linux-kernel - [RFC PATCH] drm/etnaviv: set the 32-bit coherent DMA mask later

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [day] [month] [year] [list]

Message-ID: <20250719064624.690606-1-uwu@icenowy.me>
Date: Sat, 19 Jul 2025 14:46:24 +0800
From: Icenowy Zheng <uwu@...nowy.me>
To: Lucas Stach <l.stach@...gutronix.de>,
	Russell King <linux+etnaviv@...linux.org.uk>,
	Christian Gmeiner <christian.gmeiner@...il.com>,
	David Airlie <airlied@...il.com>,
	Simona Vetter <simona@...ll.ch>
Cc: etnaviv@...ts.freedesktop.org,
	dri-devel@...ts.freedesktop.org,
	linux-kernel@...r.kernel.org,
	Icenowy Zheng <uwu@...nowy.me>
Subject: [RFC PATCH] drm/etnaviv: set the 32-bit coherent DMA mask later

The etnaviv driver wants to use 40-bit DMA mask and 32-bit coherent
allocation mask, to be able to access memory up to 40-bit, but still
limit page tables to lower 32-bit.

However, the call to of_dma_configure() next to setting the masks ruined
everything -- of_dma_configure will use the coherent allocation mask to
clip the DMA accessibility mask, leads to a setup with both masks as
32-bit.

In this situation PRIME imports stop to work because of SWIOTLB getting
activated.

Set a 40-bit coherent mask before of_dma_configure() and then the real
32-bit mask after it, to prevent main DMA mask being clipped.

Signed-off-by: Icenowy Zheng <uwu@...nowy.me>
---
This patch is marked RFC because the practice looks a little dirty.

P.S. unfortunately the testing situation isn't mainlined yet -- I am
using etnaviv as the X.org 2D accelerator on TH1520 SoC (with
xf86-video-thead DDX), and importing buffers from powervr device
(open source Vulkan driver, hacked to add xcb WSI support). Both devices
are a few patches away from mainline, and the display engine
(VeriSilicon DC8200) support is a big TODO.

 drivers/gpu/drm/etnaviv/etnaviv_drv.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.c b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
index 3e91747ed339..cded63904c65 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
@@ -650,9 +650,11 @@ static int etnaviv_pdev_probe(struct platform_device *pdev)
 	 * To make things easy, we set the dma_coherent_mask to 32
 	 * bit to make sure we are allocating the command buffers and
 	 * TLBs in the lower 4 GiB address space.
+	 *
+	 * As coherent_mask may mislead of_dma_configure(), set it to
+	 * the same with main dma mask temporarily as a hack.
 	 */
-	if (dma_set_mask(dev, DMA_BIT_MASK(40)) ||
-	    dma_set_coherent_mask(dev, DMA_BIT_MASK(32))) {
+	if (dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(40))) {
 		dev_dbg(dev, "No suitable DMA available\n");
 		return -ENODEV;
 	}
@@ -668,6 +670,13 @@ static int etnaviv_pdev_probe(struct platform_device *pdev)
 		of_node_put(first_node);
 	}

+	/*
+	 * Now as of_dma_configure() is done, we're now safe to set
+	 * coherent_dma_mask to the real value.
+	 */
+	if (dma_set_coherent_mask(&pdev->dev, DMA_BIT_MASK(32)))
+		return -ENODEV;
+
 	return component_master_add_with_match(dev, &etnaviv_master_ops, match);
 }

-- 
2.50.1