Narrow the CPPFLAGS for drm subsystems.

No need for the whole kernel to be built with -I flags for the drm
header files.

Move Linux SRCU to common.

Move Linux RCU to common.

Draft Linux tasklet implementation.

branches: 1.1.2; 1.1.6; 1.1.10; 1.1.22;
Missed this file in the linux_work move.

linux asm/barrier.h: Fix !MULTIPROCESSOR build.

drm: Eliminate __HAVE_ATOMIC_AS_MEMBAR conditionals.

Discussed on tech-kern:
https://mail-index.netbsd.org/tech-kern/2023/02/23/msg028729.html

branches: 1.19.4;
linux: Add alpha definitions of mb/wmb/rmb too for fun.

Who wants to plug an amdgpu into their alpha?

linux: Clarify error message in asm/barrier.h. Not just for drm.

linux: Add mips definitions for mb, wmb, rmb.

linux: Use dsb(sy/st/ld) from arm/cpufunc.h for mb/wmb/rmb.

Works on more Arm CPU types this way. From skrll.

drm: Fix missing barrier definitions on arm.

The definition will work on armv7; don't think we have any drmkms
drivers on armv<7.

drm: Comment *mb() vs smp_*mb() and omit default for *mb().

In general membar_*() is _not_ enough for *mb(), because membar_*()
is only for CPU/CPU synchronization on normal memory, and *mb()
requires machine-dependent I/O logic. Leave it as an `#error' case
if the architecture isn't listed here. But membar_*() is OK for
smp_*().

Fill in powerpc and sparc while here, as examples of what the
now-removed defaults failed to do. If you want to use drm on a mips
or alpha system yer gonna hafta fill this in yourself.

linux: Convert various API shims to use membar_release/acquire.

linux: Define mb/wmb/rmb as mfence/sfence/lfence on i386 too.

Not the same as membar_*; drm uses mb to order special instructions
like clflush which require mfence for ordering.

linux: For uniprocessor, define smp_*mb as __insn_barrier.

Otherwise ordering might not be guaranteed in the event of
preemption.

XXX pullup

linux: Eliminate read_barrier_depends, smp_read_barrier_depends.

These no longer exist in Linux; instead, Linux now uses READ_ONCE.

linux: Define mb/wmb/rmb for x86.

These are _not_ SMP barriers; on Linux they specifically mean
mfence/sfence/lfence. Although membar_sync and mb agree, and
membar_consumer and rmb agree, membar_producer and wmb _do not_ agree
on x86 -- Linux's wmb is specifically sfence, which is important only
for unusual memory types, whereas our membar_producer is limited to
ordinary memory.

Use adequate memory barrier for load-acquire semantics.

It is unfortunate that we chose the name membar_enter to have the
wrong semantics for this.

Move smp_mb_before/after_atomic from linux/atomic.h to asm/barrier.h.

Side-load linux/compiler.h in asm/barrier.h.

Fix mb/wmb/rmb on aarch64 to use dsb instead of dmb.

For module build assume multiprocessor.

branches: 1.3.2; 1.3.4; 1.3.6;
keep dsb in vchiq code.

dsb refers to an arm instruction, so it won't be used on MI code.

add dsb macro from vchiq

Move barrier to common so it can be shared with vchiq

linux: Make WARN_ONCE actually happen only once.

linux: Move BUILD_BUG_* to common linux/build_bug.h.

Make BUILD_BUG_ON_INVALID work on bit-field expressions.

Move linux/bug.h and asm/bug.h to common.

Define Linux byte order macros __BIG_ENDIAN/__LITTLE_ENDIAN.

branches: 1.1.2; 1.1.6;
Missed a drm2 -> common move.

branches: 1.1.2; 1.1.6;
Fix undoing of -Wpointer-arith workaround.

linux asm/vmware.h: Fill this in with hypercall asm.

Not sure if it's necessary to support vmcall or vmmcall -- if so,
need to deal with binary patching inline asm, ugh.

linux/bitops: Fix overestimate for BITS_TO_LONGS(9)

Fortunately, this seems harmless except for allocating
excessive buffer memory.

Pointed out by nonaka@, OK riastradh@.

branches: 1.16.4; 1.16.10;
Move Linux atomic bitops from linux/atomic.h to linux/bitops.h.

provide BITS_PER_TYPE


Author: Maya Rashish <maya@NetBSD.org>

Add a BIT_MASK in bits.h, move it to common so bitops.h can sideload

match linux.


Author: Maya Rashish <maya@NetBSD.org>

define BITS_PER_LONG. use compiler-defined macro instead of sizeof.


Author: Maya Rashish <maya@NetBSD.org>

Remove duplicate definition of BITS_PER_BYTE


Author: Maya Rashish <maya@NetBSD.org>

sign_extend64, GENMASK_ULL, for_each_clear_bit

Add BITS_PER_BYTE and BIT_ULL. Fix type of BIT.

Add what appears to be the fixes to CVE-2019-0154, CVE-2019-0155.
This commit requires review, but I'd also like it to be tested by others
while it is being reviewed.

CVE-2019-0155:
It was discovered that the Intel i915 graphics chipsets allowed userspace
to modify page table entries via writes to MMIO from the Blitter Command
Streamer and expose kernel memory information. A local attacker could use
this to expose sensitive information or possibly elevate privileges.

CVE-2019-0154:
It was discovered that the Intel i915 graphics chipsets could cause
a system hang when userspace performed a read from GT memory mapped
input output (MMIO) when the product is in certain low power states.
A local attacker could use this to cause a denial of service.

From upstream commits to linux-4.4.y:

branches: 1.8.2; 1.8.6; 1.8.8;
Give find_first/next_set/zero_bit a chance to work.

Self-review comments on the draft thrown together in a hurry two
weeks ago:

- I bet I screwed up for_each_set_bit or something.
- There might be several bits wrong in this __find_next_bit routine!
- for_each_set_bit is busted.
- Maybe I should test this stuff before I commit it.
- Hey, cool, buffer overflow here.
- What was I thinking.
- how did this even
- Was I drunk? I'm never drunk.
- I don't even drink.
- It wasn't even that late at night...

Move hweight8 next to its cousins.

Add find_first_bit, find_next_bit, for_each_set_bit.

Implement find_next_zero_bit. Define find_first_zero_bit in terms of it.

Add fls to <linux/bitops.h>.

Add hweight64.

... Provide GENMASK


Author: coypu <coypu@sdf.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

move bitops.h so we can include it from kernel.h

match linux side-loading of this header.


Author: coypu <coypu@sdf.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Add a BIT_MASK in bits.h, move it to common so bitops.h can sideload

match linux.


Author: Maya Rashish <maya@NetBSD.org>

Move linux/bug.h and asm/bug.h to common.

drm: Make BUILD_BUG_ON_ZERO do something.

stub BUILD_BUG_ON_ZERO, from openbsd.

This is a dual purpose macro that asserts but can also be used in
expressions. Provide the expression implementation.


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

linux: Simplify build_bug.h and add static_assert.

linux: Move BUILD_BUG_* to common linux/build_bug.h.

linux/compiler.h: Add missing barriers in READ_ONCE, WRITE_ONCE.

linux: Convert various API shims to use membar_release/acquire.

drm/vmwgfx: First draft. Passes make depend, doesn't build yet.

linux: Add __same_type and __must_be_array.

linux: Add offsetofend.

Move attributes, barrier(), and likely/unlikely to linux/compiler.h.

Define smp_store_mb, smp_store_release.

Replace ACCESS_ONCE by READ_ONCE, WRITE_ONCE.

Move <linux/compiler.h> to common.

linux: Timeout audit -- linux/completion.h.

linux: wait_for_completion

sys/kernel.h for hardclock_ticks


Author: Maya Rashish <maya@NetBSD.org>

sys/param.h for KASSERT etc


Author: Maya Rashish <maya@NetBSD.org>

Implement wait_for_completion_timeout.

hardclock_ticks -> getticks()

Add a reinit_completion function

branches: 1.5.2; 1.5.4;
IPL_VM -> IPL_SCHED

branches: 1.4.4;
Add some notes on using these shims to port Linux kernel code.

branches: 1.3.2; 1.3.4;
Use a spinlock for completions. Makes vchiq pass LOCKDEBUG where other
spinlocks where held when trying to use the completion API.

branches: 1.2.2;
Explicitly kassert the precondition c_done != 0 in _completion_claim.

Move <linux/completion.h> from drm2 to external/bsd/common.

linux/dmapool.h: Start filling out stubs.

fix various problems i've seen where cv_*wait*() return ERESTART,
which is -3 in netbsd, which we have mapped linux ERESTARTSYS to.

this has a problem because linux code often returns errors and
pointers in the same value, and pointer values between -4095 and
-1 are considered as error returns, but -3 ends up as 3 and thus
is not considered an error, and mayhem ensues.

with this in place my kabylake system seems actually stable, i
have not triggered any of my prior issues in almost 4 weeks now.


Taylor asked me to write up a description and then wrote most of
the text below for me :-)


In Linux code, we always work with ERESTARTSYS so the code meaning
start over is a positive NetBSD errno safe for PTR_ERR/ERR_PTR.
To achieve this:
1. adapt all cv_waits that return to Linux so they map ERESTART to
ERESTARTSYS, and
2. adapt all returns to userland so they convert ERESTARTSYS to
ERESTART.
Leave EINTR and all other error codes alone.

Define PTR_ERR_OR_ZERO.

branches: 1.1.2; 1.1.6; 1.1.10; 1.1.16; 1.1.22; 1.1.24; 1.1.26;
Move linux/err.h into common/include.

OK riastradh@

EHWPOISON

fix various problems i've seen where cv_*wait*() return ERESTART,
which is -3 in netbsd, which we have mapped linux ERESTARTSYS to.

this has a problem because linux code often returns errors and
pointers in the same value, and pointer values between -4095 and
-1 are considered as error returns, but -3 ends up as 3 and thus
is not considered an error, and mayhem ensues.

with this in place my kabylake system seems actually stable, i
have not triggered any of my prior issues in almost 4 weeks now.


Taylor asked me to write up a description and then wrote most of
the text below for me :-)


In Linux code, we always work with ERESTARTSYS so the code meaning
start over is a positive NetBSD errno safe for PTR_ERR/ERR_PTR.
To achieve this:
1. adapt all cv_waits that return to Linux so they map ERESTART to
ERESTARTSYS, and
2. adapt all returns to userland so they convert ERESTARTSYS to
ERESTART.
Leave EINTR and all other error codes alone.

branches: 1.3.2; 1.3.6; 1.3.24; 1.3.32;
Make it build and boot on my test machines.

Screen blanks on boot on the Ivy Bridge system with

DRM error in cpt_serr_int_handler: PCH transcoder A FIFO underrun

But after that everything is OK. Appears to be an upstream problem.
To investigate...

I think there's a cache flushing issue somewhere -- there are little
display artefacts on my T60.

Add some notes on using these shims to port Linux kernel code.

branches: 1.1.2; 1.1.4;
Move <linux/errno.h> from drm2 to common.

linux/export.h: Side-load linux/stringify.h.

Expected by some users.

Include linux/export.h in linux/printk.h, move to common

linux does this through:
printk.h -> linkage.h -> export.h

this will require less patching of drm files.

Author: Maya Rashish <maya@NetBSD.org>

linux/frame.h: STACK_FRAME_NON_STANDARD

gfpflags_allow_blocking

Define __GFP_RETRY_MAYFAIL as an alias for __GFP_NORETRY.

branches: 1.1.2; 1.1.6;
gfp.h and slab.h are now used in kernel.h, so move to common.

linux/kernel.h: S16_MIN, S32_MIN, S64_MIN

amdgpu: amdgpu_gfx.c

...plus a bogus stub for sscanf, hoping it's not that important

linux: might_sleep_if

drm: Another pass over i915. Most of i915 gem builds now.

linux: S16_MAX, S32_MAX, S64_MAX

drm: Add __is_constexpr.

Define Linux byte order macros __BIG_ENDIAN/__LITTLE_ENDIAN.

ALIGN_DOWN seems to be an alias for round_down.

Move attributes, barrier(), and likely/unlikely to linux/compiler.h.

DEFINE_STATIC_KEY_FALSE

__maybe_unused, typecheck

might_sleep, add_taint

Define empty __deprecated.

DIV_ROUND_DOWN_ULL

Define u64_to_user_ptr.

Replace ACCESS_ONCE by READ_ONCE, WRITE_ONCE.

Define clamp_val.

Actually use the type in min_t/max_t.

Otherwise we do not get the truncation we might expect. Probably not
a big deal, but we should match the semantics.

XXX pullup-7
XXX pullup-8

Add some side loads to <linux/kernel.h> to help reduce diffs.

Move IS_REACHABLE to <linux/kernel.h> to avoid double definition.

Define __BIG_ENDIAN or __LITTLE_ENDIAN depending on target endianness

Reduce diffs by side-loading some header files like Linux.

From riastradh

Implement (obsolete) simple_strtol stub; reduce diff.

branches: 1.23.2;
use __inline
add missing __printflike attributes

fix another sign compare warning:

cast 'int ret' to size_t when comparing against another size_t
and 'ret' has known to be 0 or positive.

Define __must_check to actually use warn_unused_result.

Copy va_list before using it twice.

Just define ACCESS_ONCE rather than patch it out.

Define noinline in <linux/kernel.h>.

Define DIV_ROUND_CLOSEST_ULL and div_s64.

Define min3/max3. XXX: Multiple evaluation. Hope it's OK!

Add kvasprintf to <linux/kernel.h>.

Implement Linux kasprintf.

move bitops.h so we can include it from kernel.h

match linux side-loading of this header.


Author: coypu <coypu@sdf.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

some more definitions necessary

ifdef out all the i915 gtt virtual memory stuff that errors,
unless it looks exactly like the old code and then merge the
netbsd ifdefs for it.

we don't want to use their ALIGN(, which has more arguments,
use the old alt function.
merge in the old intel_acpi.c code.
don't duplicate DECLARE_BITMAP


Author: coypu <coypu@sdf.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

more declarations of things we need.

more power management ops.


Author: coypu <coypu@sdf.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

move printk to common so we can reasonably include it from kernel.h
(linux side-loads the same)


Author: coypu <coypu@sdf.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

libkern min/max is 32-bit. Linux min/max is generic. @!*#@!$&

XXX pullup

branches: 1.8.10; 1.8.16; 1.8.18;
Tweak IS_ENABLED to allow CONFIG_FOO values.

No functional change to drm2.

LGTM from riastradh@

Add container_of to libkern.

Given x = &c->f, container_of(x, T, f) yields c, where T is the type
of c.

Discussed on tech-kern a while ago:

https://mail-index.netbsd.org/tech-kern/2013/03/21/msg015131.html

Add clamp to <linux/kernel.h>.

Add kstrtol.

branches: 1.4.2; 1.4.4; 1.4.6;
First whack at radeon driver.

No hardware to test yet, but it builds.

Make it build and boot on my test machines.

Screen blanks on boot on the Ivy Bridge system with

DRM error in cpt_serr_int_handler: PCH transcoder A FIFO underrun

But after that everything is OK. Appears to be an upstream problem.
To investigate...

I think there's a cache flushing issue somewhere -- there are little
display artefacts on my T60.

branches: 1.2.2; 1.2.4; 1.2.6;
Merge changes from drm2 <linux/kernel.h> to common <linux/kernel.h>.

Pull across some linux header files from riastradh-drm2, but put them
in a shared location.

Riastradh: nick, OK.

linux: Stub linux/kmemleak.h.

i915: more work

i915: another whack at it

i915: More progress.

linux/list.h: Add list_safe_reset_next.

Provide list_for_each_entry_safe_reverse and reorder for consistency


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

list_for_each_entry_from_reverse

list_bulk_move_tail

Omit outdated comment.

provide list_for_each_entry_from (XXX needs review)


Author: Maya Rashish <maya@NetBSD.org>

provide list_is_first


Author: Maya Rashish <maya@NetBSD.org>

list_is_last

__list_del_entry, and poison list_del.

Move struct list_head to <linux/types.h>.

Add LIST_HEAD to <linux/nbsd-namespace.h> too.

branches: 1.18.6;
Define list_splice_init.

Make sure to call pslist_entry_init.

Give hlist iteration half a chance of working, oops.

Implement list_splice_tail_init.

Add list_replace_init.

Fix hlist_for_each_entry_safe: temp is hlist_node, not element.

Fix order of arguments of hlist_add_head_rcu.

Fix hlist_for_each_entry_rcu like hlist_for_each_entry.

Define list_entry_first_or_null.

Fix hlist iteration again with more fields and null checks.

Fix hlist_for_each_entry to use the specified field.

Rewrite linux hlist with our pslist.

Upstream uses RCU for this now, so we need it.

Rewrite hashtable.h locally so it might work here.

branches: 1.5.18; 1.5.20;
Add some Linux list routines.

branches: 1.4.2; 1.4.4;
First whack at radeon driver.

No hardware to test yet, but it builds.

Make it build and boot on my test machines.

Screen blanks on boot on the Ivy Bridge system with

DRM error in cpt_serr_int_handler: PCH transcoder A FIFO underrun

But after that everything is OK. Appears to be an upstream problem.
To investigate...

I think there's a cache flushing issue somewhere -- there are little
display artefacts on my T60.

Add some notes on using these shims to port Linux kernel code.

branches: 1.1.2; 1.1.4; 1.1.6;
Pull across some linux header files from riastradh-drm2, but put them
in a shared location.

Riastradh: nick, OK.

branches: 1.1.2; 1.1.6;
Move another file from drm2 to common for <linux/kernel.h>.

linux asm/uaccess.h: strncpy_from_user

(which does not actually have strncpy semantics, naturally, from what
I can gather)

linux: Allow non-VLA structs in struct_size.

Upstream sometimes uses

struct foo {
int x;
short y[1];
};

instead of short y[] with struct_size.

linux: Move linux/overflow.h from drm2 to common.

Needed so linux/slab.h can side-load it.

drm(4): make pr_debug equivalent to aprint_debug

significantly reduces the default spam from amdgpu(4).

branches: 1.13.4;
drm: Reduce verbosity of drm kernel messages.

amdgpu_fb.c
amdgpu_fence.c
amdgpu_gart.c
amdgpu_fixpt31_32.c
amdgpu_fiji_smumgr.c
amdgpu_gem.c
amdgpu_gfx_v10_0.c
amdgpu_gfx_v6_0.c through amdgpu_gfx_v9_0.c
amdgpu_gmc.c
amdgpu_gmc_v6_0.c through amdgpu_gmc_v10_0.c
amdgpu_gtt_mgr.c
some amdgpu_h* files
some amdgpu_i* files
some amdgpu_j* files
amdgpu_kms.c
some amdgpu_m* and amdgpu_n* files

i915: hack hack hack

pr_info_ratelimited, not really

Include linux/export.h in linux/printk.h, move to common

linux does this through:
printk.h -> linkage.h -> export.h

this will require less patching of drm files.

Author: Maya Rashish <maya@NetBSD.org>

Define pr_warn.

Define struct va_format.

branches: 1.6.2; 1.6.6;
Add KERN_{EMERG,ALERT,NOTICE}

Sort KERN_* macros according to severity (worst first)

s/drm // in the KERN_* macros as this is common

Define pr_debug = printf.

provide pr_notice
avoid needing unregister_oom_notifier
use intel_register_dsm_handler(dev), as older drm does
provide dummy async_schedule
provide atomic_or
match side-loading


Author: coypu <coypu@sdf.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

move printk to common so we can reasonably include it from kernel.h
(linux side-loads the same)


Author: coypu <coypu@sdf.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Make intel_psr.c build
netbsd-ify
Stub out intel_lpe_audio.c (for now?).
Hammer intel_opregion.c into shape.
Hack up intel_lrc.c
Adapt intel_reset.c
Stub out intel_region_lmem.c
Adapt intel_ring_submission.c
Adapt intel_rps.c
Adapt intel_sdvo.c
Adapt intel_runtime_pm.c
Adapt intel_sideband.c
Adapt intel_sprite.c
Adapt intel_tv.c
Adapt intel_tc.c
Adapt intel_sseu.c
Adapt intel_timeline.c
Adapt intel_uc_fw.c and fix typo
Fix up intel_uncore.c
intel_vga.c
intel_wakeref.c
intel_vdsc.c
intel_wopcm.c
intel_workarounds.c
i915_globals.c
Fix up intelfb.c
Stub out intel_guc_log.c
Stub out intel_dsi_dcs_backlight.c
i915_perf.c
intel_hdmi.c
intel_hdcp.c


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

linux: Add cond_synchronize_rcu.

linux: rcu_replace_pointer

linux/rcupdate: Stub init_rcu_head, destroy_rcu_head.

Simplify rcu_* definitions with atomic_load/store_*.

linux/bug.h for WARN_ON


Author: Maya Rashish <maya@NetBSD.org>

Omit __insn_barrier where implied or subsumed by volatile.

Move Linux RCU to common.

Reduce code duplication: kmem_cache_create() is now exactly the same as
kmem_cache_create_dtor() except for the dtor argument, so implement
the former in terms of the latter.

Do the last change differently:

Instead of having a pre-destruct hook, put knowledge of passive
serialization into the pool allocator directly, enabled by PR_PSERIALIZE
when the pool / pool_cache is initialized. This will guarantee that
a passive serialization barrier will be performed before the object's
destructor is called, or before the page containing the object is freed
back to the system (in the case of no destructor). Note that the internal
allocator overhead is different when PR_PSERIALIZE is used (it implies
PR_NOTOUCH, because the objects must remain in a valid state).

In the DRM Linux API shim, this allows us to remove the custom page
allocator for SLAB_TYPESAFE_BY_RCU.

- For kmem_cache_create_dtor(), use a pre-destructor to issue the
synchronize_rcu() if the caller uses SLAB_TYPESAFE_BY_RCU. A
special pool allocator is not required in this case.
- Now that SLAB_TYPESAFE_BY_RCU does the right thing, no need to
call synchronize_rcu() in __i915_request_dtor().

linux: Side-load linux/overflow.h in linux/slab.h.

linux: Allow GFP_NOWAIT too.

linux: Use kmem directly for Linux kmalloc.

Take advantage of this to do LOCKDEBUG_MEM_CHECK at the point of
kfree_rcu rather than in the RCU GC thread.

i915: Defer destroying request lock and fence until page destruction.

kmem_cache is used at IPL_VM


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Fix inverted notion of whether to wait for allocation


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Define kinda bogus ARCH_KMALLOC_MINALIGN that's probably enough.

KMEM_CACHE, SLAB_TYPESAFE_BY_RCU

Define kmem_cache_shrink = pool_cache_reclaim.

branches: 1.1.2; 1.1.6;
gfp.h and slab.h are now used in kernel.h, so move to common.

Make intel_psr.c build
netbsd-ify
Stub out intel_lpe_audio.c (for now?).
Hammer intel_opregion.c into shape.
Hack up intel_lrc.c
Adapt intel_reset.c
Stub out intel_region_lmem.c
Adapt intel_ring_submission.c
Adapt intel_rps.c
Adapt intel_sdvo.c
Adapt intel_runtime_pm.c
Adapt intel_sideband.c
Adapt intel_sprite.c
Adapt intel_tv.c
Adapt intel_tc.c
Adapt intel_sseu.c
Adapt intel_timeline.c
Adapt intel_uc_fw.c and fix typo
Fix up intel_uncore.c
intel_vga.c
intel_wakeref.c
intel_vdsc.c
intel_wopcm.c
intel_workarounds.c
i915_globals.c
Fix up intelfb.c
Stub out intel_guc_log.c
Stub out intel_dsi_dcs_backlight.c
i915_perf.c
intel_hdmi.c
intel_hdcp.c


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Call it srcu_struct like linux does.


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Linux namespace for srcu.

Move Linux SRCU to common.

linux: Add offsetofend.

<linux/stringify.h> for __stringify.

linux: Add missing symbol namespacing for tasklet, sync_file.

Make intel_psr.c build
netbsd-ify
Stub out intel_lpe_audio.c (for now?).
Hammer intel_opregion.c into shape.
Hack up intel_lrc.c
Adapt intel_reset.c
Stub out intel_region_lmem.c
Adapt intel_ring_submission.c
Adapt intel_rps.c
Adapt intel_sdvo.c
Adapt intel_runtime_pm.c
Adapt intel_sideband.c
Adapt intel_sprite.c
Adapt intel_tv.c
Adapt intel_tc.c
Adapt intel_sseu.c
Adapt intel_timeline.c
Adapt intel_uc_fw.c and fix typo
Fix up intel_uncore.c
intel_vga.c
intel_wakeref.c
intel_vdsc.c
intel_wopcm.c
intel_workarounds.c
i915_globals.c
Fix up intelfb.c
Stub out intel_guc_log.c
Stub out intel_dsi_dcs_backlight.c
i915_perf.c
intel_hdmi.c
intel_hdcp.c


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Rework linux_tasklet.c a little.

Provide some more hacks for i915's grubby paws.

Add some i915drmkms additions to the tasklet API.

Make a bit more room in the columns here.

No tasklet_disable_nosync implemented at the moment.

Draft Linux tasklet implementation.

Move dma_addr_t to linux/types.h.

Move struct list_head to <linux/types.h>.

Move <linux/types.h> to common.

linux/workqueue: Draft queue_rcu_work.

i915: another whack at it

stub struct rcu_work


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Side-load linux/stringify.h in linux/workqueue.h.

(via linux/timer.h upstream)

alloc_workqueue

Stubs for init/destroy delayed work on stack.

Constify work_pending, delayed_work_pending.

drain_workqueue

Add work_pending, delayed_work_pending.

Implement return value of flush_work, flush_delayed_work.

Namespacing for system_unbound_wq.

New system_unbound_wq.

Rework Linux workqueue synchronization yet again.

- Use a low bit in the pointer to the queue, rather than whether the
pointer is null or not, to determine whether the work item is
queued/scheduled.

- _Preserve_ the pointer to the queue after we release the work.

- Release the work _before_ executing the function, not after.

This simplifies some things: we no longer have to distinguish whether
the work is queued or running in the functions where we are trying to
modify it. The pointer has to be preserved even after the work is
released so that we can flush the workqueue after the work has been
released.

Avoid the prospect of callout calls piling up.

Don't ever callout_schedule the callout while an existing call may be
in progress.

Echo some cases from cancel_delayed_work in flush_delayed_work.

Fix semantics of flush_work and flush_delayed_work.

- Change return type to void.
=> Upstream it is bool, but exactly one of hundreds of callers
actually use it, and I don't think the semantics is clear.

- Make sure to wait for whichever of the current work _and_ the next
batch queued is currently there in the workqueue.

- Don't retry a cancelled callout. Cancellation in the state
DELAYED_WORK_CANCELLED is guaranteed.

Add missing namespace for flush_scheduled_work.

Draft rewrite of Linux workqueue reimplementation.

Just use an explicit thread; don't attempt to fudge it with
workqueue(9). No doubt there are various mistakes in here, but they
should be easier to get right than the mega-kludgerific nonsense that
preceded this draft.

Add system_power_efficient_wq.

Namespace system_long_wq.

Implement flush_delayed_work.

Also Linux wants work->func to work.

Add mega-kludgerific current_work().

Add system_long_wq.

Define dummy INIT_WORK_ONSTACK and destroy_work_onstack.

branches: 1.1.2; 1.1.6; 1.1.10; 1.1.22; 1.1.24; 1.1.26;
Move linux_work.c and workqueue.h from sys/external/bsd/drm2 to
sys/external/common so that they can be used by others.

LGTM from riastradh@

These should really live outside sys/external, but that can be decided
later

linux: Fix rcu_barrier so it actually waits for everything.

linux: Rate-limit RCU GC thread as a cheap experiment.

linux: Use kmem directly for Linux kmalloc.

Take advantage of this to do LOCKDEBUG_MEM_CHECK at the point of
kfree_rcu rather than in the RCU GC thread.

Make intel_psr.c build
netbsd-ify
Stub out intel_lpe_audio.c (for now?).
Hammer intel_opregion.c into shape.
Hack up intel_lrc.c
Adapt intel_reset.c
Stub out intel_region_lmem.c
Adapt intel_ring_submission.c
Adapt intel_rps.c
Adapt intel_sdvo.c
Adapt intel_runtime_pm.c
Adapt intel_sideband.c
Adapt intel_sprite.c
Adapt intel_tv.c
Adapt intel_tc.c
Adapt intel_sseu.c
Adapt intel_timeline.c
Adapt intel_uc_fw.c and fix typo
Fix up intel_uncore.c
intel_vga.c
intel_wakeref.c
intel_vdsc.c
intel_wopcm.c
intel_workarounds.c
i915_globals.c
Fix up intelfb.c
Stub out intel_guc_log.c
Stub out intel_dsi_dcs_backlight.c
i915_perf.c
intel_hdmi.c
intel_hdcp.c


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

linux: Add cond_synchronize_rcu.

linux/rcupdate: Stub init_rcu_head, destroy_rcu_head.

Move Linux RCU to common.

Make intel_psr.c build
netbsd-ify
Stub out intel_lpe_audio.c (for now?).
Hammer intel_opregion.c into shape.
Hack up intel_lrc.c
Adapt intel_reset.c
Stub out intel_region_lmem.c
Adapt intel_ring_submission.c
Adapt intel_rps.c
Adapt intel_sdvo.c
Adapt intel_runtime_pm.c
Adapt intel_sideband.c
Adapt intel_sprite.c
Adapt intel_tv.c
Adapt intel_tc.c
Adapt intel_sseu.c
Adapt intel_timeline.c
Adapt intel_uc_fw.c and fix typo
Fix up intel_uncore.c
intel_vga.c
intel_wakeref.c
intel_vdsc.c
intel_wopcm.c
intel_workarounds.c
i915_globals.c
Fix up intelfb.c
Stub out intel_guc_log.c
Stub out intel_dsi_dcs_backlight.c
i915_perf.c
intel_hdmi.c
intel_hdcp.c


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

'struct srcu' is actually called 'struct srcu_struct'


Author: Iain Hibbert <plunky@netbsd.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Omit __insn_barrier where implied or subsumed by volatile.

Move Linux SRCU to common.

drm: Eliminate __HAVE_ATOMIC_AS_MEMBAR conditionals.

Discussed on tech-kern:
https://mail-index.netbsd.org/tech-kern/2023/02/23/msg028729.html

branches: 1.11.4;
linux: Convert various API shims to use membar_release/acquire.

linux: Fix copypasto in comment.

drm: Sprinkle miscellaneous includes.

linux: Use indirection for percpu simpleq.

Simpleq requires pointers into the head to work, but percpu(9) may
move objects around, so we have to allocate the simpleq head
separately with kmem(9).

Make intel_psr.c build
netbsd-ify
Stub out intel_lpe_audio.c (for now?).
Hammer intel_opregion.c into shape.
Hack up intel_lrc.c
Adapt intel_reset.c
Stub out intel_region_lmem.c
Adapt intel_ring_submission.c
Adapt intel_rps.c
Adapt intel_sdvo.c
Adapt intel_runtime_pm.c
Adapt intel_sideband.c
Adapt intel_sprite.c
Adapt intel_tv.c
Adapt intel_tc.c
Adapt intel_sseu.c
Adapt intel_timeline.c
Adapt intel_uc_fw.c and fix typo
Fix up intel_uncore.c
intel_vga.c
intel_wakeref.c
intel_vdsc.c
intel_wopcm.c
intel_workarounds.c
i915_globals.c
Fix up intelfb.c
Stub out intel_guc_log.c
Stub out intel_dsi_dcs_backlight.c
i915_perf.c
intel_hdmi.c
intel_hdcp.c


Author: Maya Rashish <maya@NetBSD.org>
Committer: Taylor R Campbell <riastradh@NetBSD.org>

Fix membars in tasklet_disable/enable.

Rework linux_tasklet.c a little.

Provide some more hacks for i915's grubby paws.

Omit __insn_barrier where implied or subsumed by volatile.

Add some i915drmkms additions to the tasklet API.

Sprinkle in another assertion for good measure.

Draft Linux tasklet implementation.

linux: Convert various API shims to use membar_release/acquire.

drm: Convert noisy flush_workqueue message to dtrace probe.

linux: Downgrade warning about flushing from workqueue to debug.

linux: Make flush_workqueue a noop if run from the workqueue.

I think this matches Linux semantics but this is very much not clear
and not written down anywhere obvious and why is this such a
semantically incoherent mess.

linux: Fix flush_workqueue.

Simplify mechanism: we know there's only a single thread here that
processes scheduled work in FIFO order (no multi-CPU workqueues --
would have to adapt if there were), so just schedule a work item that
notifies of completion. The previous mechanism of counting 0, 1, or
2 generation numbers was broken by an earlier change to avoid abuse
of tailqs arising from clever TAILQ_CONCAT.

linux: Update workqueue flush to reflect TAILQ_CONCAT change.

linux/workqueue: Draft queue_rcu_work.

linux/workqueue: Use atomic_load/store.

i915: another whack at it

alloc_workqueue

Constify work_pending, delayed_work_pending.

drain_workqueue

Add work_pending, delayed_work_pending.

Implement return value of flush_work, flush_delayed_work.

New system_unbound_wq.

Fix copypasta in error branch.

XXX pullup to HEAD

Fix abuse of TAILQ_CONCAT.

Other parts of this expect that the entries will be on wq->wq_queue
or wq->wq_dqueue, so we can't just move a batch of entries onto a
temporary queue. Instead, use a marker node to delimit when the
batch ends.

XXX pullup

branches: 1.44.4; 1.44.6;
- dwc2 need calling linux_workqueue_init() to avoid panic.
- use INIT_ONCE/FINI_ONCE to linux_workqueue_{init,fini}() for being called from dwc2.

TODO: dwc2 should be written as kernel module depenging on a linux module.

Workqueue lock must be at IPL_VM for use in interrupts, duh.

Don't skimp on the last batch of work when worker is dying.

This is capitalism. We can't afford to waste time the worker could
spend on labour even if we work 'em to death!

Sprinkle dtrace probes all over the Linux workqueue stuff.

Fix post-cancellation case of queue_delayed_work.

Request the callout queue immediately, as if it hadn't been cancelled
at all, rather than reschedule for the next tick.

Rework Linux workqueue synchronization yet again.

- Use a low bit in the pointer to the queue, rather than whether the
pointer is null or not, to determine whether the work item is
queued/scheduled.

- _Preserve_ the pointer to the queue after we release the work.

- Release the work _before_ executing the function, not after.

This simplifies some things: we no longer have to distinguish whether
the work is queued or running in the functions where we are trying to
modify it. The pointer has to be preserved even after the work is
released so that we can flush the workqueue after the work has been
released.

Fix flush_delayed_work semantics: run it now if it was scheduled.

Reorder fields to get better space savings than bit fields.

Document workqueue semantics as I understand it.

Avoid the prospect of callout calls piling up.

Don't ever callout_schedule the callout while an existing call may be
in progress.

Echo some cases from cancel_delayed_work in flush_delayed_work.

Don't let more than one thread claim credit for cancellation.

Factor out waiting for current work.

Merge two similar branches.

Deduplicate callout destruction.

Deduplicate callout initialization.

Handle a bazillion additional cases I hadn't considered.

There's a lot of duplication here. To be deduplicated in a
subsequent commit.

Fix semantics of flush_work and flush_delayed_work.

- Change return type to void.
=> Upstream it is bool, but exactly one of hundreds of callers
actually use it, and I don't think the semantics is clear.

- Make sure to wait for whichever of the current work _and_ the next
batch queued is currently there in the workqueue.

- Don't retry a cancelled callout. Cancellation in the state
DELAYED_WORK_CANCELLED is guaranteed.

Reverse the sense of some branches for simplicity.

Fix bugs in workqueue destruction.

Fully initialize workqueue struct. Zalloc for paranoia.

No need to retry in cancel_delayed_work_sync after all.

Deduplicate post-cancellation logic.

Don't touch a work after it is released.

Non-waiting cancellation may need to release work too.

After cancel and callout_halt, may need to reload workqueue.

Don't wait forever to flush if the worker thread has no work.

Sprinkle some assertions.

Exercise greater care in acquiring and releasing work.

Add missing break.

Add KASSERT about state after callout fires.

Work is not necessarily queued in flush_work either.

Regularize style: use initializer only if the wq is guaranteed, and
confirmed by KASSERT, to be set

Work is not necessarily queued in cancel_work.

Draft rewrite of Linux workqueue reimplementation.

Just use an explicit thread; don't attempt to fudge it with
workqueue(9). No doubt there are various mistakes in here, but they
should be easier to get right than the mega-kludgerific nonsense that
preceded this draft.

Make flush_delayed_work actually run it. Oops.

XXX pullup-7
XXX pullup-8

Disable broken workqueue_whoami garbage.

Omit vestigial cruft.

Try to avoid uninitialized garbage that gcc helpfully ignores.

Try not to ask rbtree to overwrite random fields with garbage.

Add system_power_efficient_wq.

Implement flush_delayed_work.

Also Linux wants work->func to work.

Add mega-kludgerific current_work().

Add system_long_wq.

branches: 1.1.2; 1.1.6; 1.1.10; 1.1.22; 1.1.24; 1.1.26;
Move linux_work.c and workqueue.h from sys/external/bsd/drm2 to
sys/external/common so that they can be used by others.

LGTM from riastradh@

These should really live outside sys/external, but that can be decided
later