Add AArch64 feature detection for FreeBSD #892

valpackett · 2019-08-25T14:58:26Z

and fix build on any ARM platform that's not Linux/Fuchsia.

I hope I didn't screw up any platforms.. :D

The all(target_os = "freebsd", target_arch = "aarch64") inside of an all(…, any(target_arch = "arm", target_arch = "aarch64")) is to exclude FreeBSD/arm 32-bit.

cc #387

# before
test digest::sha256::_8192                                    ... bench:      47,207 ns/iter (+/- 228) = 173 MB/s
test aead::seal_in_place::aes_128_gcm::tls12_16               ... bench:  12,084,200 ns/iter (+/- 15,248,988)

# after
test digest::sha256::_8192                                    ... bench:       8,350 ns/iter (+/- 142) = 981 MB/s
test aead::seal_in_place::aes_128_gcm::tls12_16               ... bench:     722,151 ns/iter (+/- 917,484)

briansmith

Relatively recently I've changed my policies so as to be more liberal in accepting ports to architectures and operating systems I don't generally use, so in general I'm now happy to accept a change like this. However, I do think we need to insist on having such ports include the necessary infrastructure for people to test the port independently of any help from the submitter.

Are you trying to get FreeBSD + Raspberry Pi working? If so, which Raspberry Pi?

Could you please add a script that I can run locally on Linux (and/or macOS) x86_64 that can be used to test this code, e.g. in QEMU?

And/or, if you are testing this directly on hardware, could you provide step-by-step instructions, "Purchase this device, do this to install the OS, do this to build ring's test suite for the device, do this to run the tests on the device."

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=240037 doesn't look very encouraging as far as getting a QEMU-based testing setup working.

briansmith · 2019-09-01T17:50:21Z

crypto/cpu-aarch64.c

+//
+// Only FreeBSD uses this right now.
+
+#ifdef __FreeBSD__


Is this really the best mechanism that FreeBSD has for this? They don't support the getauxval() API?

elf_aux_info (like getauxval but slightly different call signature - with a buffer length) was not present on aarch64 for a long time. It is there in -CURRENT now, but not in any release.

This mechanism is better, it seems to be what ARM intended, it's more similar to x86 cpuid. Linux supports it since 4.11 too.

briansmith · 2019-09-01T17:51:15Z

crypto/cpu-aarch64.c

+// Only FreeBSD uses this right now.
+
+#ifdef __FreeBSD__
+#pragma GCC diagnostic ignored "-Wgnu-statement-expression"


Why do we need to suppress this warning? It would be best to avoid needing to suppress it. If that's not practical, we should have a comment documenting why not.

This is because of the FreeBSD armreg.h macros, we won't need this if I switch to directly using assembly

briansmith · 2019-09-01T17:51:39Z

crypto/cpu-aarch64.c

+
+  uint64_t id_aa64isar0 = READ_SPECIALREG(ID_AA64ISAR0_EL1);
+
+  if (ID_AA64ISAR0_AES(id_aa64isar0) >= ID_AA64ISAR0_AES_BASE)


We need to follow the Google BoringSSL convention of always using braces.

briansmith · 2019-09-01T17:56:05Z

crypto/cpu-aarch64.c

+#pragma GCC diagnostic ignored "-Wgnu-statement-expression"
+#include <GFp/cpu.h>
+#include <GFp/arm_arch.h>
+#include <machine/armreg.h>


This would require us to have system header files available just to build for this target, which is something I've been trying to avoid. Maybe that's unavoidable?

READ_SPECIALREG is a macro that expands to inline asm. Maybe we should just have that inline asm here that exposes READ_SPECIALREG as a function callable from Rust, and then do the rest in Rust? IIUC this is ARM's ABI, so it's not going to change out from underneath us.

Yeah, that's a good idea.

valpackett · 2019-09-01T18:24:37Z

Are you trying to get FreeBSD + Raspberry Pi working? If so, which Raspberry Pi?

I'm on a full aarch64 desktop (Marvell MACCHIATObin + AMD GPU), typing on it right now :)

As for testing, EC2 a1 instances are super easy if you have an AWS account.

go to https://lists.freebsd.org/pipermail/freebsd-snapshots/
→ latest month
→ latest "New FreeBSD snapshots available: head" (and/or "…: stable/12") post
scroll to "FreeBSD/aarch64 EC2 AMIs are available in the following regions"
copy the AMI id for your region and use it to start an instance

briansmith · 2019-09-01T18:26:48Z

I believe I have an EspessoBin sitting around. Is AAarch64 ported to that? If so, I could try that. AWS is a little unrealistic because I'm not going to rent a (vritual) machine to test this over time.

valpackett · 2019-09-01T18:47:12Z

EspressoBin

Yeah, that works, you can boot the generic memstick image, but you need a DTB built from our tree (here I just built one) and here's an example u-boot command to boot. (That example uses USB, commands are a bit different for SD, and s/armada-3720-db.dtb/armada-3720-espressobin.dtb of course)

Also, for RPi3 and Pine64(-LTS), their specific memstick images should just work.

https://download.freebsd.org/ftp/snapshots/arm64/aarch64/ISO-IMAGES/13.0/

…hod) I agree to license my contributions to each file under the terms given at the top of each file I changed.

I agree to license my contributions to each file under the terms given at the top of each file I changed.

briansmith · 2019-09-01T21:39:02Z

I suggest we pick Raspberry Pi 3 or 4 or EspressoBin if that works for you. Then the step-by-step instructions would include basically how to build the SD card image for FreeBSD from Linux starting with the curl or wget of whatever, the dd or whatever to produce the image, how to configure the image to allow remote (probably SSH) access over ethernet, the dd or whatever to write it onto an SD card, connecting to the device over SSH, building ring's tests, getting them onto the device, running the tests on the device, and getting the results back to the host device.

valpackett · 2019-09-01T23:17:20Z

dang, you're really serious about this testing thing, why? Most projects, like Rust itself, do not test on every os/arch combination but do merge the support.

produce the image

You don't produce anything, just write the existing one o_0

building ring's tests, getting them onto the device

Why would you not build everything on the device?? 0_o

This is the most step-by-step I can get (for RPi 3):

wget -c "https://download.freebsd.org/ftp/snapshots/ISO-IMAGES/12.0/FreeBSD-12.0-STABLE-arm64-aarch64-RPI3-20190829-r351571.img.xz"
unxz FreeBSD-12.0-STABLE-arm64-aarch64-RPI3-20190829-r351571.img.xz
dd if=FreeBSD-12.0-STABLE-arm64-aarch64-RPI3-20190829-r351571.img of=/dev/… bs=1m
(there's a way to unxz to stdout and dd from stdin, but whatever)
insert the card, boot, it should get an IP address over DHCP on the ethernet port, and sshd will be running already
(i'm not going to tell you how to find the IP address… okay - watch the serial console, or watch your router)
ssh freebsd@what.ever.ip.address, password: freebsd
su, password: root
pkg install -y rust git (and whatever else you want but you shouldn't need anything else i think)
exit the root shell
git clone, cd, cargo test as usual

briansmith · 2019-09-02T01:19:33Z

dang, you're really serious about this testing thing, why? Most projects, like Rust itself, do not test on every os/arch combination but do merge the support.

To cut it short, I think this is a better way than that.

Why would you not build everything on the device?? 0_o

I have a 32 core ~4.0ghz x86-64 machine that runs an operating system I'm familiar with, which is much faster than the espressobin that I've booted like once :)

ghost · 2020-04-24T10:02:18Z

Can this be merged? I do not quite understand what the necessary condition to get this accepted is. It has been ~8 months since there was any activity here. PR still applies cleanly.

The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 git-svn-id: svn+ssh://svn.freebsd.org/ports/head@532914 35697150-7ecd-e111-bb59-0022644237b5

security/cloak: Unbreak on aarch64 The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 Approved by: ports-secteam blanket

The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892

The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 git-svn-id: svn+ssh://svn.freebsd.org/ports/head@532942 35697150-7ecd-e111-bb59-0022644237b5

The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892

www/zola: Unbreak on aarch64 The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 Approved by: ports-secteam blanket

The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 git-svn-id: svn+ssh://svn.freebsd.org/ports/head@532914 35697150-7ecd-e111-bb59-0022644237b5

The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 git-svn-id: svn+ssh://svn.freebsd.org/ports/head@532942 35697150-7ecd-e111-bb59-0022644237b5

briansmith

I'm leaning towards using #1007 instead of this one due to the concerns I raise in my other comments. WDYT?

briansmith · 2020-11-24T17:28:12Z

crypto/cpu-aarch64.c

+
+uint64_t GFp_aarch64_read_isar0(void) {
+  uint64_t val;
+  __asm __volatile("mrs %0, ID_AA64ISAR0_EL1" : "=&r" (val));


Is there a std::arch intrinsic for this now?

No, but stdarch uses this for its own detection:

https://github.com/rust-lang/stdarch/blob/a129c40f56882cae535b6861be88d848a6848bb8/crates/std_detect/src/detect/os/aarch64.rs

(but that's nightly only right now)

briansmith · 2020-11-24T17:29:56Z

src/cpu.rs

+        }
+
+        let mut features = 0;
+        let aa64isar0 = unsafe { GFp_aarch64_read_isar0() };


How do we know it is safe to execute the mrs instruction?

My understanding is that a conservative person should use getauxval() (or equivalent) to determine whether it is safe to execute mrs. Am I wrong?

It's guaranteed to be safe on FreeBSD 12.0 and newer.

HWCAP auxval (#1007) has been merged into 12.2, so I guess it's fine now too. (Used to only be available in -CURRENT.)

Thanks for the respond. I am really liking the idea of using getauxval because then it works just like Linux and we can avoid adding this tiny chunk of C code, so I think we should continue with #1007.

briansmith · 2020-11-24T17:53:05Z

Thanks for your PR. I'm going to close this as a duplicate of #1007. To be frank, it's because it's a lot easier for me to maintain it if it shares the logic of the Linux implementation. I have nothing against FreeBSD but I don't have any FreeBSD systems.

security/cloak: Unbreak on aarch64 The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 Approved by: ports-secteam blanket

www/zola: Unbreak on aarch64 The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 Approved by: ports-secteam blanket

briansmith requested changes Sep 1, 2019

View reviewed changes

valpackett added 2 commits September 1, 2019 21:51

Fix build on aarch64 on arbitrary OS (without a feature detection met…

19df2c6

…hod) I agree to license my contributions to each file under the terms given at the top of each file I changed.

Add CPU feature detection for FreeBSD/aarch64

300516a

I agree to license my contributions to each file under the terms given at the top of each file I changed.

valpackett force-pushed the fbsd-aarch64 branch from 79fd0f8 to 300516a Compare September 1, 2019 19:05

uqs pushed a commit to freebsd/freebsd-ports that referenced this pull request Apr 25, 2020

security/cloak: Unbreak on aarch64

0b8c2a6

The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892

uqs pushed a commit to freebsd/freebsd-ports that referenced this pull request Apr 25, 2020

www/zola: Unbreak on aarch64

407b16b

The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892

uqs pushed a commit to freebsd/freebsd-ports that referenced this pull request Apr 25, 2020

MFH: r532942

5900bcc

www/zola: Unbreak on aarch64 The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 Approved by: ports-secteam blanket

briansmith reviewed Nov 24, 2020

View reviewed changes

briansmith closed this Nov 24, 2020

uqs pushed a commit to freebsd/freebsd-ports that referenced this pull request Apr 1, 2021

MFH: r532942

93336e3

www/zola: Unbreak on aarch64 The ring crate is missing aarch64 support on FreeBSD. Apply [1] that adds support for it. [1] briansmith/ring#892 Approved by: ports-secteam blanket

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add AArch64 feature detection for FreeBSD #892

Add AArch64 feature detection for FreeBSD #892

valpackett commented Aug 25, 2019 •

edited

Loading

briansmith left a comment

briansmith Sep 1, 2019

valpackett Sep 1, 2019 •

edited

Loading

briansmith Sep 1, 2019

valpackett Sep 1, 2019

briansmith Sep 1, 2019

briansmith Sep 1, 2019

valpackett Sep 1, 2019

valpackett commented Sep 1, 2019

briansmith commented Sep 1, 2019

valpackett commented Sep 1, 2019

briansmith commented Sep 1, 2019

valpackett commented Sep 1, 2019

briansmith commented Sep 2, 2019

ghost commented Apr 24, 2020

briansmith left a comment

briansmith Nov 24, 2020

valpackett Nov 24, 2020

briansmith Nov 24, 2020

valpackett Nov 24, 2020

briansmith Nov 24, 2020

briansmith commented Nov 24, 2020


		uint64_t id_aa64isar0 = READ_SPECIALREG(ID_AA64ISAR0_EL1);

		if (ID_AA64ISAR0_AES(id_aa64isar0) >= ID_AA64ISAR0_AES_BASE)

Add AArch64 feature detection for FreeBSD #892

Add AArch64 feature detection for FreeBSD #892

Conversation

valpackett commented Aug 25, 2019 • edited Loading

briansmith left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

valpackett Sep 1, 2019 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

valpackett commented Sep 1, 2019

briansmith commented Sep 1, 2019

valpackett commented Sep 1, 2019

briansmith commented Sep 1, 2019

valpackett commented Sep 1, 2019

briansmith commented Sep 2, 2019

ghost commented Apr 24, 2020

briansmith left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

briansmith commented Nov 24, 2020

valpackett commented Aug 25, 2019 •

edited

Loading

valpackett Sep 1, 2019 •

edited

Loading