Asahi Lina authored 2 years ago and Daniel Almeida committed 1 month ago

The XArray is an abstract data type which behaves like a very large
array of pointers. Add a Rust abstraction for this data type.

The initial implementation uses explicit locking on get operations and
returns a guard which blocks mutation, ensuring that the referenced
object remains alive. To avoid excessive serialization, users are
expected to use an inner type that can be efficiently cloned (such as
Arc<T>), and eagerly clone and drop the guard to unblock other users
after a lookup.

Future variants may support using RCU instead to avoid mutex locking.

This abstraction also introduces a reservation mechanism, which can be
used by alloc-capable XArrays to reserve a free slot without immediately
filling it, and then do so at a later time. If the reservation is
dropped without being filled, the slot is freed again for other users,
which eliminates the need for explicit cleanup code.

Signed-off-by: Asahi Lina <lina@asahilina.net>
---

Hi everyone!

This abstraction is part of the set of dependencies for the drm/asahi
Apple M1/M2 GPU driver.

The branch at [1] contains the full series of patches rebased on
upstream leading to the complete driver, for reference on how it is
intended to be used.

Thank you everyone who helped review this on GitHub [2]! I hope I didn't
forget any CCs...

Note that I dropped the convenience `Deref` impl for `Guard`, since I
couldn't figure out how to do it safely. Suggestions welcome, or we can
leave it for a future improvement ^^

[1] https://github.com/AsahiLinux/linux/tree/gpu/rebase-20230224
[2] https://github.com/Rust-for-Linux/linux/pull/952

Changes in v3:
- Updated to the error v2/v3 series API.
- Renamed `err` to `ret` for consistency with the other instance.
- Link to v2: https://lore.kernel.org/r/20230224-rust-xarray-v2-1-4eeb0134944c@asahilina.net
Changes in v2:
- Added Pin requirement for all XArray operations, to close a
  soundness hole due to the lock in the XArray (locks are not safe to
  move while locked). Creation does not require pinning in place, since
  the lock cannot be acquired at that point.
- Added safety note to Drop impl about why we don't need to do the lock
  unlock dance to ensure soundness in case of a dropped lock guard.
- Downstream drm/asahi driver was also rebased on this version to prove
  it works (previously it was still on a pre-v1 version).
- This still depends on the Error series (v1). v2 of that will need a
  trivial rename of Error::from_kernel_errno -> Error::from_errno. If
  this version of XArray ends up looking good, I'll send a trivial v4 of
  XArray with the rename, after sending the v2 of the Error series.
- Link to v1: https://lore.kernel.org/r/20230224-rust-xarray-v1-1-80f0904ce5d3@asahilina.net

Rust driver for ARM Mali CSF-based GPUs

The skeleton is basically taken from Nova and also rust_platform_driver.rs.

So far, this is just a very early-stage experiment, but it looks promissing:

- We use the same uAPI as Panthor, although this needs a bit of work, since
  bindgen cannot translate #defines into Rust.

- The DRM registration and IOCTL boilerplate is implemented. We can test
  devquery using igt.

- Basic iomem and register set implementation, so it's possible to program
the device. We can even attest that it works already.

- IRQ handling, so we can receive notifications from the device.

- We can boot the firmware.

Notably missing (apart from literally everything else):
- Job subission logic through drm_scheduler and completion through dma_fences
- VM_BIND support through drm_gpuvm
- Devfreq, pm_idle, etc.

The name "Tyr" is inspired by Norse mythology, reflecting ARM's tradition of
naming their GPUs after Nordic mythological figures and places.

Co-developed-by: Alice Ryhl <alice.ryhl@google.com>
Signed-off-by: Alice Ryhl <alice.ryhl@google.com>
Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>

Janne Grunau authored 5 months ago and Daniel Almeida committed 2 months ago

Required by the asahi driver.

Signed-off-by: Janne Grunau <j@jannau.net>

This patch is merely a stand-in for work that Abdiel is already carrying out.
It's not meant to be upstreamed.

Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>

This abstraction adds a small subset of the regulator API. This is thought to
be sufficient for the drivers we have now.

The abstraction maintains a count of how many times `enable()` was called. This
is so it can make sure that the calls to enable balance out with the calls to
disable, as required by the C API.

Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>

Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>

Viresh Kumar authored 2 months ago and Daniel Almeida committed 2 months ago

This adds very basic bindings for the clk framework, implements only
clk_get() and clk_put(). These are the bare minimum bindings required
for many users and are simple enough to add in the first attempt.

These will be used by Rust based cpufreq / OPP core to begin with.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>

FUJITA Tomonori authored 2 months ago and Daniel Almeida committed 2 months ago

Wait until a PHY becomes ready in the probe callback by
using read_poll_timeout function.

Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Reviewed-by: Gary Guo <gary@garyguo.net>
Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>

FUJITA Tomonori authored 2 months ago and Daniel Almeida committed 2 months ago

Add read_poll_timeout functions which poll periodically until a
condition is met or a timeout is reached.

The C's read_poll_timeout (include/linux/iopoll.h) is a complicated
macro and a simple wrapper for Rust doesn't work. So this implements
the same functionality in Rust.

The C version uses usleep_range() while the Rust version uses
fsleep(), which uses the best sleep method so it works with spans that
usleep_range() doesn't work nicely with.

Unlike the C version, __might_sleep() is used instead of might_sleep()
to show proper debug info; the file name and line
number. might_resched() could be added to match what the C version
does but this function works without it.

The sleep_before_read argument isn't supported since there is no user
for now. It's rarely used in the C version.

read_poll_timeout() can only be used in a nonatomic context. This
requirement is not checked by these abstractions, but it is intended
that klint [1] or a similar tool will be used to check it in the
future.

Link: https://rust-for-linux.com/klint

 [1]
Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>

FUJITA Tomonori authored 2 months ago and Daniel Almeida committed 2 months ago

Add Rust TIMEKEEPING and TIMER abstractions to the maintainers entry
respectively.

Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>

FUJITA Tomonori authored 2 months ago and Daniel Almeida committed 2 months ago

Add a wrapper for fsleep(), flexible sleep functions in
include/linux/delay.h which typically deals with hardware delays.

The kernel supports several sleep functions to handle various lengths
of delay. This adds fsleep(), automatically chooses the best sleep
method based on a duration.

sleep functions including fsleep() belongs to TIMERS, not
TIMEKEEPING. They are maintained separately. rust/kernel/time.rs is an
abstraction for TIMEKEEPING. To make Rust abstractions match the C
side, add rust/kernel/time/delay.rs for this wrapper.

fsleep() can only be used in a nonatomic context. This requirement is
not checked by these abstractions, but it is intended that klint [1]
or a similar tool will be used to check it in the future.

Link: https://rust-for-linux.com/klint

 [1]
Reviewed-by: Gary Guo <gary@garyguo.net>
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Reviewed-by: Fiona Behrens <me@kloenk.dev>
Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>

FUJITA Tomonori authored 2 months ago and Daniel Almeida committed 2 months ago

Introduce a type representing a specific point in time. We could use
the Ktime type but C's ktime_t is used for both timestamp and
timedelta. To avoid confusion, introduce a new Instant type for
timestamp.

Rename Ktime to Instant and modify their methods for timestamp.

Implement the subtraction operator for Instant:

Delta = Instant A - Instant B

Reviewed-by: Boqun Feng <boqun.feng@gmail.com>
Reviewed-by: Gary Guo <gary@garyguo.net>
Reviewed-by: Fiona Behrens <me@kloenk.dev>
Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>

FUJITA Tomonori authored 2 months ago and Daniel Almeida committed 2 months ago

Introduce a type representing a span of time. Define our own type
because `core::time::Duration` is large and could panic during
creation.

time::Ktime could be also used for time duration but timestamp and
timedelta are different so better to use a new type.

i64 is used instead of u64 to represent a span of time; some C drivers
uses negative Deltas and i64 is more compatible with Ktime using i64
too (e.g., ktime_[us|ms]_delta() APIs return i64 so we create Delta
object without type conversion.

i64 is used instead of bindings::ktime_t because when the ktime_t
type is used as timestamp, it represents values from 0 to
KTIME_MAX, which is different from Delta.

as_millis() method isn't used in this patchset. It's planned to be
used in Binder driver.

Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Reviewed-by: Gary Guo <gary@garyguo.net>
Reviewed-by: Fiona Behrens <me@kloenk.dev>
Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>

FUJITA Tomonori authored 2 months ago and Daniel Almeida committed 2 months ago

Add PartialEq/Eq/PartialOrd/Ord trait to Ktime so two Ktime instances
can be compared to determine whether a timeout is met or not.

Use the derive implements; we directly touch C's ktime_t rather than
using the C's accessors because it is more efficient and we already do
in the existing code (Ktime::sub).

Reviewed-by: Trevor Gross <tmgross@umich.edu>
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Reviewed-by: Gary Guo <gary@garyguo.net>
Reviewed-by: Fiona Behrens <me@kloenk.dev>
Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>

FUJITA Tomonori authored 2 months ago and Daniel Almeida committed 2 months ago

Add __might_sleep_precision(), Rust friendly version of
__might_sleep(), which takes a pointer to a string with the length
instead of a null-terminated string.

Rust's core::panic::Location::file(), which gives the file name of a
caller, doesn't provide a null-terminated
string. __might_sleep_precision() uses a precision specifier in the
printk format, which specifies the length of a string; a string
doesn't need to be a null-terminated.

Modify __might_sleep() to call __might_sleep_precision() but the
impact should be negligible. strlen() isn't called in a normal case;
it's called only when printing the error (sleeping function called
from invalid context).

Note that Location::file() providing a null-terminated string for
better C interoperability is under discussion [1].

Link: https://github.com/rust-lang/libs-team/issues/466

 [1]
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Co-developed-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>

Asahi Lina authored 1 year ago and Daniel Almeida committed 2 months ago

Switch from being a refcount wrapper itself to a transparent wrapper
around `bindings::drm_device`. The refcounted type then becomes
ARef<Device<T>>.

Signed-off-by: Asahi Lina <lina@asahilina.net>

Asahi Lina authored 2 years ago and Daniel Almeida committed 2 months ago

The DRM shmem helper includes common code useful for drivers which
allocate GEM objects as anonymous shmem. Add a Rust abstraction for
this. Drivers can choose the raw GEM implementation or the shmem layer,
depending on their needs.

Signed-off-by: Asahi Lina <lina@asahilina.net>

Asahi Lina authored 2 years ago and Daniel Almeida committed 2 months ago

There doesn't seem to be a way for the Rust bindings to get a
compile-time constant reference to drm_gem_shmem_vm_ops, so we need to
duplicate that structure in Rust... this isn't nice...

Signed-off-by: Asahi Lina <lina@asahilina.net>

Asahi Lina authored 2 years ago and Daniel Almeida committed 2 months ago

drm_mm provides a simple range allocator, useful for managing virtual
address ranges. Add a Rust abstraction to expose this module to Rust
drivers.

Signed-off-by: Asahi Lina <lina@asahilina.net>

Asahi Lina authored 4 months ago and Daniel Almeida committed 2 months ago

We cannot drop ARef<GpuVmBo<T>> references within the step_*() calls,
since the destructore takes the object lock but that is already locked
here. Instead of providing a method that the callback can use to obtain
a reference (which, when dropped, would deadlock), grab a reference
ourselves and pass it explicitly into the callback as a &ref. Thus, we
can drop it without locking again.

Signed-off-by: Asahi Lina <lina@asahilina.net>

Asahi Lina authored 4 months ago and Daniel Almeida committed 2 months ago

Signed-off-by: Asahi Lina <lina@asahilina.net>

Asahi Lina authored 11 months ago and Daniel Almeida committed 2 months ago

Signed-off-by: Asahi Lina <lina@asahilina.net>

Asahi Lina authored 1 year ago and Daniel Almeida committed 2 months ago

This is just for basic usage in the DRM shmem abstractions for implied
locking, not intended as a full DMA Reservation abstraction yet.

Signed-off-by: Asahi Lina <lina@asahilina.net>

This u64/usize mismatch will probably be gone on 6.14