Created attachment 95593 [details] Web page: Korean text with shadow property When drawing text shadow in EFL port(EWebLauncher) out of memory occurs. Steps to reproduce: run EWebLauncher with attached file, zoom out, OOM occurs. The problem is that for shadowed text(e.g. in attachment) font data for some glyphs is different than for first. In this case drawGlyphs() function is invoked(FontFastPath.cpp, line 387). 374 void Font::drawGlyphBuffer(GraphicsContext* context, const GlyphBuffer& glyphBuffer, const FloatPoint& point) const 375 { 376 // Draw each contiguous run of glyphs that use the same font data. 377 const SimpleFontData* fontData = glyphBuffer.fontDataAt(0); 378 FloatSize offset = glyphBuffer.offsetAt(0); 379 FloatPoint startPoint(point); 380 float nextX = startPoint.x(); 381 int lastFrom = 0; 382 int nextGlyph = 0; 383 while (nextGlyph < glyphBuffer.size()) { 384 const SimpleFontData* nextFontData = glyphBuffer.fontDataAt(nextGlyph); 385 FloatSize nextOffset = glyphBuffer.offsetAt(nextGlyph); 386 if (nextFontData != fontData || nextOffset != offset) { /// <------------------------------------------- 387 drawGlyphs(context, fontData, glyphBuffer, lastFrom, nextGlyph - lastFrom, startPoint); 388 389 lastFrom = nextGlyph; 390 fontData = nextFontData; 391 offset = nextOffset; 392 startPoint.setX(nextX); 393 } 394 nextX += glyphBuffer.advanceAt(nextGlyph); 395 nextGlyph++; 396 } 397 398 drawGlyphs(context, fontData, glyphBuffer, lastFrom, nextGlyph - lastFrom, startPoint); 399 } Than path is as below: FontCairo.cpp: drawGlyphs()->drawGlyphsShadow()->beginShadowLayer() ContextShadowCairo.cpp: beginShadowLayer()->adjustBlurDistance() 86 PlatformContext ContextShadow::beginShadowLayer(GraphicsContext* context, const FloatRect& layerArea) 87 { 88 adjustBlurDistance(context); 89 90 double x1, x2, y1, y2; 91 cairo_clip_extents(context->platformContext()->cr(), &x1, &y1, &x2, &y2); 92 IntRect layerRect = calculateLayerBoundingRect(context, layerArea, IntRect(x1, y1, x2 - x1, y2 - y1)); 93 94 // Don't paint if we are totally outside the clip region. 95 if (layerRect.isEmpty()) 96 return 0; 97 98 m_layerImage = getScratchBuffer(layerRect.size()); 99 m_layerContext = cairo_create(m_layerImage); 100 101 // Always clear the surface first. 102 cairo_set_operator(m_layerContext, CAIRO_OPERATOR_CLEAR); 103 cairo_paint(m_layerContext); 104 cairo_set_operator(m_layerContext, CAIRO_OPERATOR_OVER); 105 106 cairo_translate(m_layerContext, m_layerContextTranslation.x(), m_layerContextTranslation.y()); 107 return m_layerContext; 108 } 181 void ContextShadow::adjustBlurDistance(GraphicsContext* context) 182 { 183 const AffineTransform transform = context->getCTM(); 184 185 // Adjust blur if we're scaling, since the radius must not be affected by transformations. 186 if (transform.isIdentity()) 187 return; 188 ........................ 206 m_blurDistance = roundf(static_cast<float>(m_blurDistance) / scale); 207 } Using m_blurDistance size of cairo context is calculated. For GTK port there is no transformation set on cairo, so transform.isIdentity() always returns true(only because of it OOM doesn't occur in GTK). But for EFL port when zoom is done following function(ewk_view.cpp) is invoked, which changes cairo_matrix. 3324 void ewk_view_paint_context_translate(Ewk_View_Paint_Context* ctxt, float x, float y) 3325 { 3326 EINA_SAFETY_ON_NULL_RETURN(ctxt); 3327 3328 ctxt->gc->translate(x, y); 3329 } In this for scale < 1.0 m_blurDistance is calculated as, e.g. for 0.5 m_blurDistance = m_blurDistance/0.5 several times, so this distance growths and growths, and cairo context is tried to be created for greater and greater rectangle, and at the end out of memory condition occurs.
Created attachment 95697 [details] Patch
I'll prepare simple local patch to prevent it. But, I'm not sure that it's way to go. This is related to not only EFL, but also GTK. could anyone help to find proper way?
Since it's an out of memory crash, the fix should be testable in a regression test. It seems surprising if cross-platform code needs to be modified for a EFL only bug.
Comment on attachment 95697 [details] Patch Seems like an EFL bug.
Created attachment 95853 [details] testpage.html This html page will make gtklauncher and eweblauncher sluggish on host pc because of memory. I tested chrome, QtTestBrowser, but they looks fine.
Created attachment 95854 [details] screenshot
(In reply to comment #4) > (From update of attachment 95697 [details]) > Seems like an EFL bug. No, It's not related to EFL. At least, it will be an issue of cairo port. But, I wonder why only Cairo(GTK and EFL I tested) have this issue. I should study other ports more.
This is general issue of cairo backend. Wrong is calculating m_blurDistance in each Font::drawGlyphs() (FontCairo.cpp) invocation. Maybe ContextShadow should be cleared at the start of drawGlyphsShadow() function.
(In reply to comment #8) > This is general issue of cairo backend. Wrong is calculating m_blurDistance in > each Font::drawGlyphs() (FontCairo.cpp) invocation. Maybe ContextShadow should > be cleared at the start of drawGlyphsShadow() function. Is it also possible to reproduce this on the WinCairo port?
(In reply to comment #9) > (In reply to comment #8) > > This is general issue of cairo backend. Wrong is calculating m_blurDistance in > > each Font::drawGlyphs() (FontCairo.cpp) invocation. Maybe ContextShadow should > > be cleared at the start of drawGlyphsShadow() function. > > Is it also possible to reproduce this on the WinCairo port? I'm not sure because I can't prepare WinCairo port. I'll try to prepare and test once more. Anyway, I found some history. adjustBlurDistance was introduced at Bug 51161 and it is only for Qt and Cairo port. However, Qt doesn't call drawGlyphs (I tested on qt-everywhere-opensource-src-4.7.3). As my understanding Bug 51161, adjustBlurDistance should be called only one time. but drawGlyphBuffer calls repeatedly adjustBlurDistance via drawGlyphs. IMO, current patch is valid or adjustBlurDistance should check whether it was applied.
In latest, this isn't reproduced.