2022 11:29:33 +0000 (12:29 +0100)]
random: rewrite header introductory comment

commit 5f75d9f3babea8ae0a2d06724656874f41d317f5 upstream.

Now that we've re-documented the various sections, we can remove the
outdated text here and replace it with a high-level overview.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: group sysctl functions
Jason A. Donenfeld [Fri, 11 Feb 2022 11:53:34 +0000 (12:53 +0100)]
random: group sysctl functions

commit 0deff3c43206c24e746b1410f11125707ad3040e upstream.

This pulls all of the sysctl-focused functions into the sixth labeled
section.

No functional changes.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: group userspace read/write functions
Jason A. Donenfeld [Fri, 11 Feb 2022 11:53:34 +0000 (12:53 +0100)]
random: group userspace read/write functions

commit a6adf8e7a605250b911e94793fd077933709ff9e upstream.

This pulls all of the userspace read/write-focused functions into the
fifth labeled section.

No functional changes.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: group entropy collection functions
Jason A. Donenfeld [Fri, 11 Feb 2022 11:53:34 +0000 (12:53 +0100)]
random: group entropy collection functions

commit 92c653cf14400946f376a29b828d6af7e01f38dd upstream.

This pulls all of the entropy collection-focused functions into the
fourth labeled section.

No functional changes.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: group entropy extraction functions
Jason A. Donenfeld [Fri, 11 Feb 2022 11:53:34 +0000 (12:53 +0100)]
random: group entropy extraction functions

commit a5ed7cb1a7732ef11959332d507889fbc39ebbb4 upstream.

This pulls all of the entropy extraction-focused functions into the
third labeled section.

No functional changes.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: group crng functions
Jason A. Donenfeld [Fri, 11 Feb 2022 11:53:34 +0000 (12:53 +0100)]
random: group crng functions

commit 3655adc7089da4f8ca74cec8fcef73ea5101430e upstream.

This pulls all of the crng-focused functions into the second labeled
section.

No functional changes.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: group initialization wait functions
Jason A. Donenfeld [Fri, 11 Feb 2022 11:53:34 +0000 (12:53 +0100)]
random: group initialization wait functions

commit 5f1bb112006b104b3e2a1e1b39bbb9b2617581e6 upstream.

This pulls all of the readiness waiting-focused functions into the first
labeled section.

No functional changes.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: remove whitespace and reorder includes
Jason A. Donenfeld [Fri, 11 Feb 2022 12:41:41 +0000 (13:41 +0100)]
random: remove whitespace and reorder includes

commit 87e7d5abad0cbc9312dea7f889a57d294c1a5fcc upstream.

This is purely cosmetic. Future work involves figuring out which of
these headers we need and which we don't.

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: remove useless header comment
Jason A. Donenfeld [Fri, 11 Feb 2022 11:28:33 +0000 (12:28 +0100)]
random: remove useless header comment

commit 6071a6c0fba2d747742cadcbb3ba26ed756ed73b upstream.

This really adds nothing at all useful.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: introduce drain_entropy() helper to declutter crng_reseed()
Jason A. Donenfeld [Fri, 11 Feb 2022 11:19:49 +0000 (12:19 +0100)]
random: introduce drain_entropy() helper to declutter crng_reseed()

commit 246c03dd899164d0186b6d685d6387f228c28d93 upstream.

In preparation for separating responsibilities, break out the entropy
count management part of crng_reseed() into its own function.

No functional changes.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: deobfuscate irq u32/u64 contributions
Jason A. Donenfeld [Thu, 10 Feb 2022 16:01:27 +0000 (17:01 +0100)]
random: deobfuscate irq u32/u64 contributions

commit b2f408fe403800c91a49f6589d95b6759ce1b30b upstream.

In the irq handler, we fill out 16 bytes differently on 32-bit and
64-bit platforms, and for 32-bit vs 64-bit cycle counters, which doesn't
always correspond with the bitness of the platform. Whether or not you
like this strangeness, it is a matter of fact.  But it might not be a
fact you well realized until now, because the code that loaded the irq
info into 4 32-bit words was quite confusing.  Instead, this commit
makes everything explicit by having separate (compile-time) branches for
32-bit and 64-bit types.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: add proper SPDX header
Jason A. Donenfeld [Thu, 10 Feb 2022 15:43:57 +0000 (16:43 +0100)]
random: add proper SPDX header

commit a07fdae346c35c6ba286af1c88e0effcfa330bf9 upstream.

Convert the current license into the SPDX notation of "(GPL-2.0 OR
BSD-3-Clause)". This infers GPL-2.0 from the text "ALTERNATIVELY, this
product may be distributed under the terms of the GNU General Public
License, in which case the provisions of the GPL are required INSTEAD OF
the above restrictions" and it infers BSD-3-Clause from the verbatim
BSD 3 clause license in the file.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: remove unused tracepoints
Jason A. Donenfeld [Thu, 10 Feb 2022 15:40:44 +0000 (16:40 +0100)]
random: remove unused tracepoints

commit 14c174633f349cb41ea90c2c0aaddac157012f74 upstream.

These explicit tracepoints aren't really used and show sign of aging.
It's work to keep these up to date, and before I attempted to keep them
up to date, they weren't up to date, which indicates that they're not
really used. These days there are better ways of introspecting anyway.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: remove ifdef'd out interrupt bench
Jason A. Donenfeld [Thu, 10 Feb 2022 15:35:24 +0000 (16:35 +0100)]
random: remove ifdef'd out interrupt bench

commit 95e6060c20a7f5db60163274c5222a725ac118f9 upstream.

With tools like kbench9000 giving more finegrained responses, and this
basically never having been used ever since it was initially added,
let's just get rid of this. There *is* still work to be done on the
interrupt handler, but this really isn't the way it's being developed.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: tie batched entropy generation to base_crng generation
Jason A. Donenfeld [Wed, 9 Feb 2022 21:46:48 +0000 (22:46 +0100)]
random: tie batched entropy generation to base_crng generation

commit 0791e8b655cc373718f0f58800fdc625a3447ac5 upstream.

Now that we have an explicit base_crng generation counter, we don't need
a separate one for batched entropy. Rather, we can just move the
generation forward every time we change crng_init state or update the
base_crng key.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: fix locking for crng_init in crng_reseed()
Dominik Brodowski [Wed, 9 Feb 2022 18:57:06 +0000 (19:57 +0100)]
random: fix locking for crng_init in crng_reseed()

commit 7191c628fe07b70d3f37de736d173d1b115396ed upstream.

crng_init is protected by primary_crng->lock. Therefore, we need
to hold this lock when increasing crng_init to 2. As we shouldn't
hold this lock for too long, only hold it for those parts which
require protection.

Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: zero buffer after reading entropy from userspace
Jason A. Donenfeld [Wed, 9 Feb 2022 17:42:13 +0000 (18:42 +0100)]
random: zero buffer after reading entropy from userspace

commit 7b5164fb1279bf0251371848e40bae646b59b3a8 upstream.

This buffer may contain entropic data that shouldn't stick around longer
than needed, so zero out the temporary buffer at the end of write_pool().

Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Jann Horn <jannh@google.com>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: remove outdated INT_MAX >> 6 check in urandom_read()
Jason A. Donenfeld [Mon, 7 Feb 2022 22:37:13 +0000 (23:37 +0100)]
random: remove outdated INT_MAX >> 6 check in urandom_read()

commit 434537ae54ad37e93555de21b6ac8133d6d773a9 upstream.

In 79a8468747c5 ("random: check for increase of entropy_count because of
signed conversion"), a number of checks were added around what values
were passed to account(), because account() was doing fancy fixed point
fractional arithmetic, and a user had some ability to pass large values
directly into it. One of things in that commit was limiting those values
to INT_MAX >> 6. The first >> 3 was for bytes to bits, and the next >> 3
was for bits to 1/8 fractional bits.

However, for several years now, urandom reads no longer touch entropy
accounting, and so this check serves no purpose. The current flow is:

urandom_read_nowarn()-->get_random_bytes_user()-->chacha20_block()

Of course, we don't want that size_t to be truncated when adding it into
the ssize_t. But we arrive at urandom_read_nowarn() in the first place
either via ordinary fops, which limits reads to MAX_RW_COUNT, or via
getrandom() which limits reads to INT_MAX.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Jann Horn <jannh@google.com>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: make more consistent use of integer types
Jason A. Donenfeld [Wed, 9 Feb 2022 13:43:25 +0000 (14:43 +0100)]
random: make more consistent use of integer types

commit 04ec96b768c9dd43946b047c3da60dcc66431370 upstream.

We've been using a flurry of int, unsigned int, size_t, and ssize_t.
Let's unify all of this into size_t where it makes sense, as it does in
most places, and leave ssize_t for return values with possible errors.

In addition, keeping with the convention of other functions in this
file, functions that are dealing with raw bytes now take void *
consistently instead of a mix of that and u8 *, because much of the time
we're actually passing some other structure that is then interpreted as
bytes by the function.

We also take the opportunity to fix the outdated and incorrect comment
in get_random_bytes_arch().

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Jann Horn <jannh@google.com>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: use hash function for crng_slow_load()
Jason A. Donenfeld [Tue, 8 Feb 2022 18:23:17 +0000 (19:23 +0100)]
random: use hash function for crng_slow_load()

commit 66e4c2b9541503d721e936cc3898c9f25f4591ff upstream.

Since we have a hash function that's really fast, and the goal of
crng_slow_load() is reportedly to "touch all of the crng's state", we
can just hash the old state together with the new state and call it a
day. This way we dont need to reason about another LFSR or worry about
various attacks there. This code is only ever used at early boot and
then never again.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: use simpler fast key erasure flow on per-cpu keys
Jason A. Donenfeld [Mon, 7 Feb 2022 14:08:49 +0000 (15:08 +0100)]
random: use simpler fast key erasure flow on per-cpu keys

commit 186873c549df11b63e17062f863654e1501e1524 upstream.

Rather than the clunky NUMA full ChaCha state system we had prior, this
commit is closer to the original "fast key erasure RNG" proposal from
<https://blog.cr.yp.to/20170723-random.html>, by simply treating ChaCha
keys on a per-cpu basis.

All entropy is extracted to a base crng key of 32 bytes. This base crng
has a birthdate and a generation counter. When we go to take bytes from
the crng, we first check if the birthdate is too old; if it is, we
reseed per usual. Then we start working on a per-cpu crng.

This per-cpu crng makes sure that it has the same generation counter as
the base crng. If it doesn't, it does fast key erasure with the base
crng key and uses the output as its new per-cpu key, and then updates
its local generation counter. Then, using this per-cpu state, we do
ordinary fast key erasure. Half of this first block is used to overwrite
the per-cpu crng key for the next call -- this is the fast key erasure
RNG idea -- and the other half, along with the ChaCha state, is returned
to the caller. If the caller desires more than this remaining half, it
can generate more ChaCha blocks, unlocked, using the now detached ChaCha
state that was just returned. Crypto-wise, this is more or less what we
were doing before, but this simply makes it more explicit and ensures
that we always have backtrack protection by not playing games with a
shared block counter.

The flow looks like this:

──extract()──► base_crng.key ◄──memcpy()───┐
                   │                       │
                   └──chacha()──────┬─► new_base_key
                                    └─► crngs[n].key ◄──memcpy()───┐
                                              │                    │
                                              └──chacha()───┬─► new_key
                                                            └─► random_bytes
                                                                      │
                                                                      └────►

There are a few hairy details around early init. Just as was done
before, prior to having gathered enough entropy, crng_fast_load() and
crng_slow_load() dump bytes directly into the base crng, and when we go
to take bytes from the crng, in that case, we're doing fast key erasure
with the base crng rather than the fast unlocked per-cpu crngs. This is
fine as that's only the state of affairs during very early boot; once
the crng initializes we never use these paths again.

In the process of all this, the APIs into the crng become a bit simpler:
we have get_random_bytes(buf, len) and get_random_bytes_user(buf, len),
which both do what you'd expect. All of the details of fast key erasure
and per-cpu selection happen only in a very short critical section of
crng_make_state(), which selects the right per-cpu key, does the fast
key erasure, and returns a local state to the caller's stack. So, we no
longer have a need for a separate backtrack function, as this happens
all at once here. The API then allows us to extend backtrack protection
to batched entropy without really having to do much at all.

The result is a bit simpler than before and has fewer foot guns. The
init time state machine also gets a lot simpler as we don't need to wait
for workqueues to come online and do deferred work. And the multi-core
performance should be increased significantly, by virtue of having hardly
any locking on the fast path.

Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Jann Horn <jannh@google.com>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agorandom: absorb fast pool into input pool after fast load
Jason A. Donenfeld [Wed, 9 Feb 2022 00:56:35 +0000 (01:56 +0100)]
random: absorb fast pool into input pool after fast load

commit c30c575db4858f0bbe5e315ff2e529c782f33a1f upstream.

During crng_init == 0, we never credit entropy in add_interrupt_
randomness(), but instead dump it directly into the primary_crng. That's
fine, except for the fact that we then wind up throwing away that
entropy later when we switch to extracting from the input pool and
xoring into (and later in this series overwriting) the primary_crng key.
The two other early init sites -- add_hwgenerator_randomness()'s use
crng_fast_load() and add_device_ randomness()'s use of crng_slow_load()
-- always additionally give their inputs to the input pool. But not
add_interrupt_randomness().

This commit fixes that shortcoming by calling mix_pool_bytes() after
crng_fast_load() in add_interrupt_randomness(). That's partially
verboten on PREEMPT_RT, where it implies taking spinlock_t from an IRQ
handler. But this also only happens during early boot and then never
again after that. Plus it's a trylock so it has the same considerations
as calling crng_fast_load(), which we're already using.

Cc: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Dominik Brodowski <linux@dominikbrodowski.net>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Suggested-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>