propagating vmgenid outward and upward
Subsystems (7)
Jason A. Donenfeld
maintainerHey folks,
Having finally wrapped up development of the initial vmgenid driver, I thought I'd pull together some thoughts on vmgenid, notification, and propagating, from disjointed conversations I've had with a few of you over the last several weeks.
The basic problem is: VMs can be cloned, forked, rewound, or snapshotted, and when this happens, a) the RNG needs to reseed itself, and b) cryptographic algorithms that are not reuse resistant need to reinitialize in one way or another. For 5.18, we're handling (a) via the new vmgenid driver, which implements a spec from Microsoft, whereby the driver receives ACPI notifications when a 16 byte unique value changes.
The vmgenid driver basically works, though it is racy, because that ACPI notification can arrive after the system is already running again. This race is even worse on Windows, where they kick the notification into a worker thread, which then publishes it upward elsewhere to another async mechanism, and eventually it hits the RNG and various userspace apps. On Linux it's not that bad -- we reseed immediately upon receiving the notification -- but it still inherits this same "push"-model deficiency, which a "pull"-model would not have.
If we had a "pull" model, rather than just expose a 16-byte unique identifier, the vmgenid virtual hardware would _also_ expose a word-sized generation counter, which would be incremented every time the unique ID changed. Then, every time we would touch the RNG, we'd simply do an inexpensive check of this memremap()'d integer, and reinitialize with the unique ID if the integer changed. In this way, the race would be entirely eliminated. We would then be able to propagate this outwards to other drivers, by just exporting an extern symbol, in the manner of jiffies, and propagate it upwards to userspace, by putting it in the vDSO, in the manner of gettimeofday. And like that, there'd be no terrible async thing and things would work pretty easily.
But that's not what we have, because Microsoft didn't collaborate with anybody on this, and now it's implemented in several hypervisors. Given that I'm already spending considerable time working on the RNG, entirely without funding, somehow I'm not super motivated to lead a cross-industry political effort to change Microsoft's vmgenid spec. Maybe somebody else has an appetite for this, but either way, those changes would be several years off at best.
So given we have a "push"-model mechanism, there are two problems to tackle, perhaps in the same way, perhaps in a different way:
A) Outwards propagation toward other kernel drivers: in this case, I have in mind WireGuard, naturally, which very much needs to clear its existing sessions when VMs are forked.
B) Upwards propagation to userspace: in this case, we handle the concerns of the Amazon engineers on this thread who broached this topic a few years ago, in which s2n, their TLS library, wants to reinitialize its userspace RNG (a silly thing, but I digress) and probably clear session keys too, for the same good reason as WireGuard.
For (A), at least wearing my WireGuard-maintainer hat, there is an easy way and there is a "race-free" way. I use scare quotes there because we're still in a "push"-model, which means it's still racy no matter what.
The faux "race-free" way involves having extern u32 rng_vm_generation; or similar in random.h, and then everything that generates a session key would snapshot this value, and every time a session key is used, a comparison would be made. This works, but given that we're going to be racy no matter what, I think I'd prefer avoiding the extra code in the hot path and extra per-session storage. It seems like that'd involve a lot of fiddly engineering for no real world benefit.
The easy way, and the way that I think I prefer, would be to just have a sync notifier_block for this, just like we have with register_pm_notifier(). From my perspective, it'd be simplest to just piggy back on the already existing PM notifier with an extra event, PM_POST_VMFORK, which would join the existing set of 7, following PM_POST_RESTORE. I think that'd be coherent. However, if the PM people don't want to play ball, we could always come up with our own notifier_block. But I don't see the need. Plus, WireGuard *already* uses the PM notifier for clearing keys, so code-wise for my use case, that'd amount adding another case for PM_POST_VMFORK, in addition to the currently existing PM_HIBERNATION_PREPARE and PM_SUSPEND_PREPARE cases, which all would be treated the same way. Ezpz. So if that sounds like an interesting thing to the PM people, I think I'd like to propose a patch for that, possibly even for 5.18, given that it'd be very straight- forward.
For (B), it's a little bit trickier. But I think our options follow the same rubric. We can expose a generation counter in the vDSO, with semantics akin to the extern integer I described above. Or we could expose that counter in a file that userspace could poll() on and receive notifications that way. Or perhaps a third way. I'm all ears here. Alex's team from Amazon last year proposed something similar to the vDSO idea, except using mmap on a sysfs file, though from what I can tell, that wound up being kind of complicated. Due to the fact that we're _already_ racy, I think I'm most inclined at this point toward the poll() approach for the same reasons as I prefer a notifier_block. But on userspace I could be convinced otherwise, and I'd be interested in totally different ideas here too.
Another thing I should note is that, while I'm not currently leaning toward it, the vDSO approach also ties into interesting discussions about userspace RNGs (generally a silly idea), and their need for things like fork detection and also learning when the kernel RNG was last reseeded. So cracking open the vDSO book might invite all sorts of other interesting questions and discussions, which may be productive or may be a humongous distraction. (Also, again, I'm not super enthusiastic about userspace RNGs.)
Also, there is an interesting question to decide with regards to userspace, which is whether the vmgenid driver should expose its unique ID to userspace, as Alex requested on an earlier thread. I am actually sort of opposed to this. That unique ID may or may not be secret and entropic; if it isn't, the crypto is designed to not be impacted negatively, but if it is, we should keep it secret. So, rather, I think the correct flow is that userspace simply calls getrandom() upon learning that the VM forked, which is guaranteed to have been reinitialized already by add_vmfork_randomness(), and that will guarantee a value that is unique to the VM, without having to actually expose that value.
So, anyway, this is more or less where my thinking on this matter is. Would be happy to hear some fresh ideas here too.
Regards,
Jason
Quoted 23 lines · Jason A. Donenfeld
Hey folks,
Having finally wrapped up development of the initial vmgenid driver, I thought I'd pull together some thoughts on vmgenid, notification, and propagating, from disjointed conversations I've had with a few of you over the last several weeks.
The basic problem is: VMs can be cloned, forked, rewound, or snapshotted, and when this happens, a) the RNG needs to reseed itself, and b) cryptographic algorithms that are not reuse resistant need to reinitialize in one way or another. For 5.18, we're handling (a) via the new vmgenid driver, which implements a spec from Microsoft, whereby the driver receives ACPI notifications when a 16 byte unique value changes.
The vmgenid driver basically works, though it is racy, because that ACPI notification can arrive after the system is already running again. This race is even worse on Windows, where they kick the notification into a worker thread, which then publishes it upward elsewhere to another async mechanism, and eventually it hits the RNG and various userspace apps. On Linux it's not that bad -- we reseed immediately upon receiving the notification -- but it still inherits this same "push"-model deficiency, which a "pull"-model would not have.
Quoted 6 lines
If we had a "pull" model, rather than just expose a 16-byte unique identifier, the vmgenid virtual hardware would _also_ expose a word-sized generation counter, which would be incremented every time the unique ID changed. Then, every time we would touch the RNG, we'd simply do an inexpensive check of this memremap()'d integer, and reinitialize with the unique ID if the integer changed.
Does the vmgenid spec (as-is) preclude the use of the 16-byte identifier like this?
After all, once you locate the identifier via the ADDR object, you could perhaps consult it every time you were about to touch the RNG. Perhaps with the help of 16-byte atomic (?) operations, you could maintain a copy of the identifier elsewhere, and detect (atomically, upon each RNG access) whether the identifier has changed, without waiting for the Notify.
We'd require (or assume) the hypervisors to modify the identifier in guest RAM before resuming execution of the guest. And we'd have to delay RNG accesses in the guest kernel until after the vmgenid device -- or its absence -- were discovered.
Just a random guesss.
Thanks
Laszlo
Quoted 98 lines
In this way, the race would
be entirely eliminated. We would then be able to propagate this outwards to other drivers, by just exporting an extern symbol, in the manner ofjiffies, and propagate it upwards to userspace, by putting it in the vDSO, in the manner of gettimeofday. And like that, there'd be no terrible async thing and things would work pretty easily.But that's not what we have, because Microsoft didn't collaborate with anybody on this, and now it's implemented in several hypervisors. Given that I'm already spending considerable time working on the RNG, entirely without funding, somehow I'm not super motivated to lead a cross-industry political effort to change Microsoft's vmgenid spec. Maybe somebody else has an appetite for this, but either way, those changes would be several years off at best.
So given we have a "push"-model mechanism, there are two problems to tackle, perhaps in the same way, perhaps in a different way:
A) Outwards propagation toward other kernel drivers: in this case, I have in mind WireGuard, naturally, which very much needs to clear its existing sessions when VMs are forked.
B) Upwards propagation to userspace: in this case, we handle the concerns of the Amazon engineers on this thread who broached this topic a few years ago, in which s2n, their TLS library, wants to reinitialize its userspace RNG (a silly thing, but I digress) and probably clear session keys too, for the same good reason as WireGuard.
For (A), at least wearing my WireGuard-maintainer hat, there is an easy way and there is a "race-free" way. I use scare quotes there because we're still in a "push"-model, which means it's still racy no matter what.
The faux "race-free" way involves having
extern u32 rng_vm_generation;or similar inrandom.h, and then everything that generates a session key would snapshot this value, and every time a session key is used, a comparison would be made. This works, but given that we're going to be racy no matter what, I think I'd prefer avoiding the extra code in the hot path and extra per-session storage. It seems like that'd involve a lot of fiddly engineering for no real world benefit.The easy way, and the way that I think I prefer, would be to just have a sync
notifier_blockfor this, just like we have withregister_pm_notifier(). From my perspective, it'd be simplest to just piggy back on the already existing PM notifier with an extra event,PM_POST_VMFORK, which would join the existing set of 7, followingPM_POST_RESTORE. I think that'd be coherent. However, if the PM people don't want to play ball, we could always come up with our ownnotifier_block. But I don't see the need. Plus, WireGuard *already* uses the PM notifier for clearing keys, so code-wise for my use case, that'd amount adding another case forPM_POST_VMFORK, in addition to the currently existingPM_HIBERNATION_PREPAREandPM_SUSPEND_PREPAREcases, which all would be treated the same way. Ezpz. So if that sounds like an interesting thing to the PM people, I think I'd like to propose a patch for that, possibly even for 5.18, given that it'd be very straight- forward.For (B), it's a little bit trickier. But I think our options follow the same rubric. We can expose a generation counter in the vDSO, with semantics akin to the extern integer I described above. Or we could expose that counter in a file that userspace could
poll()on and receive notifications that way. Or perhaps a third way. I'm all ears here. Alex's team from Amazon last year proposed something similar to the vDSO idea, except using mmap on a sysfs file, though from what I can tell, that wound up being kind of complicated. Due to the fact that we're _already_ racy, I think I'm most inclined at this point toward thepoll()approach for the same reasons as I prefer anotifier_block. But on userspace I could be convinced otherwise, and I'd be interested in totally different ideas here too.Another thing I should note is that, while I'm not currently leaning toward it, the vDSO approach also ties into interesting discussions about userspace RNGs (generally a silly idea), and their need for things like fork detection and also learning when the kernel RNG was last reseeded. So cracking open the vDSO book might invite all sorts of other interesting questions and discussions, which may be productive or may be a humongous distraction. (Also, again, I'm not super enthusiastic about userspace RNGs.)
Also, there is an interesting question to decide with regards to userspace, which is whether the vmgenid driver should expose its unique ID to userspace, as Alex requested on an earlier thread. I am actually sort of opposed to this. That unique ID may or may not be secret and entropic; if it isn't, the crypto is designed to not be impacted negatively, but if it is, we should keep it secret. So, rather, I think the correct flow is that userspace simply calls
getrandom()upon learning that the VM forked, which is guaranteed to have been reinitialized already byadd_vmfork_randomness(), and that will guarantee a value that is unique to the VM, without having to actually expose that value.So, anyway, this is more or less where my thinking on this matter is. Would be happy to hear some fresh ideas here too.
Regards,
Jason
Jason A. Donenfeld
maintainerHi Laszlo,
Quoted 7 lines · Laszlo Ersek
Quoted 6 lines
If we had a "pull" model, rather than just expose a 16-byte unique identifier, the vmgenid virtual hardware would _also_ expose a word-sized generation counter, which would be incremented every time the unique ID changed. Then, every time we would touch the RNG, we'd simply do an inexpensive check of this memremap()'d integer, and reinitialize with the unique ID if the integer changed.
Quoted 5 lines
Does the vmgenid spec (as-is) preclude the use of the 16-byte identifier like this?
After all, once you locate the identifier via the ADDR object, you could perhaps consult it every time you were about to touch the RNG.
No, you could in fact do this, and there'd be nothing wrong with that from a spec perspective. You could even vDSO it all the way through onward to userspace. However, doing a 16-byte atomic memcmp on each-and-every packet is really a non-starter. For that kind of "check it in the hot path" thing to be viable, you really want it to be a counter that is word-sized. The "pull"-model involves pulling on every single packet in order to be better than the "push"-model. Anyway, even with a word-sized counter, it's unclear whether the costs of checking on every packet would be worth it to everyone, but at least it's more tenable than a 16-byte whammy.
Jason
Quoted 27 lines · Jason A. Donenfeld
Hi Laszlo,
Quoted 7 lines · Laszlo Ersek
Quoted 6 lines
If we had a "pull" model, rather than just expose a 16-byte unique identifier, the vmgenid virtual hardware would _also_ expose a word-sized generation counter, which would be incremented every time the unique ID changed. Then, every time we would touch the RNG, we'd simply do an inexpensive check of this memremap()'d integer, and reinitialize with the unique ID if the integer changed.
Quoted 5 lines
Does the vmgenid spec (as-is) preclude the use of the 16-byte identifier like this?
After all, once you locate the identifier via the ADDR object, you could perhaps consult it every time you were about to touch the RNG.
No, you could in fact do this, and there'd be nothing wrong with that from a spec perspective. You could even vDSO it all the way through onward to userspace. However, doing a 16-byte atomic memcmp on each-and-every packet is really a non-starter. For that kind of "check it in the hot path" thing to be viable, you really want it to be a counter that is word-sized. The "pull"-model involves pulling on every single packet in order to be better than the "push"-model. Anyway, even with a word-sized counter, it's unclear whether the costs of checking on every packet would be worth it to everyone, but at least it's more tenable than a 16-byte whammy.
Quoted 1 line
Jason
Hmm okay, so it's a performance optimization... some batching then? Do you really need to worry about every packet? Every 64 packets not enough? Packets are after all queued at NICs etc, and VM fork can happen after they leave wireguard ...
Signature · 1 line
MST
Jason A. Donenfeld
maintainerHi Michael,
Quoted 4 lines · Michael S. Tsirkin
Hmm okay, so it's a performance optimization... some batching then? Do you really need to worry about every packet? Every 64 packets not enough? Packets are after all queued at NICs etc, and VM fork can happen after they leave wireguard ...
Unfortunately, yes, this is an "every packet" sort of thing -- if the race is to be avoided in a meaningful way. It's really extra bad:
ChaCha20 and AES-CTR work by xoring a secret stream of bytes with plaintext to produce a ciphertext. If you use that same secret stream and xor it with a second plaintext and transmit that too, an attacker can combine the two different ciphertexts to learn things about the original plaintext.
But, anyway, it seems like the race is here to stay given what we have _currently_ available with the virtual hardware. That's why I'm focused on trying to get something going that's the least bad with what we've currently got, which is racy by design. How vitally important is it to have something that doesn't race in the far future? I don't know, really. It seems plausible that that ACPI notifier triggers so early that nothing else really even has a chance, so the race concern is purely theoretical. But I haven't tried to measure that so I'm not sure.
Jason
Quoted 8 lines · Jason A. Donenfeld
Hi Michael,
Quoted 4 lines · Michael S. Tsirkin
Hmm okay, so it's a performance optimization... some batching then? Do you really need to worry about every packet? Every 64 packets not enough? Packets are after all queued at NICs etc, and VM fork can happen after they leave wireguard ...
Quoted 7 lines
Unfortunately, yes, this is an "every packet" sort of thing -- if the race is to be avoided in a meaningful way. It's really extra bad:
ChaCha20 and AES-CTR work by xoring a secret stream of bytes with plaintext to produce a ciphertext. If you use that same secret stream and xor it with a second plaintext and transmit that too, an attacker can combine the two different ciphertexts to learn things about the original plaintext.
So what about the point about packets queued then? You don't fish packets out of qdisc queues, do you?
Quoted 11 lines
But, anyway, it seems like the race is here to stay given what we have _currently_ available with the virtual hardware. That's why I'm focused on trying to get something going that's the least bad with what we've currently got, which is racy by design. How vitally important is it to have something that doesn't race in the far future? I don't know, really. It seems plausible that that ACPI notifier triggers so early that nothing else really even has a chance, so the race concern is purely theoretical. But I haven't tried to measure that so I'm not sure.
Jason
Quoted 17 lines · Michael S. Tsirkin
Quoted 8 lines · Jason A. Donenfeld
Hi Michael,
Quoted 4 lines · Michael S. Tsirkin
Hmm okay, so it's a performance optimization... some batching then? Do you really need to worry about every packet? Every 64 packets not enough? Packets are after all queued at NICs etc, and VM fork can happen after they leave wireguard ...
Quoted 7 lines
Unfortunately, yes, this is an "every packet" sort of thing -- if the race is to be avoided in a meaningful way. It's really extra bad:
ChaCha20 and AES-CTR work by xoring a secret stream of bytes with plaintext to produce a ciphertext. If you use that same secret stream and xor it with a second plaintext and transmit that too, an attacker can combine the two different ciphertexts to learn things about the original plaintext.
Quoted 2 lines
So what about the point about packets queued then? You don't fish packets out of qdisc queues, do you?
Oh pls ignore it, I think I got it. Resending same packet is not a problem, producing a new one is.
Quoted 11 lines
Quoted 11 lines
But, anyway, it seems like the race is here to stay given what we have _currently_ available with the virtual hardware. That's why I'm focused on trying to get something going that's the least bad with what we've currently got, which is racy by design. How vitally important is it to have something that doesn't race in the far future? I don't know, really. It seems plausible that that ACPI notifier triggers so early that nothing else really even has a chance, so the race concern is purely theoretical. But I haven't tried to measure that so I'm not sure.
Jason
So how about measuring the performance impact of reading the 16 byte vmgenid then? This could be a performance option, too - some people might want extra security, some might not care. And I feel if linux DTRT and reads the 16 bytes then hypervisor vendors will be motivated to improve and add a 4 byte unique one. As long as linux is interrupt driven there's no motivation for change.
Signature · 1 line
MST
Quoted 26 lines · Jason A. Donenfeld
Hi Michael,
Quoted 4 lines · Michael S. Tsirkin
Hmm okay, so it's a performance optimization... some batching then? Do you really need to worry about every packet? Every 64 packets not enough? Packets are after all queued at NICs etc, and VM fork can happen after they leave wireguard ...
Unfortunately, yes, this is an "every packet" sort of thing -- if the race is to be avoided in a meaningful way. It's really extra bad:
ChaCha20 and AES-CTR work by xoring a secret stream of bytes with plaintext to produce a ciphertext. If you use that same secret stream and xor it with a second plaintext and transmit that too, an attacker can combine the two different ciphertexts to learn things about the original plaintext.But, anyway, it seems like the race is here to stay given what we have _currently_ available with the virtual hardware. That's why I'm focused on trying to get something going that's the least bad with what we've currently got, which is racy by design. How vitally important is it to have something that doesn't race in the far future? I don't know, really. It seems plausible that that ACPI notifier triggers so early that nothing else really even has a chance, so the race concern is purely theoretical. But I haven't tried to measure that so I'm not sure.
Quoted 1 line
Jason
I got curious, and wrote a dumb benchmark:
#include <stdio.h>
#include <assert.h>
#include <limits.h>
#include <string.h>
struct lng {
unsigned long long l1;
unsigned long long l2;
};struct shrt {
unsigned long s;
};struct lng l = { 1, 2 };
struct shrt s = { 3 };
static void test1(volatile struct shrt *sp)
{
if (sp->s != s.s) {
printf("short mismatch!\n");
s.s = sp->s;
}
}
static void test2(volatile struct lng *lp)
{
if (lp->l1 != l.l1 || lp->l2 != l.l2) {
printf("long mismatch!\n");
l.l1 = lp->l1;
l.l2 = lp->l2;
}
}int main(int argc, char **argv)
{
volatile struct shrt sv = { 4 };
volatile struct lng lv = { 5, 6 };
if (argc > 1) {
printf("test 1\n");
for (int i = 0; i < 10000000; ++i)
test1(&sv);
} else {
printf("test 2\n");
for (int i = 0; i < 10000000; ++i)
test2(&lv);
}
return 0;
}Results (built with -O2, nothing fancy):
[mst@tuck ~]$ perf stat -r 1000 ./a.out 1 > /dev/null
Performance counter stats for './a.out 1' (1000 runs):
5.12 msec task-clock:u # 0.945 CPUs utilized ( +- 0.07% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
52 page-faults:u # 10.016 K/sec ( +- 0.07% )
20,190,800 cycles:u # 3.889 GHz ( +- 0.01% )
50,147,371 instructions:u # 2.48 insn per cycle ( +- 0.00% )
20,032,224 branches:u # 3.858 G/sec ( +- 0.00% )
1,604 branch-misses:u # 0.01% of all branches ( +- 0.26% )0.00541882 +- 0.00000847 seconds time elapsed ( +- 0.16% )
[mst@tuck ~]$ perf stat -r 1000 ./a.out > /dev/null
Performance counter stats for './a.out' (1000 runs):
7.75 msec task-clock:u # 0.947 CPUs utilized ( +- 0.12% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
52 page-faults:u # 6.539 K/sec ( +- 0.07% )
30,205,916 cycles:u # 3.798 GHz ( +- 0.01% )
80,147,373 instructions:u # 2.65 insn per cycle ( +- 0.00% )
30,032,227 branches:u # 3.776 G/sec ( +- 0.00% )
1,621 branch-misses:u # 0.01% of all branches ( +- 0.23% )0.00817982 +- 0.00000965 seconds time elapsed ( +- 0.12% )
So yes, the overhead is higher by 50% which seems a lot but it's from a very small number, so I don't see why it's a show stopper, it's not by a factor of 10 such that we should sacrifice safety by default. Maybe a kernel flag that removes the read replacing it with an interrupt will do.
In other words, premature optimization is the root of all evil.
Signature · 1 line
MST
Jason A. Donenfeld
maintainerHey Michael,
Thanks for the benchmark.
Quoted 7 lines · Michael S. Tsirkin
So yes, the overhead is higher by 50% which seems a lot but it's from a very small number, so I don't see why it's a show stopper, it's not by a factor of 10 such that we should sacrifice safety by default. Maybe a kernel flag that removes the read replacing it with an interrupt will do.
In other words, premature optimization is the root of all evil.
Unfortunately I don't think it's as simple as that for several reasons.
First, I'm pretty confident a beefy Intel machine can mostly hide non-dependent comparisons in the memory access and have the problem mostly go away. But this is much less the case on, say, an in-order MIPS32r2, which isn't just "some crappy ISA I'm using for the sake of argument," but actually the platform on which a lot of networking and WireGuard stuff runs, so I do care about it. There, we have 4 reads/comparisons which can't pipeline nearly as well.
There's also the atomicity aspect, which I think makes your benchmark not quite accurate. Those 16 bytes could change between the first and second word (or between the Nth and N+1th word for N<=3 on 32-bit). What if in that case the word you read second doesn't change, but the word you read first did? So then you find yourself having to do a hi-lo-hi dance. And then consider the 32-bit case, where that's even more annoying. This is just one of those things that comes up when you compare the semantics of a "large unique ID" and "word-sized counter", as general topics. (My suggestion is that vmgenid provide both.)
Finally, there's a slightly storage aspect, where adding 16 bytes to a per-key struct is a little bit heavier than adding 4 bytes and might bust a cache line without sufficient care, care which always has some cost in one way or another.
So I just don't know if it's realistic to impose a 16-byte per-packet comparison all the time like that. I'm familiar with WireGuard obviously, but there's also cifs and maybe even wifi and bluetooth, and who knows what else, to care about too. Then there's the userspace discussion. I can't imagine a 16-byte hotpath comparison being accepted as implementable.
Quoted 4 lines
And I feel if linux
DTRT and reads the 16 bytes then hypervisor vendors will be motivated to improve and add a 4 byte unique one. As long as linux is interrupt driven there's no motivation for change.
I reeeeeally don't want to get pulled into the politics of this on the hypervisor side. I assume an improved thing would begin with QEMU and Firecracker or something collaborating because they're both open source and Amazon people seem interested. And then pressure builds for Microsoft and VMware to do it on their side. And then we get this all nicely implemented in the kernel. In the meantime, though, I'm not going to refuse to address the problem entirely just because the virtual hardware is less than perfect; I'd rather make the most with what we've got while still being somewhat reasonable from an implementation perspective.
Jason
Quoted 15 lines · Jason A. Donenfeld
Hey Michael,
Thanks for the benchmark.
Quoted 7 lines · Michael S. Tsirkin
So yes, the overhead is higher by 50% which seems a lot but it's from a very small number, so I don't see why it's a show stopper, it's not by a factor of 10 such that we should sacrifice safety by default. Maybe a kernel flag that removes the read replacing it with an interrupt will do.
In other words, premature optimization is the root of all evil.
Unfortunately I don't think it's as simple as that for several reasons.
Quoted 7 lines
First, I'm pretty confident a beefy Intel machine can mostly hide non-dependent comparisons in the memory access and have the problem mostly go away. But this is much less the case on, say, an in-order MIPS32r2, which isn't just "some crappy ISA I'm using for the sake of argument," but actually the platform on which a lot of networking and WireGuard stuff runs, so I do care about it. There, we have 4 reads/comparisons which can't pipeline nearly as well.
Sure. Want to try running some benchmarks on that platform? Presumably you have access to such a box, right?
Quoted 10 lines
There's also the atomicity aspect, which I think makes your benchmark not quite accurate. Those 16 bytes could change between the first and second word (or between the Nth and N+1th word for N<=3 on 32-bit). What if in that case the word you read second doesn't change, but the word you read first did? So then you find yourself having to do a hi-lo-hi dance.
And then consider the 32-bit case, where that's even
more annoying. This is just one of those things that comes up when you compare the semantics of a "large unique ID" and "word-sized counter", as general topics. (My suggestion is that vmgenid provide both.)
I don't see how this matters for any applications at all. Feel free to present a case that would be race free with a word but not a 16 byte value, I could not imagine one. It's human to err of course.
Quoted 6 lines
Finally, there's a slightly storage aspect, where adding 16 bytes to a per-key struct is a little bit heavier than adding 4 bytes and might bust a cache line without sufficient care, care which always has some cost in one way or another.
Quoted 6 lines
So I just don't know if it's realistic to impose a 16-byte per-packet comparison all the time like that. I'm familiar with WireGuard obviously, but there's also cifs and maybe even wifi and bluetooth, and who knows what else, to care about too. Then there's the userspace discussion. I can't imagine a 16-byte hotpath comparison being accepted as implementable.
I think this hinges on benchmarking results. Want to start with my silly benchmark at least? If you can't measure an order of magnitude gain then I think any effect on wireguard will be in the noise.
Quoted 9 lines
Quoted 4 lines
And I feel if linux
DTRT and reads the 16 bytes then hypervisor vendors will be motivated to improve and add a 4 byte unique one. As long as linux is interrupt driven there's no motivation for change.I reeeeeally don't want to get pulled into the politics of this on the hypervisor side. I assume an improved thing would begin with QEMU and Firecracker or something collaborating because they're both open source and Amazon people seem interested.
I think it would begin with a benchmark showing there's even any measureable performance to be gained by switching the semantics.
Quoted 9 lines
And then pressure builds for
Microsoft and VMware to do it on their side. And then we get this all nicely implemented in the kernel. In the meantime, though, I'm not going to refuse to address the problem entirely just because the virtual hardware is less than perfect; I'd rather make the most with what we've got while still being somewhat reasonable from an implementation perspective.Jason
Right but given you are trading security off for performance, it matters a lot what the performance gain is.
Signature · 1 line
MST
Jason A. Donenfeld
maintainerHi Michael,
Quoted 11 lines · Michael S. Tsirkin
Quoted 10 lines
There's also the atomicity aspect, which I think makes your benchmark not quite accurate. Those 16 bytes could change between the first and second word (or between the Nth and N+1th word for N<=3 on 32-bit). What if in that case the word you read second doesn't change, but the word you read first did? So then you find yourself having to do a hi-lo-hi dance.
And then consider the 32-bit case, where that's even
more annoying. This is just one of those things that comes up when you compare the semantics of a "large unique ID" and "word-sized counter", as general topics. (My suggestion is that vmgenid provide both.)
Quoted 3 lines
I don't see how this matters for any applications at all. Feel free to present a case that would be race free with a word but not a 16 byte value, I could not imagine one. It's human to err of course.
Word-size reads happen all at once on systems that Linux supports, whereas this is not the case for 16 bytes (with a few niche exceptions like cmpxchg16b and such). If you read the counter atomically, you can check to see whether it's changed just after encrypting but before transmitting and not transmit if it has changed, and voila, no race. With 16 bytes, synchronization of that read is pretty tricky (though maybe not all together impossible), because, as I mentioned, the first word might have changed by the time you read a matching second word. I'm sure you're familiar with the use of seqlocks in the kernel for solving a somewhat related problem.
Jason
Quoted 29 lines · Jason A. Donenfeld
Hi Michael,
Quoted 11 lines · Michael S. Tsirkin
Quoted 10 lines
There's also the atomicity aspect, which I think makes your benchmark not quite accurate. Those 16 bytes could change between the first and second word (or between the Nth and N+1th word for N<=3 on 32-bit). What if in that case the word you read second doesn't change, but the word you read first did? So then you find yourself having to do a hi-lo-hi dance.
And then consider the 32-bit case, where that's even
more annoying. This is just one of those things that comes up when you compare the semantics of a "large unique ID" and "word-sized counter", as general topics. (My suggestion is that vmgenid provide both.)Quoted 3 lines
I don't see how this matters for any applications at all. Feel free to present a case that would be race free with a word but not a 16 byte value, I could not imagine one. It's human to err of course.
Word-size reads happen all at once on systems that Linux supports, whereas this is not the case for 16 bytes (with a few niche exceptions like cmpxchg16b and such). If you read the counter atomically, you can check to see whether it's changed just after encrypting but before transmitting and not transmit if it has changed, and voila, no race. With 16 bytes, synchronization of that read is pretty tricky (though maybe not all together impossible), because, as I mentioned, the first word might have changed by the time you read a matching second word. I'm sure you're familiar with the use of seqlocks in the kernel for solving a somewhat related problem.
Quoted 1 line
Jason
I just don't see how "value changed while it was read" is so different from "value changed one clock after it was read". Since we don't detect the latter I don't see why we should worry about the former. What I don't have here is how would a code reading the value look. It might help to write some pseudo code to show that, but I'd say it makes more sense to just code the read up even just so the overhead of the current interface can be roughtly measured.
Signature · 1 line
MST
Jason A. Donenfeld
maintainerHi Michael,
Quoted 3 lines · Michael S. Tsirkin
I just don't see how "value changed while it was read" is so different from "value changed one clock after it was read". Since we don't detect the latter I don't see why we should worry about the former.
The "barrier" is at the point where the plaintext has been chosen AND the nonce for a given keypair has been selected. So, if you have plaintext in a buffer, and a key in a buffer, and the nonce for that encryption in a buffer, and then after those are all selected, you check to see if the vmgenid has changed since the birth of that key, then you're all set. If it changes _after_ that point of check (your "one clock after"), it doesn't matter: you'll just be double-transmitting the same ciphertext, which is something that flaky wifi sometimes does _anyway_ (and attackers can do intentionally), so network protocols already are resilient to replay. This is the same case you asked about earlier, and then answered yourself, when you were wondering about reaching down into qdiscs.
Jason
Quoted 20 lines · Jason A. Donenfeld
Hi Michael,
Quoted 3 lines · Michael S. Tsirkin
I just don't see how "value changed while it was read" is so different from "value changed one clock after it was read". Since we don't detect the latter I don't see why we should worry about the former.
The "barrier" is at the point where the plaintext has been chosen AND the nonce for a given keypair has been selected. So, if you have plaintext in a buffer, and a key in a buffer, and the nonce for that encryption in a buffer, and then after those are all selected, you check to see if the vmgenid has changed since the birth of that key, then you're all set. If it changes _after_ that point of check (your "one clock after"), it doesn't matter: you'll just be double-transmitting the same ciphertext, which is something that flaky wifi sometimes does _anyway_ (and attackers can do intentionally), so network protocols already are resilient to replay. This is the same case you asked about earlier, and then answered yourself, when you were wondering about reaching down into qdiscs.
Quoted 1 line
Jason
So writing some code:
1:
put plaintext in a buffer
put a key in a buffer
put the nonce for that encryption in a buffer
if vm gen id != stored vm gen id
stored vm gen id = vm gen id
goto 1
I think this is race free, but I don't see why does it matter whether we read gen id atomically or not.
Signature · 1 line
MST
Jason A. Donenfeld
maintainerHi Michael,
Quoted 7 lines · Michael S. Tsirkin
So writing some code:
1:
put plaintext in a buffer
put a key in a buffer
put the nonce for that encryption in a buffer
Quoted 6 lines
if vm gen id != stored vm gen id
stored vm gen id = vm gen id
goto 1I think this is race free, but I don't see why does it matter whether we read gen id atomically or not.
Because that 16 byte read of vmgenid is not atomic. Let's say you read the first 8 bytes, and then the VM is forked. In the forked VM, the next 8 bytes are the same as last time, but the first 8 bytes, which you already read, have changed. In that case, your != becomes a ==, and the test fails.
This is one of those fundamental things of "unique ID" vs "generation counter word".
Anyway, per your request in your last email, I wrote some code for this, which may or may not be totally broken, and only works on 64-bit x86, which is really the best possible case in terms of performance. And even so, it's not great.
Jason
--------8<------------------------
drivers/net/wireguard/noise.c +1 -0
@@ -106,6 +106,7 @@ static struct noise_keypair *keypair_create(struct wg_peer *peer) keypair->entry.type =INDEX_HASHTABLE_KEYPAIR ; keypair->entry.peer = peer;kref_init (&keypair->refcount); + keypair->vmgenid =vmgenid_read_atomic ();return keypair; }
drivers/net/wireguard/noise.h +3 -0
@@ -27,10 +27,13 @@ struct noise_symmetric_key {bool is_valid; }; +extern __uint128_t vmgenid_read_atomic (void ); +struct noise_keypair {struct index_hashtable_entry entry;struct noise_symmetric_key sending;atomic64_t sending_counter; +__uint128_t vmgenid;struct noise_symmetric_key receiving;struct noise_replay_counter receiving_counter;__le32 remote_index;
drivers/net/wireguard/send.c +3 -0
@@ -381,6 +381,9 @@ void wg_packet_send_staged_packets(struct wg_peer *peer) goto out_invalid; } +if (keypair->vmgenid !=vmgenid_read_atomic ()) + goto out_invalid; + packets.prev->next =NULL ;wg_peer_get (keypair->entry.peer);PACKET_CB (packets.next)->keypair = keypair;
drivers/virt/vmgenid.c +16 -0
@@ -21,6 +21,21 @@ struct vmgenid_state {u8 this_id[VMGENID_SIZE ]; }; +static __uint128_t *val; + +__uint128_t vmgenid_read_atomic (void ) +{ +__uint128_t ret =0 ; +if (!val) +return 0 ; + asmvolatile ("lock cmpxchg16b %1" + :"+A" (ret) + :"m" (*val),"b" (0 ),"c" (0 ) + :"cc" ); +return ret; +} +EXPORT_SYMBOL (vmgenid_read_atomic); +static int vmgenid_add (struct acpi_device *device) {struct acpi_buffer parsed = {ACPI_ALLOCATE_BUFFER };
@@ -50,6 +65,7 @@ static int vmgenid_add(struct acpi_device *device) phys_addr = (obj->package.elements[0 ].integer.value <<0 ) | (obj->package.elements[1 ].integer.value <<32 ); state->next_id =devm_memremap (&device->dev, phys_addr,VMGENID_SIZE ,MEMREMAP_WB ); + val = (__uint128_t *)state->next_id;if (IS_ERR (state->next_id)) { ret =PTR_ERR (state->next_id); goto out;
Quoted 14 lines · Jason A. Donenfeld
Hi Michael,
Quoted 11 lines · Michael S. Tsirkin
So writing some code:
1:
put plaintext in a buffer
put a key in a buffer
put the nonce for that encryption in a bufferif vm gen id != stored vm gen id
stored vm gen id = vm gen id
goto 1
Quoted 5 lines
Quoted 2 lines
I think this is race free, but I don't see why does it matter whether we read gen id atomically or not.
Because that 16 byte read of vmgenid is not atomic. Let's say you read the first 8 bytes, and then the VM is forked.
But at this point when VM was forked plaintext key and nonce are all in buffer, and you previously indicated a fork at this point is harmless. You wrote "If it changes _after_ that point of check ... it doesn't matter:"
Quoted 4 lines
In the forked VM, the next
8 bytes are the same as last time, but the first 8 bytes, which you already read, have changed. In that case, your != becomes a ==, and the test fails.
Yes I'm aware what an atomic read is. If the read is not atomic a part of value can change ;)
Signature · 1 line
MST
Jason A. Donenfeld
maintainerHi Michael,
Quoted 7 lines · Michael S. Tsirkin
Quoted 2 lines
Because that 16 byte read of vmgenid is not atomic. Let's say you read the first 8 bytes, and then the VM is forked.
But at this point when VM was forked plaintext key and nonce are all in buffer, and you previously indicated a fork at this point is harmless. You wrote "If it changes _after_ that point of check ... it doesn't matter:"
Ahhh, fair point. I think you're right.
Alright, so all we're talking about here is an ordinary 16-byte read, and 16 bytes of storage per keypair, and a 16-byte comparison.
Still seems much worse than just having a single word...
Jason
Quoted 11 lines · Jason A. Donenfeld
Hi Michael,
Quoted 7 lines · Michael S. Tsirkin
Quoted 2 lines
Because that 16 byte read of vmgenid is not atomic. Let's say you read the first 8 bytes, and then the VM is forked.
But at this point when VM was forked plaintext key and nonce are all in buffer, and you previously indicated a fork at this point is harmless. You wrote "If it changes _after_ that point of check ... it doesn't matter:"
Quoted 8 lines
Ahhh, fair point. I think you're right.
Alright, so all we're talking about here is an ordinary 16-byte read, and 16 bytes of storage per keypair, and a 16-byte comparison.
Still seems much worse than just having a single word...
Jason
And it is, I saw a 30% higher overhead, it is however 30% of a very low number ;)
Signature · 1 line
MST
Quoted 11 lines · Jason A. Donenfeld
Hi Michael,
Quoted 7 lines · Michael S. Tsirkin
Quoted 2 lines
Because that 16 byte read of vmgenid is not atomic. Let's say you read the first 8 bytes, and then the VM is forked.
But at this point when VM was forked plaintext key and nonce are all in buffer, and you previously indicated a fork at this point is harmless. You wrote "If it changes _after_ that point of check ... it doesn't matter:"
Quoted 8 lines
Ahhh, fair point. I think you're right.
Alright, so all we're talking about here is an ordinary 16-byte read, and 16 bytes of storage per keypair, and a 16-byte comparison.
Still seems much worse than just having a single word...
Jason
Oh I forgot about __int128.
#include <stdio.h>
#include <assert.h>
#include <limits.h>
#include <string.h>
struct lng {
__int128 l;
};struct shrt {
unsigned long s;
};struct lng l = { 1 };
struct shrt s = { 3 };
static void test1(volatile struct shrt *sp)
{
if (sp->s != s.s) {
printf("short mismatch!\n");
s.s = sp->s;
}
}
static void test2(volatile struct lng *lp)
{
if (lp->l != l.l) {
printf("long mismatch!\n");
l.l = lp->l;
}
}int main(int argc, char **argv)
{
volatile struct shrt sv = { 4 };
volatile struct lng lv = { 5 };
if (argc > 1) {
printf("test 1\n");
for (int i = 0; i < 100000000; ++i)
test1(&sv);
} else {
printf("test 2\n");
for (int i = 0; i < 100000000; ++i)
test2(&lv);
}
return 0;
}with that the compiler has an easier time to produce optimal code, so the difference is smaller.
Note: compiled with
gcc -O2 -mno-sse -mno-sse2 -ggdb bench3.c
since with sse there's no difference at all.
[mst@tuck ~]$ perf stat -r 100 ./a.out 1 > /dev/null
Performance counter stats for './a.out 1' (100 runs):
94.55 msec task-clock:u # 0.996 CPUs utilized ( +- 0.09% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
52 page-faults:u # 548.914 /sec ( +- 0.21% )
400,459,851 cycles:u # 4.227 GHz ( +- 0.03% )
500,147,935 instructions:u # 1.25 insn per cycle ( +- 0.00% )
200,032,462 branches:u # 2.112 G/sec ( +- 0.00% )
1,810 branch-misses:u # 0.00% of all branches ( +- 0.73% )0.0949732 +- 0.0000875 seconds time elapsed ( +- 0.09% )
[mst@tuck ~]$
[mst@tuck ~]$ perf stat -r 100 ./a.out > /dev/null
Performance counter stats for './a.out' (100 runs):
110.19 msec task-clock:u # 1.136 CPUs utilized ( +- 0.18% )
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
52 page-faults:u # 537.743 /sec ( +- 0.22% )
428,518,442 cycles:u # 4.431 GHz ( +- 0.07% )
900,147,986 instructions:u # 2.24 insn per cycle ( +- 0.00% )
200,032,505 branches:u # 2.069 G/sec ( +- 0.00% )
2,139 branch-misses:u # 0.00% of all branches ( +- 0.77% )0.096956 +- 0.000203 seconds time elapsed ( +- 0.21% )
Signature · 1 line
MST
Quoted 70 lines · Jason A. Donenfeld
Hi Michael,
Quoted 7 lines · Michael S. Tsirkin
So writing some code:
1:
put plaintext in a buffer
put a key in a buffer
put the nonce for that encryption in a bufferQuoted 6 lines
if vm gen id != stored vm gen id
stored vm gen id = vm gen id
goto 1I think this is race free, but I don't see why does it matter whether we read gen id atomically or not.
Because that 16 byte read of vmgenid is not atomic. Let's say you read the first 8 bytes, and then the VM is forked. In the forked VM, the next 8 bytes are the same as last time, but the first 8 bytes, which you already read, have changed. In that case, your != becomes a ==, and the test fails.
This is one of those fundamental things of "unique ID" vs "generation counter word".
Anyway, per your request in your last email, I wrote some code for this, which may or may not be totally broken, and only works on 64-bit x86, which is really the best possible case in terms of performance. And even so, it's not great.
Jason
--------8<------------------------
drivers/net/wireguard/noise.c +1
-0@@ -106,6 +106,7 @@ static struct noise_keypair *keypair_create(struct wg_peer *peer) keypair->entry.type =INDEX_HASHTABLE_KEYPAIR ; keypair->entry.peer = peer;kref_init (&keypair->refcount); + keypair->vmgenid =vmgenid_read_atomic ();return keypair; }drivers/net/wireguard/noise.h +3
-0@@ -27,10 +27,13 @@ struct noise_symmetric_key {bool is_valid; }; +extern __uint128_t vmgenid_read_atomic (void ); +struct noise_keypair {struct index_hashtable_entry entry;struct noise_symmetric_key sending;atomic64_t sending_counter; +__uint128_t vmgenid;struct noise_symmetric_key receiving;struct noise_replay_counter receiving_counter;__le32 remote_index;drivers/net/wireguard/send.c +0
-0
drivers/net/wireguard/send.c void wg_packet_send_staged_packets(struct wg_peer *peer) · line 389
Quoted 9 lines
drivers/net/wireguard/send.c +3
-0@@ -381,6 +381,9 @@ void wg_packet_send_staged_packets(struct wg_peer *peer) goto out_invalid; } +if (keypair->vmgenid !=vmgenid_read_atomic ()) + goto out_invalid; + packets.prev->next =NULL ;wg_peer_get (keypair->entry.peer);PACKET_CB (packets.next)->keypair = keypair;
I don't think we care about an atomic read here. All data is in buffer by this point, if it did not fork before that then we are ok, even if it forks during the read.
We probably do need a memory barrier to make sure all writes complete before the read of vmgenid, I'm not sure which kind - I think hypervisor can be trusted to do a full CPU barrier on fork so probably just a compiler barrier.
Quoted 34 lines
drivers/virt/vmgenid.c +16
-0@@ -21,6 +21,21 @@ struct vmgenid_state {u8 this_id[VMGENID_SIZE ]; }; +static __uint128_t *val; + +__uint128_t vmgenid_read_atomic (void ) +{ +__uint128_t ret =0 ; +if (!val) +return 0 ; + asmvolatile ("lock cmpxchg16b %1" + :"+A" (ret) + :"m" (*val),"b" (0 ),"c" (0 ) + :"cc" ); +return ret; +} +EXPORT_SYMBOL (vmgenid_read_atomic); +static int vmgenid_add (struct acpi_device *device) {struct acpi_buffer parsed = {ACPI_ALLOCATE_BUFFER };@@ -50,6 +65,7 @@ static int vmgenid_add(struct acpi_device *device) phys_addr = (obj->package.elements[0 ].integer.value <<0 ) | (obj->package.elements[1 ].integer.value <<32 ); state->next_id =devm_memremap (&device->dev, phys_addr,VMGENID_SIZE ,MEMREMAP_WB ); + val = (__uint128_t *)state->next_id;if (IS_ERR (state->next_id)) { ret =PTR_ERR (state->next_id); goto out;
Quoted 23 lines · Jason A. Donenfeld
Hey folks,
Having finally wrapped up development of the initial vmgenid driver, I thought I'd pull together some thoughts on vmgenid, notification, and propagating, from disjointed conversations I've had with a few of you over the last several weeks.
The basic problem is: VMs can be cloned, forked, rewound, or snapshotted, and when this happens, a) the RNG needs to reseed itself, and b) cryptographic algorithms that are not reuse resistant need to reinitialize in one way or another. For 5.18, we're handling (a) via the new vmgenid driver, which implements a spec from Microsoft, whereby the driver receives ACPI notifications when a 16 byte unique value changes.
The vmgenid driver basically works, though it is racy, because that ACPI notification can arrive after the system is already running again. This race is even worse on Windows, where they kick the notification into a worker thread, which then publishes it upward elsewhere to another async mechanism, and eventually it hits the RNG and various userspace apps. On Linux it's not that bad -- we reseed immediately upon receiving the notification -- but it still inherits this same "push"-model deficiency, which a "pull"-model would not have.
Quoted 11 lines
If we had a "pull" model, rather than just expose a 16-byte unique identifier, the vmgenid virtual hardware would _also_ expose a word-sized generation counter, which would be incremented every time the unique ID changed. Then, every time we would touch the RNG, we'd simply do an inexpensive check of this memremap()'d integer, and reinitialize with the unique ID if the integer changed. In this way, the race would be entirely eliminated. We would then be able to propagate this outwards to other drivers, by just exporting an extern symbol, in the manner of
jiffies, and propagate it upwards to userspace, by putting it in the vDSO, in the manner of gettimeofday. And like that, there'd be no terrible async thing and things would work pretty easily.
I am not sure what the difference is though. So we have a 16 byte unique value and you would prefer a dword counter. How is the former not a superset of the later? I'm not sure how safe it is to expose it to userspace specifically, but rest of text talks about exposing it to a kernel driver so maybe not an issue? So what makes interrupt driven required, and why not just remap and read existing vmgenid in the pull manner? What did I miss?
Quoted 90 lines
But that's not what we have, because Microsoft didn't collaborate with anybody on this, and now it's implemented in several hypervisors. Given that I'm already spending considerable time working on the RNG, entirely without funding, somehow I'm not super motivated to lead a cross-industry political effort to change Microsoft's vmgenid spec. Maybe somebody else has an appetite for this, but either way, those changes would be several years off at best.
So given we have a "push"-model mechanism, there are two problems to tackle, perhaps in the same way, perhaps in a different way:
A) Outwards propagation toward other kernel drivers: in this case, I have in mind WireGuard, naturally, which very much needs to clear its existing sessions when VMs are forked.
B) Upwards propagation to userspace: in this case, we handle the concerns of the Amazon engineers on this thread who broached this topic a few years ago, in which s2n, their TLS library, wants to reinitialize its userspace RNG (a silly thing, but I digress) and probably clear session keys too, for the same good reason as WireGuard.
For (A), at least wearing my WireGuard-maintainer hat, there is an easy way and there is a "race-free" way. I use scare quotes there because we're still in a "push"-model, which means it's still racy no matter what.
The faux "race-free" way involves having
extern u32 rng_vm_generation;or similar inrandom.h, and then everything that generates a session key would snapshot this value, and every time a session key is used, a comparison would be made. This works, but given that we're going to be racy no matter what, I think I'd prefer avoiding the extra code in the hot path and extra per-session storage. It seems like that'd involve a lot of fiddly engineering for no real world benefit.The easy way, and the way that I think I prefer, would be to just have a sync
notifier_blockfor this, just like we have withregister_pm_notifier(). From my perspective, it'd be simplest to just piggy back on the already existing PM notifier with an extra event,PM_POST_VMFORK, which would join the existing set of 7, followingPM_POST_RESTORE. I think that'd be coherent. However, if the PM people don't want to play ball, we could always come up with our ownnotifier_block. But I don't see the need. Plus, WireGuard *already* uses the PM notifier for clearing keys, so code-wise for my use case, that'd amount adding another case forPM_POST_VMFORK, in addition to the currently existingPM_HIBERNATION_PREPAREandPM_SUSPEND_PREPAREcases, which all would be treated the same way. Ezpz. So if that sounds like an interesting thing to the PM people, I think I'd like to propose a patch for that, possibly even for 5.18, given that it'd be very straight- forward.For (B), it's a little bit trickier. But I think our options follow the same rubric. We can expose a generation counter in the vDSO, with semantics akin to the extern integer I described above. Or we could expose that counter in a file that userspace could
poll()on and receive notifications that way. Or perhaps a third way. I'm all ears here. Alex's team from Amazon last year proposed something similar to the vDSO idea, except using mmap on a sysfs file, though from what I can tell, that wound up being kind of complicated. Due to the fact that we're _already_ racy, I think I'm most inclined at this point toward thepoll()approach for the same reasons as I prefer anotifier_block. But on userspace I could be convinced otherwise, and I'd be interested in totally different ideas here too.Another thing I should note is that, while I'm not currently leaning toward it, the vDSO approach also ties into interesting discussions about userspace RNGs (generally a silly idea), and their need for things like fork detection and also learning when the kernel RNG was last reseeded. So cracking open the vDSO book might invite all sorts of other interesting questions and discussions, which may be productive or may be a humongous distraction. (Also, again, I'm not super enthusiastic about userspace RNGs.)
Also, there is an interesting question to decide with regards to userspace, which is whether the vmgenid driver should expose its unique ID to userspace, as Alex requested on an earlier thread. I am actually sort of opposed to this. That unique ID may or may not be secret and entropic; if it isn't, the crypto is designed to not be impacted negatively, but if it is, we should keep it secret. So, rather, I think the correct flow is that userspace simply calls
getrandom()upon learning that the VM forked, which is guaranteed to have been reinitialized already byadd_vmfork_randomness(), and that will guarantee a value that is unique to the VM, without having to actually expose that value.So, anyway, this is more or less where my thinking on this matter is. Would be happy to hear some fresh ideas here too.
Regards,
Jason
Jason A. Donenfeld
maintainerHi Michael,
Quoted 12 lines · Michael S. Tsirkin
Quoted 11 lines
If we had a "pull" model, rather than just expose a 16-byte unique identifier, the vmgenid virtual hardware would _also_ expose a word-sized generation counter, which would be incremented every time the unique ID changed. Then, every time we would touch the RNG, we'd simply do an inexpensive check of this memremap()'d integer, and reinitialize with the unique ID if the integer changed. In this way, the race would be entirely eliminated. We would then be able to propagate this outwards to other drivers, by just exporting an extern symbol, in the manner of
jiffies, and propagate it upwards to userspace, by putting it in the vDSO, in the manner of gettimeofday. And like that, there'd be no terrible async thing and things would work pretty easily.
Quoted 3 lines
I am not sure what the difference is though. So we have a 16 byte unique value and you would prefer a dword counter. How is the former not a superset of the later?
Laszlo just asked the same question, which I answered here:
<https://lore.kernel.org/lkml/Yh5JwK6toc%[email protected]/>. You have to read the full 16 bytes. You can't safely just read the first 4 or 8 or something, because it's a "unique ID" rather than a counter. That seems like a needlessly expensive thing to do on each-and-every packet.
Quoted 5 lines
I'm not sure how safe it is to expose it to
userspace specifically, but rest of text talks about exposing it to a kernel driver so maybe not an issue? So what makes interrupt driven required, and why not just remap and read existing vmgenid in the pull manner? What did I miss?
I don't really understand your question, but guessing your meaning: I'm not talking about exposing the actual 16-byte value to any other drivers, but just notifying them that their sessions should be dropped. If it's easier to think about this in code, grep for wg_pm_notification(), and consider that it'd be changing this code:
if (action != PM_HIBERNATION_PREPARE && action != PM_SUSPEND_PREPARE)
return 0;into:
if (action != PM_HIBERNATION_PREPARE && action != PM_SUSPEND_PREPARE &&
action != PM_VMFORK_POST)
return 0;But perhaps I misunderstood this part of your question?
Jason
Quoted 7 lines · Jason A. Donenfeld
The easy way, and the way that I think I prefer, would be to just have a sync
notifier_blockfor this, just like we have withregister_pm_notifier(). From my perspective, it'd be simplest to just piggy back on the already existing PM notifier with an extra event,PM_POST_VMFORK, which would join the existing set of 7, followingPM_POST_RESTORE. I think that'd be coherent. However, if the PM people don't want to play ball, we could always come up with our own
Quoted 8 lines
notifier_block. But I don't see the need. Plus, WireGuard *already* uses the PM notifier for clearing keys, so code-wise for my use case, that'd amount adding another case forPM_POST_VMFORK, in addition to the currently existingPM_HIBERNATION_PREPAREandPM_SUSPEND_PREPAREcases, which all would be treated the same way. Ezpz. So if that sounds like an interesting thing to the PM people, I think I'd like to propose a patch for that, possibly even for 5.18, given that it'd be very straight- forward.
A notifier block like this makes sense, but why tie onto the PM_ stuff? This isn't power management issues, it's a system-wide change that I am sure others will want to know about that doesn't reflect any power changes.
As much as I hate adding new notifiers in the kernel, that might be all you need here.
thanks,
greg k-h
Jason A. Donenfeld
maintainerHi Greg,
Quoted 7 lines · Greg KH
A notifier block like this makes sense, but why tie onto the
PM_stuff? This isn't power management issues, it's a system-wide change that I am sure others will want to know about that doesn't reflect any power changes.As much as I hate adding new notifiers in the kernel, that might be all you need here.
You might indeed be right. I guess I was thinking that "resuming from suspend" and "resuming from a VM fork" are kind of the same thing. There _is_ a certain kind of similarity between the two. I was hoping if the similarity was a strong enough one, maybe it'd make sense to do them together rather than adding another notifier. But I suppose you disagree, and it sounds like Rafael might too --
<https://lore.kernel.org/lkml/CAJZ5v0g+GihH_b9YvwuHzdrUVNGXOeabOznDC1vK6qLi8gtSTQ@mail.gmail.com/>.
Code-wise for me with WireGuard it's of course appealing to treat them the same, since it's like a one line change, but if I need to add a new notifier call there, it's not the end of the world.
Jason
Quoted 11 lines · Jason A. Donenfeld
Hi Greg,
Quoted 7 lines · Greg KH
A notifier block like this makes sense, but why tie onto the
PM_stuff? This isn't power management issues, it's a system-wide change that I am sure others will want to know about that doesn't reflect any power changes.As much as I hate adding new notifiers in the kernel, that might be all you need here.
Quoted 7 lines
You might indeed be right. I guess I was thinking that "resuming from suspend" and "resuming from a VM fork" are kind of the same thing. There _is_ a certain kind of similarity between the two. I was hoping if the similarity was a strong enough one, maybe it'd make sense to do them together rather than adding another notifier. But I suppose you disagree, and it sounds like Rafael might too --
<https://lore.kernel.org/lkml/CAJZ5v0g+GihH_b9YvwuHzdrUVNGXOeabOznDC1vK6qLi8gtSTQ@mail.gmail.com/>.
Hey, nice, we agree! :)
Quoted 3 lines
Code-wise for me with WireGuard it's of course appealing to treat them the same, since it's like a one line change, but if I need to add a new notifier call there, it's not the end of the world.
I know there are other places in the kernel that would like to be notified when they have been moved to another machine so that they can do things like determine if the CPU functionality has changed (or not), and perhaps do other types of device reconfiguration. Right now the kernel does not have any way of knowing this, so it makes sense that if the platform (i.e. ACPI here) has a way of creating such a event, it should and then we can start tieing in other subsystems to use it as-needed.
thanks,
greg k-h
Jason A. Donenfeld
maintainerQuoted 21 lines · Greg KH
Quoted 11 lines · Jason A. Donenfeld
Hi Greg,
Quoted 7 lines · Greg KH
A notifier block like this makes sense, but why tie onto the
PM_stuff? This isn't power management issues, it's a system-wide change that I am sure others will want to know about that doesn't reflect any power changes.As much as I hate adding new notifiers in the kernel, that might be all you need here.
Quoted 7 lines
You might indeed be right. I guess I was thinking that "resuming from suspend" and "resuming from a VM fork" are kind of the same thing. There _is_ a certain kind of similarity between the two. I was hoping if the similarity was a strong enough one, maybe it'd make sense to do them together rather than adding another notifier. But I suppose you disagree, and it sounds like Rafael might too --
<https://lore.kernel.org/lkml/CAJZ5v0g+GihH_b9YvwuHzdrUVNGXOeabOznDC1vK6qLi8gtSTQ@mail.gmail.com/>.
Quoted 1 line
Hey, nice, we agree! :)
It is now done and posted here:
https://lore.kernel.org/lkml/[email protected]/
Jason
Jason A. Donenfeld
maintainerHey again,
Quoted 4 lines · Jason A. Donenfeld
For (B), it's a little bit trickier. But I think our options follow the same rubric. We can expose a generation counter in the vDSO, with semantics akin to the extern integer I described above. Or we could expose that counter in a file that userspace could
poll()on and receive
Quoted 8 lines
notifications that way. Or perhaps a third way. I'm all ears here. Alex's team from Amazon last year proposed something similar to the vDSO idea, except using mmap on a sysfs file, though from what I can tell, that wound up being kind of complicated. Due to the fact that we're _already_ racy, I think I'm most inclined at this point toward the
poll()approach for the same reasons as I prefer anotifier_block. But on userspace I could be convinced otherwise, and I'd be interested in totally different ideas here too.
I implemented the poll() case here in 15 lines of code and found it remarkably simple to do:
https://lore.kernel.org/lkml/[email protected]/
This is just a PoC/RFC for the sake of having something tangible to look at for this thread. It is notable to me, though, that implementing this was so minimal.
Regards,
Jason
Quoted 15 lines · Jason A. Donenfeld
Hey folks,
Having finally wrapped up development of the initial vmgenid driver, I thought I'd pull together some thoughts on vmgenid, notification, and propagating, from disjointed conversations I've had with a few of you over the last several weeks.
The basic problem is: VMs can be cloned, forked, rewound, or snapshotted, and when this happens, a) the RNG needs to reseed itself, and b) cryptographic algorithms that are not reuse resistant need to reinitialize in one way or another. For 5.18, we're handling (a) via the new vmgenid driver, which implements a spec from Microsoft, whereby the driver receives ACPI notifications when a 16 byte unique value changes.
Quoted 2 lines
The vmgenid driver basically works, though it is racy, because that ACPI notification can arrive after the system is already running again. This
I believe enough people already pointed out that this assumption is incorrect. The thing that is racy about VMGenID is the interrupt based notification. The actual identifier is updated before the VM resumes from its clone operation, so if you match on that you will know whether you are in a new or old world. And that is enough to create transactions: Save the identifier before a "crypto transaction", validate before you finish, if they don't match, abort, reseed and replay.
Quoted 92 lines
race is even worse on Windows, where they kick the notification into a worker thread, which then publishes it upward elsewhere to another async mechanism, and eventually it hits the RNG and various userspace apps. On Linux it's not that bad -- we reseed immediately upon receiving the notification -- but it still inherits this same "push"-model deficiency, which a "pull"-model would not have.
If we had a "pull" model, rather than just expose a 16-byte unique identifier, the vmgenid virtual hardware would _also_ expose a word-sized generation counter, which would be incremented every time the unique ID changed. Then, every time we would touch the RNG, we'd simply do an inexpensive check of this memremap()'d integer, and reinitialize with the unique ID if the integer changed. In this way, the race would be entirely eliminated. We would then be able to propagate this outwards to other drivers, by just exporting an extern symbol, in the manner of
jiffies, and propagate it upwards to userspace, by putting it in the vDSO, in the manner of gettimeofday. And like that, there'd be no terrible async thing and things would work pretty easily.But that's not what we have, because Microsoft didn't collaborate with anybody on this, and now it's implemented in several hypervisors. Given that I'm already spending considerable time working on the RNG, entirely without funding, somehow I'm not super motivated to lead a cross-industry political effort to change Microsoft's vmgenid spec. Maybe somebody else has an appetite for this, but either way, those changes would be several years off at best.
So given we have a "push"-model mechanism, there are two problems to tackle, perhaps in the same way, perhaps in a different way:
A) Outwards propagation toward other kernel drivers: in this case, I have in mind WireGuard, naturally, which very much needs to clear its existing sessions when VMs are forked.
B) Upwards propagation to userspace: in this case, we handle the concerns of the Amazon engineers on this thread who broached this topic a few years ago, in which s2n, their TLS library, wants to reinitialize its userspace RNG (a silly thing, but I digress) and probably clear session keys too, for the same good reason as WireGuard.
For (A), at least wearing my WireGuard-maintainer hat, there is an easy way and there is a "race-free" way. I use scare quotes there because we're still in a "push"-model, which means it's still racy no matter what.
The faux "race-free" way involves having
extern u32 rng_vm_generation;or similar inrandom.h, and then everything that generates a session key would snapshot this value, and every time a session key is used, a comparison would be made. This works, but given that we're going to be racy no matter what, I think I'd prefer avoiding the extra code in the hot path and extra per-session storage. It seems like that'd involve a lot of fiddly engineering for no real world benefit.The easy way, and the way that I think I prefer, would be to just have a sync
notifier_blockfor this, just like we have withregister_pm_notifier(). From my perspective, it'd be simplest to just piggy back on the already existing PM notifier with an extra event,PM_POST_VMFORK, which would join the existing set of 7, followingPM_POST_RESTORE. I think that'd be coherent. However, if the PM people don't want to play ball, we could always come up with our ownnotifier_block. But I don't see the need. Plus, WireGuard *already* uses the PM notifier for clearing keys, so code-wise for my use case, that'd amount adding another case forPM_POST_VMFORK, in addition to the currently existingPM_HIBERNATION_PREPAREandPM_SUSPEND_PREPAREcases, which all would be treated the same way. Ezpz. So if that sounds like an interesting thing to the PM people, I think I'd like to propose a patch for that, possibly even for 5.18, given that it'd be very straight- forward.For (B), it's a little bit trickier. But I think our options follow the same rubric. We can expose a generation counter in the vDSO, with semantics akin to the extern integer I described above. Or we could expose that counter in a file that userspace could
poll()on and receive notifications that way. Or perhaps a third way. I'm all ears here. Alex's team from Amazon last year proposed something similar to the vDSO idea, except using mmap on a sysfs file, though from what I can tell, that wound up being kind of complicated. Due to the fact that we're _already_ racy, I think I'm most inclined at this point toward thepoll()approach for the same reasons as I prefer anotifier_block. But on userspace I could be convinced otherwise, and I'd be interested in totally different ideas here too.Another thing I should note is that, while I'm not currently leaning toward it, the vDSO approach also ties into interesting discussions about userspace RNGs (generally a silly idea), and their need for things like fork detection and also learning when the kernel RNG was last reseeded. So cracking open the vDSO book might invite all sorts of other interesting questions and discussions, which may be productive or may be a humongous distraction. (Also, again, I'm not super enthusiastic about userspace RNGs.)
Quoted 11 lines
Also, there is an interesting question to decide with regards to userspace, which is whether the vmgenid driver should expose its unique ID to userspace, as Alex requested on an earlier thread. I am actually sort of opposed to this. That unique ID may or may not be secret and entropic; if it isn't, the crypto is designed to not be impacted negatively, but if it is, we should keep it secret. So, rather, I think the correct flow is that userspace simply calls
getrandom()upon learning that the VM forked, which is guaranteed to have been reinitialized already byadd_vmfork_randomness(), and that will guarantee a value that is unique to the VM, without having to actually expose that value.
If you follow the logic at the beginning of the mail, you can create something race free if you consume the hardware VMGenID counter. You can not make it race free if you rely on the interrupt mechanism.
So following that train of thought, if you expose the hardware VMGenID to user space, you could allow user space to act race free based on VMGenID. That means consumers of user space RNGs could validate whether the ID is identical between the beginning of the crypto operation and the end.
That said, there are 2 pieces to the puzzle of user space notification: Polling and event based. The part above solves the polling use cases - user space libraries that just want to know whether they are now in a new world.
However, there are more complicated cases as well. What do you do with Samba for example? It needs to generate a new SID after the clone. That's a super heavy operation. Do you want to have smbd constantly poll on the VMGenID just to see whether it needs to kick off some administrative actions?
For the event based approach, we're in the same boat as "S3 resume" - we need a global notification mechanism that the state of the system changed and act accordingly. That's where the systemd proposal[1] comes in: Create inhibitors and scriptlets that get spawned when we want to suspend and then resume-cloned later. I'm personally even ok if we just limit that whole use case to cloning while you're in S3 only.
In that case, all we would need from the kernel is an easily readable GenID that changes before systemd runs again after suspend: Systemd wakes up after resume, checks if the GenID changed and if so, invokes the unquiescing target in after the resume one.
For this particular use case we're not in the fast path, so we could make GenID reading a syscall which checks against VMGenID. But that won't cut it for the polling use case.
I'm also not a super big fan of putting all that logic into systemd. It means applications need to create their own notification mechanisms to pass that cloning notification into actual processes. Don't we have any mechanism that applications and libraries could use to natively get an event when the GenID changes?
Alex
[1] https://github.com/systemd/systemd/issues/20222
Quoted 6 lines
So, anyway, this is more or less where my thinking on this matter is. Would be happy to hear some fresh ideas here too.
Regards,
Jason
Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879
Jason A. Donenfeld
maintainerHi Alex,
Quoted 7 lines · Alexander Graf
Quoted 2 lines
The vmgenid driver basically works, though it is racy, because that ACPI notification can arrive after the system is already running again. This
I believe enough people already pointed out that this assumption is incorrect. The thing that is racy about VMGenID is the interrupt based notification.
I'm having a hard time figuring out what's different between your statement and mine. I said that the race is due to the notification. You said that the race is due to the notification. What subtle thing am I missing here that would lead you to say that my assumption is incorrect? Or did you just misread?
Quoted 5 lines
The actual identifier is updated before the VM resumes
from its clone operation, so if you match on that you will know whether you are in a new or old world. And that is enough to create
transactions: Save the identifier before a "crypto transaction", validate before you finish, if they don't match, abort, reseed and replay.
Right. But more than just transactions, it's useful to preventing key reuse vulnerabilities, in which case, you store the current identifier just before an ephemeral key is generated, and then subsequently check to see that the identifier hasn't changed before transmitting anything related to that key.
Quoted 3 lines
If you follow the logic at the beginning of the mail, you can create something race free if you consume the hardware VMGenID counter. You can not make it race free if you rely on the interrupt mechanism.
Yes, as mentioned and discussed in depth before. However, your use of the word "counter" is problematic. Vmgenid is not a counter. It's a unique identifier. That means you can't compare it with a single word comparison but have to compare all of the 16 bytes. That seems potentially expensive. It's for that reason that I suggested augmenting the vmgenid spec with an additional word-sized _counter_ that could be mapped into the kernels and into userspaces.
Quoted 5 lines
So following that train of thought, if you expose the hardware VMGenID to user space, you could allow user space to act race free based on VMGenID. That means consumers of user space RNGs could validate whether the ID is identical between the beginning of the crypto operation and the end.
Right.
Quoted 5 lines
However, there are more complicated cases as well. What do you do with Samba for example? It needs to generate a new SID after the clone. That's a super heavy operation. Do you want to have smbd constantly poll on the VMGenID just to see whether it needs to kick off some administrative actions?
Were it a single word-sized integer, mapped into memory, that wouldn't be much of a problem at all. It could constantly read this before and after every operation. The problem is that it's 16 bytes and understandably applications don't want to deal with that clunkiness.
Quoted 2 lines
In that case, all we would need from the kernel is an easily readable GenID that changes
Actually, no, you need even less than that. All that's required is a sysfs/procfs file that can be poll()'d on. It doesn't need to have any content. When poll() returns readable, the VM has been forked. Then userspace rngs and other things like that can call getrandom() to receive a fresh value to mix into whatever their operation is. Since all we're talking about here is _event notification_, all we need is that event, which is what poll() provides.
Quoted 5 lines
I'm also not a super big fan of putting all that logic into systemd. It means applications need to create their own notification mechanisms to pass that cloning notification into actual processes. Don't we have any mechanism that applications and libraries could use to natively get an event when the GenID changes?
Yes. poll() can do this. For the purposes of discussion, I've posted an implementation of this idea here:
https://lore.kernel.org/lkml/[email protected]/
What I'm sort of leaning toward is doing something like that patch, and then later if vmgenid ever grows an additional word-sized counter, moving to explore the race-free approach. Given the amount of programming required to actually implement the race-free approach (transactions and careful study of each case), the poll() file approach might be a medium-grade compromise for the time being. Evidently that's what Microsoft decided too.
Jason
Quoted 6 lines · Jason A. Donenfeld
Hi Alex,
Quoted 3 lines · Alexander Graf
Quoted 2 lines
The vmgenid driver basically works, though it is racy, because that ACPI notification can arrive after the system is already running again. This
Quoted 8 lines
Quoted 3 lines
I believe enough people already pointed out that this assumption is incorrect. The thing that is racy about VMGenID is the interrupt based notification.
I'm having a hard time figuring out what's different between your statement and mine. I said that the race is due to the notification. You said that the race is due to the notification. What subtle thing am I missing here that would lead you to say that my assumption is incorrect? Or did you just misread?
The subtle difference is that you don't need to rely on the notification to learn about the world switch. If you instead read VMGenID explicitly without waiting for the notification, you're guaranteed to always know whether you were cloned. That means the actual VMGenID interface is not always racy. Just the notification part is.
So you *can* build a race-free VMGenID mechanism. You just can't build a race-free *event based* VMGenID mechanism if you want to allow cloning at arbitrary points in time.
Quoted 15 lines
Quoted 6 lines
The actual identifier is updated before the VM resumes
from its clone operation, so if you match on that you will know whether you are in a new or old world. And that is enough to create
transactions: Save the identifier before a "crypto transaction", validate before you finish, if they don't match, abort, reseed and replay.Right. But more than just transactions, it's useful to preventing key reuse vulnerabilities, in which case, you store the current identifier just before an ephemeral key is generated, and then subsequently check to see that the identifier hasn't changed before transmitting anything related to that key.
Quoted 3 lines
If you follow the logic at the beginning of the mail, you can create something race free if you consume the hardware VMGenID counter. You can not make it race free if you rely on the interrupt mechanism.
Quoted 7 lines
Yes, as mentioned and discussed in depth before. However, your use of the word "counter" is problematic. Vmgenid is not a counter. It's a unique identifier. That means you can't compare it with a single word comparison but have to compare all of the 16 bytes. That seems potentially expensive. It's for that reason that I suggested augmenting the vmgenid spec with an additional word-sized _counter_ that could be mapped into the kernels and into userspaces.
I think Michael's benchmarks show quite well that it's not all that expensive. We're talking 2x 64bit compares within the same cache line. You already realized that we don't need them to be atomic - just properly barriered.
So what if we created a vsyscall that takes a buffer of "up to 16 bytes". If we later realize that an additional page with a 4 byte counter is a viable performance optimization, we can work with MS to add that to the spec. But the user space interface would stay identical.
Quoted 13 lines
Quoted 6 lines
So following that train of thought, if you expose the hardware VMGenID to user space, you could allow user space to act race free based on VMGenID. That means consumers of user space RNGs could validate whether the ID is identical between the beginning of the crypto operation and the end.
Right.
Quoted 5 lines
However, there are more complicated cases as well. What do you do with Samba for example? It needs to generate a new SID after the clone. That's a super heavy operation. Do you want to have smbd constantly poll on the VMGenID just to see whether it needs to kick off some administrative actions?
Quoted 4 lines
Were it a single word-sized integer, mapped into memory, that wouldn't be much of a problem at all. It could constantly read this before and after every operation. The problem is that it's 16 bytes and understandably applications don't want to deal with that clunkiness.
I don't think applications should be in the business of mapping arbitrary locations of the vdso space to match on them. We need to build an interface that is fast and flexible enough so they can just say "here is a buffer, tell me if the ID changed". That's practically all you need - on init you run that once. Later on, you invoke it every time between finishing a crypto operation and putting it on the wire.
Quoted 20 lines
Quoted 3 lines
In that case, all we would need from the kernel is an easily readable GenID that changes
Actually, no, you need even less than that. All that's required is a sysfs/procfs file that can be poll()'d on. It doesn't need to have any content. When
poll()returns readable, the VM has been forked. Then userspace rngs and other things like that can callgetrandom()to receive a fresh value to mix into whatever their operation is. Since all we're talking about here is_eventnotification_, all we need is that event, which is whatpoll()provides.Quoted 5 lines
I'm also not a super big fan of putting all that logic into systemd. It means applications need to create their own notification mechanisms to pass that cloning notification into actual processes. Don't we have any mechanism that applications and libraries could use to natively get an event when the GenID changes?
Yes.
poll()can do this. For the purposes of discussion, I've posted an implementation of this idea here:
https://lore.kernel.org/lkml/[email protected]/
Quoted 7 lines
What I'm sort of leaning toward is doing something like that patch, and then later if vmgenid ever grows an additional word-sized counter, moving to explore the race-free approach. Given the amount of programming required to actually implement the race-free approach (transactions and careful study of each case), the
poll()file approach might be a medium-grade compromise for the time being. Evidently that's what Microsoft decided too.
I agree on the slightly racy compromise and that it's a step into the right direction. Doing this is a no brainer IMHO and I like the proc based poll approach.
I have an additional problem you might have an idea for with the poll based path. In addition to the clone notification, I'd need to know at which point everyone who was listening to a clone notification is finished acting up it. If I spawn a tiny VM to do "work", I want to know when it's safe to hand requests into it. How do I find out when that point in time is?
As far as the race-free approach goes, I wouldn't get hung up on 4 byte vs 16 byte UUID to match against. Outside of FUD that this might potentially have performance impact (4 byte reads will have impact too!), there's nothing that would keep us from implementing that interface in addition to the poll today.
I'm happy to see all of these things evolve incrementally though. We can start with the poll interface and then later implement a vsyscall that allows transactions in hot paths.
Alex
Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879
Quoted 1 line · Alexander Graf
I agree on the slightly racy compromise
Thought hard about this, I think I agree, and I guess as a minimum we can start with at least the ACPI+RNG patch, right? That will already address wireguard ...
Signature · 1 line
MST
Jason A. Donenfeld
maintainerHey Alex,
Quoted 3 lines · Alexander Graf
I agree on the slightly racy compromise and that it's a step into the right direction. Doing this is a no brainer IMHO and I like the proc based poll approach.
Alright. I'm going to email a more serious patch for that in the next few hours and you can have a look. Let's do that for 5.19.
Quoted 6 lines
I have an additional problem you might have an idea for with the poll based path. In addition to the clone notification, I'd need to know at which point everyone who was listening to a clone notification is finished acting up it. If I spawn a tiny VM to do "work", I want to know when it's safe to hand requests into it. How do I find out when that point in time is?
Seems tricky to solve. Even a count of current waiters and a generation number won't be sufficient, since it wouldn't take into account users who haven't _yet_ gotten to waiting. But maybe it's not the right problem to solve? Or somehow not necessary? For example, if the problem is a bit more constrained a solution becomes easier: you have a fixed/known set of readers that you know about, and you guarantee that they're all waiting before the fork. Then after the fork, they all do something to alert you in their poll()er, and you count up how many alerts you get until it matches the number of expected waiters. Would that work? It seems like anything more general than that is just butting heads with the racy compromise we're already making.
Jason
Quoted 30 lines · Jason A. Donenfeld
Hey Alex,
Quoted 3 lines · Alexander Graf
I agree on the slightly racy compromise and that it's a step into the right direction. Doing this is a no brainer IMHO and I like the proc based poll approach.
Alright. I'm going to email a more serious patch for that in the next few hours and you can have a look. Let's do that for 5.19.
Quoted 6 lines
I have an additional problem you might have an idea for with the poll based path. In addition to the clone notification, I'd need to know at which point everyone who was listening to a clone notification is finished acting up it. If I spawn a tiny VM to do "work", I want to know when it's safe to hand requests into it. How do I find out when that point in time is?
Seems tricky to solve. Even a count of current waiters and a generation number won't be sufficient, since it wouldn't take into account users who haven't _yet_ gotten to waiting. But maybe it's not the right problem to solve? Or somehow not necessary? For example, if the problem is a bit more constrained a solution becomes easier: you have a fixed/known set of readers that you know about, and you guarantee that they're all waiting before the fork. Then after the fork, they all do something to alert you in their poll()er, and you count up how many alerts you get until it matches the number of expected waiters. Would that work? It seems like anything more general than that is just butting heads with the racy compromise we're already making.
Quoted 1 line
Jason
I have some ideas here ... but can you explain the use-case a bit more?
Signature · 1 line
MST