Fix #2596

fitz/fitz.i

@@ -4625,6 +4625,7 @@ if basestate:
                     old_annots[k] = v
                 page._erase()  # remove the page
                 page = None
+                TOOLS.store_shrink(100)
                 page = self.load_page(pno)  # reload the page
 
                 # copy annot refs over to the new dictionary

src/__init__.py

@@ -5040,8 +5040,23 @@ def reload_page(self, page: "struct Page *") -> "struct Page *":
         pno = page.number  # save the page number
         for k, v in page._annot_refs.items():  # save the annot dictionary
             old_annots[k] = v
+
+        # We need to set `page.this` to `None` here in order to force the
+        # `mupdf::FzPage` to be removed from the MuPDF document's internal
+        # list, so that when `self.load_page()` is called below, its call of
+        # `fz_load_page()` will create a new `mupdf::FzPage`.
+        #
+        # This only works if the MuPDF `fz_page`'s reference count is 1. User
+        # code will have to work fairly hard to break this, by somehow making
+        # a copy of the mupdf::FzPage itself; for example making a copy in
+        # Python will usually create a new Python object that refers to the
+        # same mupdf::FzPage and this not increment the MuPDF reference count.
+        #
+        assert page.this.m_internal.refs == 1
+        page.this = None
         page._erase()  # remove the page
         page = None
+        TOOLS.store_shrink(100)
         page = self.load_page(pno)  # reload the page
 
         # copy annot refs over to the new dictionary

tests/test_general.py

@@ -539,3 +539,17 @@ def test_2692():
                 clip=fitz.Rect(0,0,10,10),
                 )
 
+
+def test_2596():
+    """Cconfirm correctly abandoning cache when reloading a page."""
+    doc = fitz.Document(f"{scriptdir}/resources/test_2596.pdf")
+    page = doc[0]
+    pix0 = page.get_pixmap()  # render the page
+    _ = doc.tobytes(garbage=3)  # save with garbage collection
+
+    # Note this will invalidate cache content for this page.
+    # Reloading the page now empties the cache, so rendering
+    # will deliver the same pixmap
+    page = doc.reload_page(page)
+    pix1 = page.get_pixmap()
+    assert pix1.samples == pix0.samples