Allow nvkm_client_map to map subregions.

Linux ioremap does not care about overlapping mapped regions, but
bus_space_map does.  Since nouveau ioremaps the entire register space
of BAR 0, and separately some subregions of it, we need to convey the
bus addresses in question to it in order to bus_space_subregion.

Kinda kludgey, but we don't care about running this in userland or
anything.

While here: initialize object->map.tag.  Not physically necessary on
x86 as long as the x86_io_mem_tag is nonull, so this won't fix any
symptoms on x86, but it is wrong to leave it null.

(riastradh)

Let's not truncate registers to 8-bit when we read them, shall we?

(riastradh)

Use bus_space now that we have the tag and handle.

Fix possible byte order issue while here.

(riastradh)

Work around ioctl ABI mistake in i915 drm.

A flags argument was added to the I915_GEM_MMAP ioctl.  The version
of xf86-video-intel we imported was compiled to use the new ioctl
structure including the flags argument, but without initializing the
flags argument, so it would be userland stack garbage causing
i915_gem_mmap_ioctl to fail.

Newer xf86-video-intel will query the mmap version first before
trying to use either the older ioctl structure, with the original
shorter argument and thus no stack garbage (which will be padded with
to the full length in the kernel by drm_ioctl), or the newer ioctl
structure, with the argument initialized.

libdrm does not appear to be affected by this -- it zeroes the longer
structure before passing it in.

Patch from mrg@, diagnosis and explanation by me.

(riastradh)

Fix order of arguments to kcopy.  @#&!$!*#

Add const qualifiers to src to compiler will detect this.

(riastradh)

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...

(riastradh)

Use kcopy, not memcpy, in case of failure to page in dst.  From chs@.

(riastradh)

Forgot again to restore reference consumed by uvm_map.

Add missing errno comment while here.

(riastradh)

Zero-pad truncated drm ioctl commands on input.

This way we don't act on uninitialized stack garbage if user invokes
a drm ioctl with a short input, which, uhhh, we have been doing for a
long time.  #@^&*$!@&@#*@!

XXX pullup-7
XXX pullup-8

(riastradh)

For reasons beyond me, map does size/offset but unmap does start/end.

(riastradh)

Narrow scope of locks in i915 error checks to reduce deadlocks.

Not really sure why I had put some of these locks here...

(riastradh)

Need irq_lock to wait on irq_queue.  From mrg@.

(riastradh)

Use uvm_map directly and make sure we unmap correct length.

(riastradh)

Fix mapping of non-page-aligned stuff in vmap_batch.

(riastradh)

uvm_map consumes reference on success, so be sure to acquire another.

(riastradh)

Remove wrong comment.

(riastradh)

Caller must also hold vblank_time_lock here.

(riastradh)

Fix scope of locks in drm_handle_vblank.

(riastradh)

Fix intel pipe update locking and whining.

(riastradh)

Rework vblank locking.  New drm_vblank_get/put_locked subroutines.

(riastradh)

KASSERT -> KASSERTMSG

(riastradh)

Fix locking in amd_sched_wakeup and amd_sched_entity_in.

(riastradh)

kfifo is used under a spin lock, so its lock must be a spin lock.

From mrg@.

(riastradh)

Make mutex destruction match initialization.

From mrg@.

(riastradh)

At mrg@'s suggestion, make amdgpu win over radeon.

(riastradh)

Destroy glob->lock if we init it.

(riastradh)

Linux vfree accepts NULL as noop.  Match semantics.

From mrg@.

(riastradh)

KASSERT -> KASSERTMSG, to help debug panic observed by mrg@.

(riastradh)

Call drm_modeset_lock_fini in drm_plane_cleanup.

From mrg@.

(riastradh)

Linux fence_put accepts NULL as noop.  Match this.

From mrg@.

(riastradh)

Call ww_mutex_destroy in drm_modeset_lock_fini.

From mrg@.

(riastradh)

Initialize connector->kdev, which amdgpu uses.

(riastradh)

Truncate workqueue names to fit in the maximum 16 characters.

(riastradh)

Prefix radeon .c files with `radeon_'.

(riastradh)

Make amdgpu cik build.

- Sprinkle __UNCONST and __UNVOLATILE judiciously.
- Ifdef out debugfs stuff and other unused things.
- Add includes.
- Convert void * to char * before arithmetic.
- Add prototype for amdgpu_cik_gpu_check_soft_reset.
  . XXX Should really be in a header file, but whatever...
- Mark unused variables.
- sprintf -> snprintf

(riastradh)

Disable PCIe stuff until I can figure out what it does.

(riastradh)

Just disable -Wshadow and -Wmissing-field-initializers for all amdgpu.

(riastradh)

Options are lowercase in config conditions.

(riastradh)

Update amdgpu2netbsd.  Make idempotent.

(riastradh)

New option AMDGPU_CIK for devices covered by radeon and amdgpu.

(riastradh)

Make extremely noisy fence trace output conditional.

(riastradh)

Notify radeon_fence_default_wait on fence signal.

(riastradh)

Make sure all of struct drm_file gets initialized.

(riastradh)

Treat pre-enabled devices as 1 enablecnt.

(riastradh)

For now, unconditionally set nouveau verbosity to debug.

(riastradh)

Paranoia: let dev be null, in match routines.

(riastradh)

Reduce diff a little.

This code was already #ifndef __NetBSD__; the corresponding logic
lives in nouveau_module.c.

(riastradh)

Clarify nvkm_device_pci_create arguments.  Use bogus parent for match prints.

(riastradh)

Another zero __ffs64 bug.

(riastradh)

Mark it on the tree so we remove it during destruction.

(riastradh)

Don't forget to add the base address.

(riastradh)

Don't pass zero to __ffs64 -- it finds zero indigestible.

(riastradh)

This can happen while cold.  Wait up to 100ms for it.

(riastradh)

Device hangs on pci disable in match; try skipping enable/disable.

(riastradh)

Hook up the nouveau pci driver so it does anything.

(riastradh)

Zero the whole thing for good measure.

(riastradh)

Make sure the resources are marked unmapped initially.

(riastradh)

Move the Linux PCI shim into its very own .c file.

The poor thing has deserved this for a very long time, even though it
is full of kludges.  Also I'm tired of recompiling everything every
time I touch it.

(riastradh)

Destroy spin locks we created.

(riastradh)

Fix sense of test to make linux_pci_dev_destroy work.

(riastradh)

Zero-initialize idr cache.

(riastradh)

Initialize more of struct pci_device.

(riastradh)

Sync drm_close_file: call drm_property_release_user_blobs.

(riastradh)

Release the magic before releasing the master.

Releasing the master frees the idr into which we are trying to
release the magic.  Use of magic that has been freed sounds like a
dangerous proposition.

(riastradh)

Make gc static.  No need to use this private state elsewhere.

(riastradh)

Draft rewrite of idr preload.

Previous idr code assumed every caller would definitely call
idr_preload, idr_alloc, idr_preload_end, but some callers skip
idr_alloc if an intermediate failure happens first, and would
therefore leak idr nodes.

Use a per-lwp single-node cache instead, and print warnings about
leakers.

(riastradh)

Sprinkle idr_preload/idr_preload_end.

Try to do these outside all mutex locks.

Tricky case is i915_gem_context, where the struct mutex is held for a
long time.

(riastradh)

Clarify with comments and assertions.

(riastradh)

There may not be an old list.

Nihil ex nihilo -- don't dereference null.

(riastradh)

amdgpu.kmod depends on drmkms.kmod, not on amdgpu.kmod.

(riastradh)

Lotsa kasserts about referencedness of frences.

(riastradh)

Acquire references in add_shared/excl.

(riastradh)

Call linux_pci_dev_init/destroy in viadrmums(4) too.

(riastradh)

Let upstream code attempt match for nouveau.

(riastradh)

Call linux_pci_dev_destroy on detach.

(riastradh)

Implement linux_pci_dev_destroy.

(riastradh)

Pass the parent explicitly to linux_pci_dev_init.

This will allow us to use it in match routines.

(riastradh)

Move linux_pci_dev_init into callers of drm_pci_attach.

The order in nouveau has to be

- linux_pci_dev_init
- nvkm_device_pci_new
- rest of drm_pci_attach

(riastradh)

Locks are not like negatives: doubles don't cancel out.

(riastradh)

Disable broken workqueue_whoami garbage.

(riastradh)

Update amdgpu2netbsd script.

(riastradh)

Rename amdgpu files to avoid clashes with radeon.

(riastradh)

Don't hork on DEBUG builds.

(riastradh)

Note radeon/amdgpu file collisions TODO in README.

(riastradh)

Tired of fine-grained commits.  Have an amdgpu(4).

(riastradh)

const const const const const

(riastradh)

Convert amdgpu scheduler to drm waitqueues and NetBSD kthreads.

(riastradh)

Don't abuse struct fence internals for irreverent dolphins.

(riastradh)

Delete upstream trace point header files that get in the way.

(riastradh)

Add some amdgpu goo.

(riastradh)

Empty stub <linux/irq.h>.

Not sure why upstream uses it -- the file in Linux exhorts against
its usage in any MI code.

(riastradh)

Draft reservation_object_get_fences_rcu.

(riastradh)

Add fence_is_later and fence_wait_any_timeout.

(riastradh)

Fix prototype of function in comment.

(riastradh)

Rewrite <linux/kfifo.h>.

pcq isn't flexible enough for its needs now.  Just use a spinlocked
circular buffer.

(riastradh)

Add atomic_long.

(riastradh)

Note need_mmap_used is __diagused, not __unused.

(riastradh)

Define nsecs_to_jiffies.

(riastradh)

Need <sys/kernel.h> for boottime.

(riastradh)

Teach drm_sysctl about uint, kludgily.

(riastradh)

Add a copy of __user (empty) to <linux/types.h> for expedience.

(riastradh)

Define list_splice_init.

(riastradh)

Make sure to call pslist_entry_init.

(riastradh)

Fence release callback may be absent; default to fence_free.

(riastradh)

Fix ordering of bus_dmamap_sync.

The ordering is:

...
- driver writes to kernel pointer
- sync
- driver triggers device use of newly written data (*)
- sync
- driver writes to kernel pointer
- sync
...

The operation (*) is a DMA write.  Hence the syncs should be:

...
- driver writes to kernel pointer
- bus_dmamap_sync(BUS_DMASYNC_PREWRITE)
- driver triggers device use of newly written data
- bus_dmamap_sync(BUS_DMASYNC_POSTWRITE)
- bus_dmamap_sync(BUS_DMASYNC_PREWRITE)
- driver writes to kernel pointer
- bus_dmamap_sync(BUS_DMASYNC_POSTWRITE)
- bus_dmamap_sync(BUS_DMASYNC_PREWRITE)
...

(riastradh)

GART might not be mapped with bus_dma.

Still need membar and tlb flush in that case.

(riastradh)

get pd_rom_bst and pd_rom_found_bsh _after_ pci_map_rom

fixes early attach panic.

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

(riastradh)

Apparently one can reserve_shared only once in a row.  Simplify.

(riastradh)

Make drmkms_pci.kmod build again.

(riastradh)

Upstream nixed __OS_HAS_AGP; use CONFIG_AGP instead.

(riastradh)

Reverse backwards namespacing.

(riastradh)

Add missing namespacing.

(riastradh)

Newline-terminate error message.

(riastradh)

Move hweight8 next to its cousins.

(riastradh)

Make i915drmkms.kmod build again.

(riastradh)

Put override directory first in CPPFLAGS.

(riastradh)

Move CPPFLAGS to top, CWARNFLAGS below.

(riastradh)

Omit commented warn flag, evidently needless.

(riastradh)

Make drmkms.kmod build again on amd64.

(riastradh)

Define CONFIG_DRM_FBDEV_EMULATION=1 in generic drmkms code.

This was accidentally pulled in by i915drmkms, but we need it
generically or else we need to patch upstream differently.

(riastradh)

Sort.  Put CWARNFLAGS next to source lists.

(riastradh)

Guard #include "pci.h" with _KERNEL_OPT.

(riastradh)

Add some components to drmkms_linux.kmod.

(riastradh)

Sort.

(riastradh)

Sort.

(riastradh)

Nix extraneous whitespace.

(riastradh)

Implement kvfree by free(9); assume kmalloc/vmalloc use malloc(9).

(riastradh)

Insert membars needed for non-x86.

XXX We should really define membar_enter_postatomic and
membar_exit_preatomic to avoid relying on not typoing
__HAVE_ATOMIC_AS_MEMBAR.

(riastradh)

With <asm/cpufeature.h>, we can now use cpu_has_pat.

(riastradh)

Define cpu_has_pat for x86.

(riastradh)

Make this a proper boolean.  Skip ISSET.  Add some spacing.

(riastradh)

Rename nouveau_devices_init/fini -> nvkm_devices_init/fini.

Matches newer upstream naming conventions better.

(riastradh)

Fold local intel_uncore_destroy into upstream intel_uncore_fini.

(riastradh)

Move hint setting nearer to uvm_map for clarity.  CTASSERT page size.

Elsewhere this code says 4096, rather than PAGE_SIZE.  PAGE_SIZE is
the smallest unit that uvm_map accepts, so it had better match.

(riastradh)

Initialize kva hint for uvm_map to avoid using uninitialized memory.

(riastradh)

Update usage() to include documented flags

(sevan)

Remove temporary diagnostic now that I know I forgot uio_offset.

(riastradh)

Put file, line, func in DRM_ERROR messages.

(riastradh)

Don't forget to initialize uio_offset.

(riastradh)

Turn on driver debug messages for now

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

(riastradh)

Transfer from the correct offset.

(riastradh)

Temporary diagnostic kludge: confirm the ubc_uiomove worked as expected.

(riastradh)

Guess what uvm criterion means we should discard backing pages.

(riastradh)

Stub SYS_RESTART for reboot notifications so we can remove panic("XXX").

(riastradh)

Attempt to match Linux semantics for membars implied by atomics.

This is kind of moot at the moment because we're mostly x86-only for
drmkms, but this might help in the future if we ever went beyond x86.

(riastradh)

Nix memory barriers that were nixed upstream.  Add smp_mb__after_atomic.

(riastradh)

Fix return types of test_and_{set,clear,change}_bit.

(riastradh)

Don't forget to set .uio_offset and to call UIO_SETUP_SYSSPACE.

(riastradh)

correct typo

shows up as null deref in finish_csr_load.

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

(riastradh)

i915 needs firmware now too, so defer to mountroot.

(riastradh)

Don't forget to call kthread_exit.

(riastradh)

Free the bus crap.

(riastradh)

For the kludges we don't have or need a device (parent) pointer.

(riastradh)

Nix vestigial temporary.

(riastradh)

Omit vestigial cruft.

(riastradh)

Try to avoid uninitialized garbage that gcc helpfully ignores.

(riastradh)

Conditional here shouldn't be necessary now.

(riastradh)

Copy va_list before using it twice.

(riastradh)

Can't and needn't xcall while cold.

(riastradh)

Fix vestigial call to drm_fb_helper_set_config.

(riastradh)

Use atomics to manage fence->flags.

Outside users use this field too, and expect it to be managed with
atomics.

(riastradh)

Update comment to reflect reality.

(riastradh)

Remove temporary drm_fb_helper_set_config stub from old code.

(riastradh)

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

(riastradh)

Wait for the read to be valid!

(riastradh)

Stub out intel_dsi_init.

(riastradh)

Fix fence API usage: lock is _not_ held here, so acquire it.

(riastradh)

Draft implementation of the Linux reservation and fence APIs.

(riastradh)

Implement some of the Linux RCU API.

(riastradh)

Add missing -width; remove unnecessary .Pp.

(wiz)

Improve the "Map" section.

(maxv)

Document ALGs.

(maxv)

When the root= boot arg is missing, attempt to autodetect the root device
by comparing a hash of a disk device's MBR with the hash passed in from
the bootloader.

(jmcneill)

Add evbarm_cpu_rootconf callback for hooking into cpu_rootconf

(jmcneill)

Pass a hash of the MBR and the booted partition to the kernel.

(jmcneill)

Add kasan interceptors for strcpy/strcmp/strlen.

(maxv)