[Spice-devel] [PATCH 02/12] qxl-wddm-dod: Use rendering offload thread
Christophe de Dinechin
cdupontd at redhat.com
Tue Mar 21 08:50:28 UTC 2017
> On 20 Mar 2017, at 19:03, Frediano Ziglio <fziglio at redhat.com <mailto:fziglio at redhat.com>> wrote:
>
>
> Side note: the .gitattribute in that project seems all wrong. It makes source files a pain to edit under Linux or macOS. Why not mark the source files as text?
> To avoid cr/lf conversions.
But we want CR/LF conversions for text files. That’s the whole point.
> If you can find good flags to make git Windows and git Unix to work you are welcome.
> Got mad to try, you fix a tool and you break another.
I did that recently on another project. text=auto does work. I guess there was some other problem at the time. You mention multiple tools. Do you remember which one broke?
> I have no problems with Linux, my editor just handle DOS files as DOS.
So does Emacs, but it shows additional ^M within the file if CR/LF is inconsistent throughout the file, as it the case here, presumably because the .gitattributes was added in the middle of the life of the project without first making the files consistent.
Thanks,
Christophe
>
> On 20 Mar 2017, at 13:08, Frediano Ziglio <fziglio at redhat.com <mailto:fziglio at redhat.com>> wrote:
>
>
> Instead of sending drawable commands down from presentation
> callback, collect drawables objects and pass them to dedicated
> thread for processing. This reduce peak load of presentation
> callback.
>
> Signed-off-by: Javier Celaya <javier.celaya at flexvdi.com <mailto:javier.celaya at flexvdi.com>>
> Signed-off-by: Yuri Benditovich <yuri.benditovich at daynix.com <mailto:yuri.benditovich at daynix.com>>
> ---
> qxldod/QxlDod.cpp | 43 +++++++++++++++++++++++++++++++++----------
> qxldod/QxlDod.h | 4 ++--
> 2 files changed, 35 insertions(+), 12 deletions(-)
>
> diff --git a/qxldod/QxlDod.cpp b/qxldod/QxlDod.cpp
> index b952bf9..01de9b3 100755
> --- a/qxldod/QxlDod.cpp
> +++ b/qxldod/QxlDod.cpp
> @@ -3794,11 +3794,20 @@ QxlDevice::ExecutePresentDisplayOnly(
> SIZE_T sizeRects = NumDirtyRects*sizeof(RECT);
> SIZE_T size = sizeof(DoPresentMemory) + sizeMoves + sizeRects;
>
> + QXLDrawable **pDrawables = reinterpret_cast<QXLDrawable **>(new
> (NonPagedPoolNx) BYTE[sizeof(QXLDrawable *)*(NumDirtyRects + NumMoves +
> 1)]);
>
> here would be
>
> QXLDrawable **pDrawables = new (NonPagedPoolNx) QXLDrawable *[NumDirtyRects + NumMoves + 1];
>
> is non paged memory needed? Both functions (producer and consumer) are in paged areas.
>
> + UINT nIndex = 0;
> +
> + if (!pDrawables)
> + {
> + return STATUS_NO_MEMORY;
> + }
> +
> DoPresentMemory* ctx = reinterpret_cast<DoPresentMemory*>
> (new (NonPagedPoolNx) BYTE[size]);
>
>
> DoPresentMemory* ctx = new (NonPagedPoolNx) DoPresentMemory;
>
> That would be nicer, but apparently, there is extra stuff padded to it:
>
> SIZE_T size = sizeof(DoPresentMemory) + sizeMoves + sizeRects;
>
>
> Also, this part of the code was not changed. It was like this before. But is it necessary ? It’s really borderline with respect to alignment. In one case, we have BYTE-aligned memory, in the other it’s at least sizeof(void *).
>
> You could use placement new. Assuming non-paged pool:
>
> BYTE *storage = new(NonPagedPoolNx) BYTE[size];
> DoPresentMemory *ctx = new(storage) DoPresentMemory;
>
> There’s no constructor, apparently. So it does not make much of a difference.
> Looking at the code it looks quite weird, a ctx is allocated in the function, then at the end freed,
> I found these comments:
>
> // Alternate between synch and asynch execution, for demonstrating
> // that a real hardware implementation can do either
> ...
>
> // Save Mdl to unmap and unlock the pages in worker thread
> ...
>
> // copy moves and update pointer
> ...
>
> // copy dirty rects and update pointer
>
>
> apparently this is all due to the initial code (from a Microsoft example) that was marshalling this
> call to a worker thread.
> However now that we are going to introduce a worker thread this looks misleading.
> I would remove comments and code. For instance there's no need to copy Moves & DirtyRect
> and DoPresentMemory can be allocated just in the stack (or even better removed).
>
> same comments as above
>
> if (!ctx)
> {
> + delete[] reinterpret_cast<BYTE*>(pDrawables);
>
> delete[] pDrawables;
>
> return STATUS_NO_MEMORY;
> }
>
> @@ -3828,6 +3837,8 @@ QxlDevice::ExecutePresentDisplayOnly(
> PMDL mdl = IoAllocateMdl((PVOID)SrcAddr, sizeToMap, FALSE, FALSE,
> NULL);
> if(!mdl)
> {
> + delete[] reinterpret_cast<BYTE*>(ctx);
>
> There were several leaks of ctx in case of failure before. I suggest a separate patch to fix that. There are many other places where the current patch does not fix them, for instance VgaDevice::GetModeList can leak m_ModeInfo if m_ModeNumbers can’t be allocated.
> Can you write some patches about these? These specifically seem not really related.
>
> + delete[] reinterpret_cast<BYTE*>(pDrawables);
>
> similar to above, in this case "delete ctx;”
>
> There is a risk if, indeed, we store more than an object in it. delete[] and delete are implemented differently by some runtimes (I don’t recall about the Win driver runtime). It is not guaranteed that delete ctx would work reliably if we allocated more than sizeof(DoPresentMemory) bytes. Being able to deal with variable size is the very reason for delete[].
>
>
>
> return STATUS_INSUFFICIENT_RESOURCES;
> }
>
> @@ -3844,6 +3855,8 @@ QxlDevice::ExecutePresentDisplayOnly(
> {
> Status = GetExceptionCode();
> IoFreeMdl(mdl);
> + delete[] reinterpret_cast<BYTE*>(ctx);
> + delete[] reinterpret_cast<BYTE*>(pDrawables);
>
>
> ditto
>
> return Status;
> }
>
> @@ -3857,6 +3870,8 @@ QxlDevice::ExecutePresentDisplayOnly(
> Status = STATUS_INSUFFICIENT_RESOURCES;
> MmUnlockPages(mdl);
> IoFreeMdl(mdl);
> + delete[] reinterpret_cast<BYTE*>(ctx);
> + delete[] reinterpret_cast<BYTE*>(pDrawables);
>
> ditto
>
>
>
> return Status;
> }
>
> @@ -3922,7 +3937,9 @@ QxlDevice::ExecutePresentDisplayOnly(
> DbgPrint(TRACE_LEVEL_INFORMATION, ("--- %d SourcePoint.x = %ld,
> SourcePoint.y = %ld, DestRect.bottom = %ld, DestRect.left = %ld,
> DestRect.right = %ld, DestRect.top = %ld\n",
> i , pSourcePoint->x, pSourcePoint->y, pDestRect->bottom,
> pDestRect->left, pDestRect->right, pDestRect->top));
>
> - CopyBits(*pDestRect, *pSourcePoint);
> + pDrawables[nIndex] = CopyBits(*pDestRect, *pSourcePoint);
> +
> + if (pDrawables[nIndex]) nIndex++;
> }
>
> // Copy all the dirty rects from source image to video frame buffer.
> @@ -3936,11 +3953,13 @@ QxlDevice::ExecutePresentDisplayOnly(
> DbgPrint(TRACE_LEVEL_INFORMATION, ("--- %d pDirtyRect->bottom = %ld,
> pDirtyRect->left = %ld, pDirtyRect->right = %ld, pDirtyRect->top =
> %ld\n",
> i, pDirtyRect->bottom, pDirtyRect->left, pDirtyRect->right,
> pDirtyRect->top));
>
> - BltBits(&DstBltInfo,
> + pDrawables[nIndex] = BltBits(&DstBltInfo,
> &SrcBltInfo,
> 1,
> pDirtyRect,
> &sourcePoint);
> +
> + if (pDrawables[nIndex]) nIndex++;
> }
>
> // Unmap unmap and unlock the pages.
> @@ -3951,6 +3970,10 @@ QxlDevice::ExecutePresentDisplayOnly(
> }
> delete [] reinterpret_cast<BYTE*>(ctx);
>
> + pDrawables[nIndex] = NULL;
> +
> + PostToWorkerThread(pDrawables);
> +
> return STATUS_SUCCESS;
> }
>
> @@ -4364,7 +4387,7 @@ VOID QxlDevice::SetImageId(InternalImage *internal,
> }
> }
>
> -void QxlDevice::CopyBits(const RECT& rect, const POINT& sourcePoint)
> +QXLDrawable *QxlDevice::CopyBits(const RECT& rect, const POINT& sourcePoint)
> {
>
> This CopyBits and BltBits are not doing anymore the operation, should
> be renamed to something like PrepareCopyBits (better names welcome)
>
> But we still need the driver entry points, right?
> These are method, not really related to entry points.
>
>
> PAGED_CODE();
> QXLDrawable *drawable;
> @@ -4373,18 +4396,18 @@ void QxlDevice::CopyBits(const RECT& rect, const
> POINT& sourcePoint)
>
> if (!(drawable = Drawable(QXL_COPY_BITS, &rect, NULL, 0))) {
> DbgPrint(TRACE_LEVEL_ERROR, ("Cannot get Drawable.\n"));
> - return;
> + return NULL;
> }
>
> drawable->u.copy_bits.src_pos.x = sourcePoint.x;
> drawable->u.copy_bits.src_pos.y = sourcePoint.y;
>
> - PushDrawable(drawable);
> -
> DbgPrint(TRACE_LEVEL_VERBOSE, ("<--- %s\n", __FUNCTION__));
> +
> + return drawable;
> }
>
> -VOID QxlDevice::BltBits (
> +QXLDrawable *QxlDevice::BltBits (
> BLT_INFO* pDst,
> CONST BLT_INFO* pSrc,
> UINT NumRects,
> @@ -4407,7 +4430,7 @@ VOID QxlDevice::BltBits (
>
> if (!(drawable = Drawable(QXL_DRAW_COPY, pRects, NULL, 0))) {
> DbgPrint(TRACE_LEVEL_ERROR, ("Cannot get Drawable.\n"));
> - return;
> + return NULL;
> }
>
> CONST RECT* pRect = &pRects[0];
> @@ -4480,9 +4503,9 @@ VOID QxlDevice::BltBits (
> drawable->surfaces_rects[0].top,
> drawable->surfaces_rects[0].bottom,
> drawable->u.copy.src_bitmap));
>
> - PushDrawable(drawable);
> -
> DbgPrint(TRACE_LEVEL_VERBOSE, ("<--- %s\n", __FUNCTION__));
> +
> + return drawable;
> }
>
> VOID QxlDevice::PutBytesAlign(QXLDataChunk **chunk_ptr, UINT8 **now_ptr,
> diff --git a/qxldod/QxlDod.h b/qxldod/QxlDod.h
> index 4a62680..f441f4b 100755
> --- a/qxldod/QxlDod.h
> +++ b/qxldod/QxlDod.h
> @@ -495,12 +495,12 @@ public:
> BOOLEAN IsBIOSCompatible() { return FALSE; }
> protected:
> NTSTATUS GetModeList(DXGK_DISPLAY_INFORMATION* pDispInfo);
> - VOID BltBits (BLT_INFO* pDst,
> + QXLDrawable *BltBits (BLT_INFO* pDst,
> CONST BLT_INFO* pSrc,
> UINT NumRects,
> _In_reads_(NumRects) CONST RECT *pRects,
> POINT* pSourcePoint);
> - void CopyBits(const RECT& rect, const POINT& sourcePoint);
> + QXLDrawable *CopyBits(const RECT& rect, const POINT& sourcePoint);
> QXLDrawable *Drawable(UINT8 type,
> CONST RECT *area,
> CONST RECT *clip,
>
> Frediano
>
> _______________________________________________
> Spice-devel mailing list
> Spice-devel at lists.freedesktop.org <mailto:Spice-devel at lists.freedesktop.org>
> https://lists.freedesktop.org/mailman/listinfo/spice-devel <https://lists.freedesktop.org/mailman/listinfo/spice-devel>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/spice-devel/attachments/20170321/50c8e8ce/attachment-0001.html>
More information about the Spice-devel
mailing list