On Mon, 7 Oct 2024 at 10:25, Akihiko Odaki <akihiko.od...@daynix.com> wrote:
> > > Theoretically, It should be safe to assume the callee creates > > autoreleasepool by themselves as needed in general. We have bunch of > > code to call Objective-C APIs without creating autoreleasepool in the > > caller. Practically, [PGDeviceDescriptor new] is likely to be > > implemented with ARC, which wraps methods in autoreleasepool as > > necessary. > > > > > > As far as I'm aware, ARC does NOT automatically insert autorelease pool > > blocks. The reason you rarely need to create autoreleasepool blocks in > > "plain" Objective-C programming is that Cocoa/CFRunloop/libdispatch > > event handlers run each event in an autoreleasepool. So you don't need > > to create them explicitly when using dispatch_async and similar, or when > > running code on the main thread (which runs inside NSApplicationMain/ > > CFRunloopRun/dispatch_main). > > My statement regarding ARC was wrong; It seems ARC just retains a value > annoted as autoreleased. > Yes, the runtime maintains a sort of shadow stack of pointers to autoreleased objects for each thread. When you enter a pool block, it remembers the position in the shadow stack. On exiting the pool, all objects added since entering the pool are popped from the shadow stack and released. Hence, if objects are added to the stack outside of ANY autorelease pool, those are never popped and released. > > > > As far as I'm aware, if you don't explicitly define autoreleasepools in > > raw threads created with the pthreads API, any autoreleased objects will > > leak. At least I've not found any specification/documentation > > contradicting this. And most code in Qemu runs on such raw threads, so > > we need to play it safe with regard to autorelease semantics. > > I decided to dig deeper and found this documentation: > > https://developer.apple.com/library/archive/documentation/Cocoa/Conceptual/MemoryMgmt/Articles/mmAutoreleasePools.html > > It says: > > Cocoa always expects code to be executed within an autorelease pool > > block, otherwise autoreleased objects do not get released and your > > application leaks memory. > > So yes, we must wrap everything in @autoreleasepool at least for Cocoa. > It is probably a good practice to wrap all Objective-C code in > @autoreleasepool. > Yes; any method or function returning an Objective-C object pointer using autorelease semantics - either annotated or by naming convention - will use the pool. But unless it's all your own code all the way down, you don't know if anything inside a method whose own return value is not autoreleased might use autorelease semantics. (The naming convention says most methods return autoreleased, except those named create*, new*, or similar.) > Whether the existing Qemu Objective-C code is safe in this regard I > > don't know for certain, but I've certainly paid attention to this aspect > > when modifying ui/cocoa.m in the past, and indeed most of that code runs > > on the main thread. Note also how I wrap the apple_gfx_render_new_frame > > call in a pool when it can't be guaranteed it's running on a dispatch > > queue because the command buffer inside that uses autorelease semantics. > > It is more about event loop rather than thread. Resources allocated > before [NSApp run] will leak even if they are autoreleased in the main > thread. apple_gfx_common_realize() is one of such functions that run in > the main thread before [NSApp run]. In ui/cocoa, cocoa_display_init() > runs before [NSApp run]. Fortunately we already have NSAutoreleasePool > for this function. > Yes. Unless it's very clear that the code will run only in a Cocoa event handler or dispatch queue context, you need to use an autoreleasepool. > > > > Functions that uses a method that returns autorelease resources > should > > be wrapped with autoreleasepool instead of assuming the caller > creates > > autoreleasepool for them. > > > > > > I'm treating apple_gfx_common_realize as an internal API, and I don't > > think expecting its callers to wrap it in an autoreleasepool block is > > unreasonable. I can certainly explicitly document this in a comment. > > We don't have a comment for cocoa_display_init() and it's more about > generic macOS programming so it's not necessary. > > > > > > diff --git a/hw/display/apple-gfx.m > b/hw/display/apple-gfx.m > > > > new file mode 100644 > > > > index 00000000000..837300f9cd4 > > > > --- /dev/null > > > > +++ b/hw/display/apple-gfx.m > > > > @@ -0,0 +1,536 @@ > > > > +/* > > > > + * QEMU Apple ParavirtualizedGraphics.framework device > > > > + * > > > > + * Copyright © 2023 Amazon.com, Inc. or its affiliates. > All > > > Rights Reserved. > > > > + * > > > > + * This work is licensed under the terms of the GNU GPL, > > version > > > 2 or later. > > > > + * See the COPYING file in the top-level directory. > > > > + * > > > > + * ParavirtualizedGraphics.framework is a set of > > libraries that > > > macOS provides > > > > + * which implements 3d graphics passthrough to the host > > as well as a > > > > + * proprietary guest communication channel to drive it. > This > > > device model > > > > + * implements support to drive that library from within > QEMU. > > > > + */ > > > > + > > > > +#include "apple-gfx.h" > > > > +#include "trace.h" > > > > +#include "qemu/main-loop.h" > > > > +#include "ui/console.h" > > > > +#include "monitor/monitor.h" > > > > +#include "qapi/error.h" > > > > +#include "migration/blocker.h" > > > > +#include <mach/mach_vm.h> > > > > +#import > <ParavirtualizedGraphics/ParavirtualizedGraphics.h> > > > > + > > > > +static const PGDisplayCoord_t apple_gfx_modes[] = { > > > > + { .x = 1440, .y = 1080 }, > > > > + { .x = 1280, .y = 1024 }, > > > > +}; > > > > + > > > > +typedef struct PGTask_s { // Name matches forward > > declaration in > > > PG header > > > > > > Let's name it AppleGFXTask. It is a common practice to have > > the same > > > tag > > > name and typedef in QEMU. > > > > > > > > > This is defining a forward-declared type from framework headers > > which is > > > opaque from the framework's point of view. We do not get to > > choose its > > > struct name. The alternative is having casts wherever these > > objects are > > > being passed between our code and the framework. (See the > > original v1/v2 > > > vmapple patch series for how messy this is.) > > > > I got the idea. Let's not avoid the typedef then to clarify the > naming > > is not under our control. > > > > > > I'm not sure what you mean by this double negative. Are you saying, > > don't add our own typedef for struct PGTask_s at all, just use the > > framework-supplied PGTask_t where appropriate? > > Yes. > > > > > > > > > > > > +static void > > apple_gfx_render_frame_completed(AppleGFXState *s, > > > void *vram, > > > > + > > id<MTLTexture> texture) > > > > +{ > > > > + --s->pending_frames; > > > > + assert(s->pending_frames >= 0); > > > > + > > > > + if (vram != s->vram) { > > > > + /* Display mode has changed, drop this old frame. > */ > > > > + assert(texture != s->texture); > > > > + g_free(vram); > > > > > > This management of buffers looks a bit convoluted. I suggest > > > remembering > > > the width and height instead of pointers and comparing them. > > This way > > > you can free resources in set_mode(). > > > > > > > > > Yeah, I suppose that works, I can change that around. > > > > > > > + } else { > > > > + copy_mtl_texture_to_surface_mem(texture, vram); > > > > > > Writing vram outside BQL may result in tearing. > > > > > > > > > As far as I can tell(*), QXL does the same. I couldn't find any > > examples > > > of double-buffering in any of the existing display devices, which > > would > > > be the only way to do async updates efficiently and without > > tearing. In > > > any case, this solution is still vastly better than a regular VGA > > > device, which suffers from very visible tearing with macOS on the > > guest > > > side anyway. And in an ideal world, we'd pass through the rendered > > > texture directly to the Cocoa UI code rather than copying out > > only for > > > the CPU to draw it back into a window surface which is then > > passed to > > > the GPU for host side rendering. But I felt this patch is already > > very, > > > very large, and if anyone cares, we can fix imperfections in > > subsequent > > > updates. > > > > > > (*)The rendering code in that device is also fairly complex, so I > > may be > > > misreading it. > > > > QXL always modifies the surface with BQL. The surface is modified > with > > qxl_blit(), which is called by qxl_render_update_area_unlocked(). > > qxl_render_update_area_unlocked() is called by either of > > qxl_render_update() and qxl_render_update_area_bh(). Both of them are > > called with BQL. The name includes "unlocked", but it means it is > > called > > without holding QXL-internal lock. > > > > Most devices works entirely with BQL so they don't perform double > > buffering. apple-gfx can do the same. > > > > > > I think we can safely move apple-gfx's framebuffer state management back > > inside the BQL, yes. I just figured that copying dozens of megabytes of > > framebuffer data on every frame while holding the lock was not going to > > help BQL contention. Especially as PVG does not have a concept of dirty > > areas, so we must copy the whole framebuffer every time. (Unless we were > > to implement dirty area detection ourselves.) > > > > Unfortunately, implementing double-buffering would require a major > > rework of Qemu's whole surface management, console code, and probably > > most of the UI implementations. I'm guessing the OpenGL fast-path > > sidesteps all of this, so replicating that with Metal would probably be > > the easier way forward. (Although doing all this graphics stuff inside > > the BQL generally seems like a major architectural flaw; I suppose most > > enterprise use of Qemu does not involve the framebuffer, so it's not > > shown up in BQL contention profiling there. It certainly does in desktop > > use, although at least on macOS hosts there are far worse culprits in > > that regard.) > > We don't need double-buffering; instead we can have one thread that > renders the UI and also > [PGDisplay -encodeCurrentFrameToCommandBuffer:texture:region:]. > Rendering the UI in machine event loop in one thread is a bad idea in > the first place so it will make sense to extract all UI work into a > separate thread. This still requires a lot of work because the UI code > assumes BQL everywhere. > I'm not reworking the UI layer in this patch set, so it sounds like the best we can do is to acquire the BQL for updating the console framebuffer and just take the lock contention hit. > In my understanding BQL is less problematic for KVM users because QEMU > can (sometimes) avoid locking BQL in the vCPU threads. > Yeah, the current hvf implementation is unfortunately still very BQL heavy. Part of that is down to the interrupt controller: with the KVM accel, you typically use the interrupt controller provided by the kernel. With the hvf accel, you get a lot of VM exits for servicing interrupt controller I/O; at least on x86, the APIC implementation only works inside the BQL. HVF on x86-64 does technically have an in-kernel APIC implementation; unfortunately, it has some severe bugs which prevent using it in the general case. (EOI signalling to the IOAPIC does not work/FB14425590; the behaviour for coalescing and retriggering of level-triggered interrupts is also flawed/FB14425412.) On the aarch64 side, an in-kernel GIC implementation was only just added to hvf with macOS 15. I have submitted a patch that at least adds x2apic support to the hvf accel and thus significantly reduces BQL pressure a few times, but the maintainers of the x86 hvf backend don't seem to be active at the moment. > > > > > > > > > + if (s->gfx_update_requested) { > > > > + s->gfx_update_requested = false; > > > > + dpy_gfx_update_full(s->con); > > > > + graphic_hw_update_done(s->con); > > > > + s->new_frame_ready = false; > > > > > > This assignment is unnecessary as s->new_frame_ready is > > always false > > > when s->gfx_update_requested. If you want to make sure > > > s->gfx_update_requested and s->new_frame_ready are mutually > > exclusive, > > > use one enum value instead of having two bools. > > > > > > > > > I'll need to refresh my memory and get back to you on this one, > it's > > > been so many months since I actively worked on this code. > > > > > > > + } else { > > > > + s->new_frame_ready = true; > > > > + } > > > > + } > > > > + if (s->pending_frames > 0) { > > > > + apple_gfx_render_new_frame(s); > > > > + } > > > > +} > > > > + > > > > +static void apple_gfx_fb_update_display(void *opaque) > > > > +{ > > > > + AppleGFXState *s = opaque; > > > > + > > > > + dispatch_async(s->render_queue, ^{ > > > > + if (s->pending_frames > 0) { > > > > > > It should check for s->new_frame_ready as > > > apple_gfx_render_frame_completed() doesn't check if > > > s->pending_frames > 0 before calling > > graphic_hw_update_done(), which is > > > inconsistent. > > > > > > > > > pending_frames is about guest-side frames that are queued to be > > rendered > > > by the host GPU. > > > new_frame_ready being true indicates that the contents of the Qemu > > > console surface has been updated with new frame data since the > last > > > gfx_update. > > > gfx_update_requested indicates that gfx_update is currently > > awaiting an > > > async completion (graphic_hw_update_done) but the surface has not > > > received a new frame content, but the GPU is stily busy drawing > one. > > > > > > apple_gfx_render_frame_completed is scheduled exactly once per > > pending > > > frame, so pending_frames > 0 is an invariant there. (Hence the > > assert.)> > I don't think there is any inconsistency here, but I'll > > double check. > > > It's possible that there's an easier way to express the state > > machine, > > > and I'll take a look at that. > > > > I meant that apple_gfx_render_frame_completed() does not check if the > > frame is the last one currently pending. > apple_gfx_fb_update_display() > > ignores a new ready frame when there is a more pending frame, but > > apple_gfx_render_frame_completed() unconditionally fires > > graphic_hw_update_done() even if there is a more pending frame. And I > > think apple_gfx_render_frame_completed() is right and > > apple_gfx_fb_update_display() is wrong in such a situation. > > > > > > OK, got it. And yes, I agree. > > > > > > > > Checking if s->pending_frames > 0 both in > > apple_gfx_fb_update_display() > > > and apple_gfx_render_frame_completed() is also problematic as > > that can > > > defer graphic_hw_update_done() indefinitely if we are getting > new > > > frames > > > too fast. > > > > > > > > > I see what you mean about this part. I'll have to test it, but I > > guess > > > we should reverse the priority, like this: > > > > > > if (s->new_frame_ready) { > > > dpy_gfx_update_full(s->con); > > > s->new_frame_ready = false; > > > graphic_hw_update_done(s->con); > > > } else if (s->pending_frames > 0) { > > > s->gfx_update_requested = true; > > > } else { > > > graphic_hw_update_done(s->con); > > > } > > > > > > 1. If we already have a frame, ready to be displayed, return it > > immediately. > > > 2. If the guest has reported that it's completed a frame and the > > GPU is > > > currently busy rendering it, defer graphic_hw_update_done until > > it's done. > > > 3. If the guest reports no changes to its display, indicate this > > back to > > > Qemu as a no-op display update graphic_hw_update_done() with no > > > dpy_gfx_update* call. > > > > Yes, that looks correct. > > > > > > > > > + s->gfx_update_requested = true; > > > > + } else { > > > > + if (s->new_frame_ready) { > > > > + dpy_gfx_update_full(s->con); > > > > + s->new_frame_ready = false; > > > > + } > > > > + graphic_hw_update_done(s->con); > > > > + }> + }); > > > > +} > > > > + > > > > +static const GraphicHwOps apple_gfx_fb_ops = { > > > > + .gfx_update = apple_gfx_fb_update_display, > > > > + .gfx_update_async = true, > > > > +}; > > > > + > > > > +static void update_cursor(AppleGFXState *s) > > > > +{ > > > > + dpy_mouse_set(s->con, s->pgdisp.cursorPosition.x, > > > > + s->pgdisp.cursorPosition.y, s- > > >cursor_show); > > > > +} > > > > + > > > > +static void set_mode(AppleGFXState *s, uint32_t width, > > uint32_t > > > height) > > > > +{ > > > > + void *vram = NULL; > > > > + DisplaySurface *surface; > > > > + MTLTextureDescriptor *textureDescriptor; > > > > + id<MTLTexture> texture = nil; > > > > + __block bool no_change = false; > > > > + > > > > + dispatch_sync(s->render_queue, > > > > > > Calling dispatch_sync() while holding BQL may result in > deadlock. > > > > > > Only if any code executed on the same dispatch queue acquires the > > lock > > > either directly or transitively. I believe I have ensure this is > not > > > done on the reqnder_queue, have you found anywhere this is the > case? > > > > The documentation is not clear what threads a dispatch queue runs > > on. We > > can have a deadlock if they lock the BQL. > > > > > > dispatch_sync is a synchronisation primitive (it waits for and asserts > > exclusive access to the given queue), it doesn't actually do any thread > > scheduling. Work scheduled asynchronously to non-main dispatch queues > > will otherwise run on libdispatch pool threads. Running blocks on > > dispatch queues will not preempt and schedule it on other threads which > > may or may not be holding some locks. > > What if all pool threads are waiting for BQL? > The pool is elastic, so if threads are blocked (in syscalls) and there are queued tasks ready to run, the runtime will create more threads as necessary. > > > > So the only way this code will deadlock is if any code scheduled to > > render_queue directly or transitively acquires the BQL. None of it does, > > although updating the console while holding the BQL rather complicates > this. > > > > > > > > > + ^{ > > > > + if (s->surface && > > > > + width == surface_width(s->surface) && > > > > + height == surface_height(s->surface)) { > > > > + no_change = true; > > > > + } > > > > + }); > > > > + > > > > + if (no_change) { > > > > + return; > > > > + } > > > > + > > > > + vram = g_malloc0(width * height * 4); > > > > + surface = qemu_create_displaysurface_from(width, > height, > > > PIXMAN_LE_a8r8g8b8, > > > > + width * 4, > > vram); > > > > + > > > > + @autoreleasepool { > > > > + textureDescriptor = > > > > + [MTLTextureDescriptor > > > > + > > > texture2DDescriptorWithPixelFormat:MTLPixelFormatBGRA8Unorm > > > > + width:width > > > > + height:height > > > > + mipmapped:NO]; > > > > + textureDescriptor.usage = s- > > >pgdisp.minimumTextureUsage; > > > > + texture = [s->mtl > > > newTextureWithDescriptor:textureDescriptor]; > > > > + } > > > > + > > > > + s->using_managed_texture_storage = > > > > + (texture.storageMode == MTLStorageModeManaged); > > > > + > > > > + dispatch_sync(s->render_queue, > > > > + ^{ > > > > + id<MTLTexture> old_texture = nil; > > > > + void *old_vram = s->vram; > > > > + s->vram = vram; > > > > + s->surface = surface; > > > > + > > > > + dpy_gfx_replace_surface(s->con, surface); > > > > + > > > > + old_texture = s->texture; > > > > + s->texture = texture; > > > > + [old_texture release]; > > > > > > You can just do: > > > [s->texture release]; > > > s->texture = texture; > > > > > > This will make s->texture dangling between the two > > statements, but that > > > don't matter since s->texture is not an atomic variable that > > can be > > > safely observed from another thread anyway. > > > > > > > + > > > > + if (s->pending_frames == 0) { > > > > + g_free(old_vram); > > > > + } > > > > + }); > > > > +} > > > > + > > > > +static void create_fb(AppleGFXState *s) > > > > +{ > > > > + s->con = graphic_console_init(NULL, 0, > > &apple_gfx_fb_ops, s); > > > > + set_mode(s, 1440, 1080); > > > > + > > > > + s->cursor_show = true; > > > > +} > > > > + > > > > +static size_t apple_gfx_get_default_mmio_range_size(void) > > > > +{ > > > > + size_t mmio_range_size; > > > > + @autoreleasepool { > > > > + PGDeviceDescriptor *desc = [PGDeviceDescriptor > new]; > > > > + mmio_range_size = desc.mmioLength; > > > > + [desc release]; > > > > + } > > > > + return mmio_range_size; > > > > +} > > > > + > > > > +void apple_gfx_common_init(Object *obj, AppleGFXState *s, > > const > > > char* obj_name) > > > > > > This function can be merged into apple_gfx_common_realize(). > > > > > > > > > Probably. I'll try it. > > > > > > Upon further inspection, we need to call > > cocoa_enable_runloop_on_main_thread() during the init phase, not > > realize(). So we can't get rid of this entirely. Is there any value in > > moving the other init code into _realize()? > > Calling cocoa_enable_runloop_on_main_thread() should be avoided even in > apple_gfx_common_init(). QEMU can plug a device at runtime instead of > initialization time, and in such a case, apple_gfx_common_init() will > run after calling qemu_main. > Good point, it looks like I forgot to mark the mmio variant of apple-gfx devices as non-hotpluggable; I've done that now. I strongly suspect the guest driver doesn't support hotplugging anyway. > I had a closer look and found it has a memory_region_init_io() call, > which should remain in apple_gfx_common_init(). This leads to the same > conclusion that we cannot remove this function so let's only move > migrate_add_blocker() to apple_gfx_common_realize() to report its error. > Sounds good. > > > > > > +{ > > > > + Error *local_err = NULL; > > > > + int r; > > > > + size_t mmio_range_size = > > > apple_gfx_get_default_mmio_range_size(); > > > > + > > > > + trace_apple_gfx_common_init(obj_name, > mmio_range_size); > > > > + memory_region_init_io(&s->iomem_gfx, obj, > > &apple_gfx_ops, s, > > > obj_name, > > > > + mmio_range_size); > > > > + s->iomem_gfx.disable_reentrancy_guard = true; > > > > > > Why do you disable reentrancy guard? > > > > > > > > > Perhaps with the proposed AIO_WAIT_WHILE based I/O scheme, this > > won't be > > > an issue anymore, but the guard would otherwise keep dropping > MMIOs > > > which immediately caused the PV graphics device to stop making > > progress. > > > The MMIO APIs in the PVG framework are thread- and reentrancy- > > safe, so > > > we certainly don't need to serialise them on our side. > > > > It's better to understand why such reentrancy happens. Reentrancy > > itself > > is often a sign of bug. > > > > > > > > > + > > > > + /* TODO: PVG framework supports serialising device > state: > > > integrate it! */ > > > > + if (apple_gfx_mig_blocker == NULL) { > > > > + error_setg(&apple_gfx_mig_blocker, > > > > + "Migration state blocked by apple-gfx > > display > > > device"); > > > > + r = migrate_add_blocker(&apple_gfx_mig_blocker, > > &local_err); > > > > + if (r < 0) { > > > > + error_report_err(local_err); > > > > > > Please report the error to the caller of > > apple_gfx_common_realize() > > > instead. > > > > > > > + } > > > > + } > > > > +}> + > > > > +static void > > > apple_gfx_register_task_mapping_handlers(AppleGFXState *s, > > > > + > > > PGDeviceDescriptor *desc) > > > > +{ > > > > + desc.createTask = ^(uint64_t vmSize, void * _Nullable > * > > > _Nonnull baseAddress) { > > > > + AppleGFXTask *task = apple_gfx_new_task(s, > vmSize); > > > > + *baseAddress = (void*)task->address; > > > > > > nit: please write as (void *) instead of (void*). > > > > > > > + trace_apple_gfx_create_task(vmSize, *baseAddress); > > > > + return task; > > > > + }; > > > > + > > > > > > > +{ > > > > + PGDisplayDescriptor *disp_desc = [PGDisplayDescriptor > > new]; > > > > + > > > > + disp_desc.name <http://disp_desc.name> <http:// > > disp_desc.name <http://disp_desc.name>> = @"QEMU display"; > > > > + disp_desc.sizeInMillimeters = NSMakeSize(400., > > 300.); /* A > > > 20" display */ > > > > + disp_desc.queue = dispatch_get_main_queue(); > > > > + disp_desc.newFrameEventHandler = ^(void) { > > > > + trace_apple_gfx_new_frame(); > > > > + dispatch_async(s->render_queue, ^{ > > > > + /* Drop frames if we get too far ahead. */ > > > > + if (s->pending_frames >= 2) > > > > + return; > > > > + ++s->pending_frames; > > > > + if (s->pending_frames > 1) { > > > > + return; > > > > + } > > > > + @autoreleasepool { > > > > + apple_gfx_render_new_frame(s); > > > > + } > > > > + }); > > > > + }; > > > > + disp_desc.modeChangeHandler = ^(PGDisplayCoord_t > > sizeInPixels, > > > > + OSType pixelFormat) { > > > > + trace_apple_gfx_mode_change(sizeInPixels.x, > > sizeInPixels.y); > > > > + set_mode(s, sizeInPixels.x, sizeInPixels.y); > > > > + }; > > > > + disp_desc.cursorGlyphHandler = ^(NSBitmapImageRep > *glyph, > > > > + PGDisplayCoord_t > > hotSpot) { > > > > + uint32_t bpp = glyph.bitsPerPixel; > > > > + size_t width = glyph.pixelsWide; > > > > + size_t height = glyph.pixelsHigh; > > > > + size_t padding_bytes_per_row = glyph.bytesPerRow > > - width > > > * 4; > > > > + const uint8_t* px_data = glyph.bitmapData; > > > > + > > > > + trace_apple_gfx_cursor_set(bpp, width, height); > > > > + > > > > + if (s->cursor) { > > > > + cursor_unref(s->cursor); > > > > + s->cursor = NULL; > > > > + } > > > > + > > > > + if (bpp == 32) { /* Shouldn't be anything else, > > but just > > > to be safe...*/ > > > > + s->cursor = cursor_alloc(width, height); > > > > + s->cursor->hot_x = hotSpot.x; > > > > + s->cursor->hot_y = hotSpot.y; > > > > + > > > > + uint32_t *dest_px = s->cursor->data; > > > > + > > > > + for (size_t y = 0; y < height; ++y) { > > > > + for (size_t x = 0; x < width; ++x) { > > > > + /* NSBitmapImageRep's red & blue > channels > > > are swapped > > > > + * compared to QEMUCursor's. */ > > > > + *dest_px = > > > > + (px_data[0] << 16u) | > > > > + (px_data[1] << 8u) | > > > > + (px_data[2] << 0u) | > > > > + (px_data[3] << 24u); > > > > + ++dest_px; > > > > + px_data += 4; > > > > + } > > > > + px_data += padding_bytes_per_row; > > > > + } > > > > + dpy_cursor_define(s->con, s->cursor); > > > > + update_cursor(s); > > > > + } > > > > + }; > > > > + disp_desc.cursorShowHandler = ^(BOOL show) { > > > > + trace_apple_gfx_cursor_show(show); > > > > + s->cursor_show = show; > > > > + update_cursor(s); > > > > + }; > > > > + disp_desc.cursorMoveHandler = ^(void) { > > > > + trace_apple_gfx_cursor_move(); > > > > + update_cursor(s); > > > > + }; > > > > + > > > > + return disp_desc; > > > > +} > > > > + > > > > +static NSArray<PGDisplayMode*>* > > > apple_gfx_prepare_display_mode_array(void) > > > > +{ > > > > + PGDisplayMode *modes[ARRAY_SIZE(apple_gfx_modes)]; > > > > + NSArray<PGDisplayMode*>* mode_array = nil; > > > > + int i; > > > > + > > > > + for (i = 0; i < ARRAY_SIZE(apple_gfx_modes); i++) { > > > > + modes[i] = > > > > + [[PGDisplayMode alloc] > > > initWithSizeInPixels:apple_gfx_modes[i] refreshRateInHz:60.]; > > > > + } > > > > + > > > > + mode_array = [NSArray arrayWithObjects:modes > > > count:ARRAY_SIZE(apple_gfx_modes)]; > > > > + > > > > + for (i = 0; i < ARRAY_SIZE(apple_gfx_modes); i++) { > > > > + [modes[i] release]; > > > > + modes[i] = nil; > > > > + } > > > > + > > > > + return mode_array; > > > > +} > > > > + > > > > +static id<MTLDevice> copy_suitable_metal_device(void) > > > > +{ > > > > + id<MTLDevice> dev = nil; > > > > + NSArray<id<MTLDevice>> *devs = MTLCopyAllDevices(); > > > > + > > > > + /* Prefer a unified memory GPU. Failing that, pick a > non- > > > removable GPU. */ > > > > + for (size_t i = 0; i < devs.count; ++i) { > > > > + if (devs[i].hasUnifiedMemory) { > > > > + dev = devs[i]; > > > > + break; > > > > + } > > > > + if (!devs[i].removable) { > > > > + dev = devs[i]; > > > > + } > > > > + } > > > > + > > > > + if (dev != nil) { > > > > + [dev retain]; > > > > + } else { > > > > + dev = MTLCreateSystemDefaultDevice(); > > > > + } > > > > + [devs release]; > > > > + > > > > + return dev; > > > > +} > > > > + > > > > +void apple_gfx_common_realize(AppleGFXState *s, > > > PGDeviceDescriptor *desc) > > > > +{ > > > > + PGDisplayDescriptor *disp_desc = nil; > > > > + > > > > + QTAILQ_INIT(&s->tasks); > > > > + s->render_queue = dispatch_queue_create("apple- > > gfx.render", > > > > + > > DISPATCH_QUEUE_SERIAL); > > > > + s->mtl = copy_suitable_metal_device(); > > > > + s->mtl_queue = [s->mtl newCommandQueue]; > > > > + > > > > + desc.device = s->mtl; > > > > + > > > > + apple_gfx_register_task_mapping_handlers(s, desc); > > > > + > > > > + s->pgdev = PGNewDeviceWithDescriptor(desc); > > > > + > > > > + disp_desc = apple_gfx_prepare_display_handlers(s); > > > > + s->pgdisp = [s->pgdev > newDisplayWithDescriptor:disp_desc > > > > + port:0 > > > serialNum:1234]; > > > > + [disp_desc release]; > > > > + s->pgdisp.modeList = > > apple_gfx_prepare_display_mode_array(); > > > > + > > > > + create_fb(s); > > > > +} > > > > diff --git a/hw/display/meson.build > b/hw/display/meson.build > > > > index 7db05eace97..70d855749c0 100644 > > > > --- a/hw/display/meson.build > > > > +++ b/hw/display/meson.build > > > > @@ -65,6 +65,8 @@ system_ss.add(when: 'CONFIG_ARTIST', > > if_true: > > > files('artist.c')) > > > > > > > > system_ss.add(when: 'CONFIG_ATI_VGA', if_true: > > [files('ati.c', > > > 'ati_2d.c', 'ati_dbg.c'), pixman]) > > > > > > > > +system_ss.add(when: 'CONFIG_MAC_PVG', if_true: > > > [files('apple-gfx.m'), pvg, metal]) > > > > +system_ss.add(when: 'CONFIG_MAC_PVG_VMAPPLE', if_true: > > > [files('apple-gfx-vmapple.m'), pvg, metal]) > > > > > > > > if config_all_devices.has_key('CONFIG_VIRTIO_GPU') > > > > virtio_gpu_ss = ss.source_set() > > > > diff --git a/hw/display/trace-events > b/hw/display/trace-events > > > > index 781f8a33203..1809a358e36 100644 > > > > --- a/hw/display/trace-events > > > > +++ b/hw/display/trace-events > > > > @@ -191,3 +191,29 @@ dm163_bits_ppi(unsigned dest_width) > > > "dest_width : %u" > > > > dm163_leds(int led, uint32_t value) "led %d: 0x%x" > > > > dm163_channels(int channel, uint8_t value) "channel %d: > > 0x%x" > > > > dm163_refresh_rate(uint32_t rr) "refresh rate %d" > > > > + > > > > +# apple-gfx.m > > > > +apple_gfx_read(uint64_t offset, uint64_t res) > > > "offset=0x%"PRIx64" res=0x%"PRIx64 > > > > +apple_gfx_write(uint64_t offset, uint64_t val) > > > "offset=0x%"PRIx64" val=0x%"PRIx64 > > > > +apple_gfx_create_task(uint32_t vm_size, void *va) > > "vm_size=0x%x > > > base_addr=%p" > > > > +apple_gfx_destroy_task(void *task) "task=%p" > > > > +apple_gfx_map_memory(void *task, uint32_t range_count, > > uint64_t > > > virtual_offset, uint32_t read_only) "task=%p range_count=0x%x > > > virtual_offset=0x%"PRIx64" read_only=%d" > > > > +apple_gfx_map_memory_range(uint32_t i, uint64_t phys_addr, > > > uint64_t phys_len) "[%d] phys_addr=0x%"PRIx64" > > phys_len=0x%"PRIx64 > > > > +apple_gfx_remap(uint64_t retval, uint64_t source, uint64_t > > > target) "retval=%"PRId64" source=0x%"PRIx64" target=0x%"PRIx64 > > > > +apple_gfx_unmap_memory(void *task, uint64_t > virtual_offset, > > > uint64_t length) "task=%p virtual_offset=0x%"PRIx64" > > length=0x%"PRIx64 > > > > +apple_gfx_read_memory(uint64_t phys_address, uint64_t > length, > > > void *dst) "phys_addr=0x%"PRIx64" length=0x%"PRIx64" dest=%p" > > > > +apple_gfx_raise_irq(uint32_t vector) "vector=0x%x" > > > > +apple_gfx_new_frame(void) "" > > > > +apple_gfx_mode_change(uint64_t x, uint64_t y) "x=%"PRId64" > > > y=%"PRId64 > > > > +apple_gfx_cursor_set(uint32_t bpp, uint64_t width, > uint64_t > > > height) "bpp=%d width=%"PRId64" height=0x%"PRId64 > > > > +apple_gfx_cursor_show(uint32_t show) "show=%d" > > > > +apple_gfx_cursor_move(void) "" > > > > +apple_gfx_common_init(const char *device_name, size_t > > mmio_size) > > > "device: %s; MMIO size: %zu bytes" > > > > + > > > > +# apple-gfx-vmapple.m > > > > +apple_iosfc_read(uint64_t offset, uint64_t res) > > > "offset=0x%"PRIx64" res=0x%"PRIx64 > > > > +apple_iosfc_write(uint64_t offset, uint64_t val) > > > "offset=0x%"PRIx64" val=0x%"PRIx64 > > > > +apple_iosfc_map_memory(uint64_t phys, uint64_t len, > > uint32_t ro, > > > void *va, void *e, void *f) "phys=0x%"PRIx64" len=0x%"PRIx64" > > ro=%d > > > va=%p e=%p f=%p" > > > > +apple_iosfc_unmap_memory(void *a, void *b, void *c, void > *d, > > > void *e, void *f) "a=%p b=%p c=%p d=%p e=%p f=%p" > > > > +apple_iosfc_raise_irq(uint32_t vector) "vector=0x%x" > > > > + > > > > diff --git a/meson.build b/meson.build > > > > index 10464466ff3..f09df3f09d5 100644 > > > > --- a/meson.build > > > > +++ b/meson.build > > > > @@ -741,6 +741,8 @@ socket = [] > > > > version_res = [] > > > > coref = [] > > > > iokit = [] > > > > +pvg = [] > > > > +metal = [] > > > > emulator_link_args = [] > > > > midl = not_found > > > > widl = not_found > > > > @@ -762,6 +764,8 @@ elif host_os == 'darwin' > > > > coref = dependency('appleframeworks', modules: > > 'CoreFoundation') > > > > iokit = dependency('appleframeworks', modules: 'IOKit', > > > required: false) > > > > host_dsosuf = '.dylib' > > > > + pvg = dependency('appleframeworks', modules: > > > 'ParavirtualizedGraphics') > > > > + metal = dependency('appleframeworks', modules: 'Metal') > > > > elif host_os == 'sunos' > > > > socket = [cc.find_library('socket'), > > > > cc.find_library('nsl'), > > > > > > >
