agp/intel: Reinforce the barrier after GTT updates

After changing the timing between GTT updates and execution on the GPU,
we started seeing sporadic failures on Ironlake. These were narrowed
down to being an insufficiently strong enough barrier/delay after
updating the GTT and scheduling execution on the GPU. By forcing the
uncached read, and adding the missing barrier for the singular
insert_page (relocation paths), the sporadic failures go away.

Fixes: 983d308 ("agp/intel: Serialise after GTT updates")
Fixes: 3497971 ("agp/intel: Flush chipset writes after updating a single PTE")
Signed-off-by: Chris Wilson <[email protected]>
Acked-by: Andi Shyti <[email protected]>
Cc: [email protected] # v4.0+
Link: https://patchwork.freedesktop.org/patch/msgid/[email protected]

drivers/char/agp/intel-gtt.c

@@ -846,6 +846,7 @@ void intel_gtt_insert_page(dma_addr_t addr,
 			   unsigned int flags)
 {
 	intel_private.driver->write_entry(addr, pg, flags);
+	readl(intel_private.gtt + pg);
 	if (intel_private.driver->chipset_flush)
 		intel_private.driver->chipset_flush();
 }
@@ -871,7 +872,7 @@ void intel_gtt_insert_sg_entries(struct sg_table *st,
 			j++;
 		}
 	}
-	wmb();
+	readl(intel_private.gtt + j - 1);
 	if (intel_private.driver->chipset_flush)
 		intel_private.driver->chipset_flush();
 }
@@ -1105,6 +1106,7 @@ static void i9xx_cleanup(void)
 
 static void i9xx_chipset_flush(void)
 {
+	wmb();
 	if (intel_private.i9xx_flush_page)
 		writel(1, intel_private.i9xx_flush_page);
 }