drm/omap: partial workaround for DRA7xx DMM errata i878

Errata i878 says that MPU should not be used to access RAM and DMM at
the same time. As it's not possible to prevent MPU accessing RAM, we
need to access DMM via a proxy.

This patch changes DMM driver to access DMM registers via sDMA. Instead
of doing a normal readl/writel call to read/write a register, we use
sDMA to copy 4 bytes from/to the DMM registers.

This patch provides only a partial workaround for i878, as not only DMM
register reads/writes are affected, but also accesses to the DMM mapped
buffers (framebuffers, usually).

Signed-off-by: Tomi Valkeinen <tomi.valkeinen@ti.com>
Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>

2 files changed, 154 insertions, 2 deletions

diff --git a/drivers/gpu/drm/omapdrm/omap_dmm_priv.h b/drivers/gpu/drm/omapdrm/omap_dmm_priv.h
index c2785cc98dc9..60bb3f9297bc 100644
--- a/drivers/gpu/drm/omapdrm/omap_dmm_priv.h
+++ b/drivers/gpu/drm/omapdrm/omap_dmm_priv.h
@@ -159,6 +159,7 @@ struct dmm_platform_data {
 
 struct dmm {
         struct device *dev;
+        dma_addr_t phys_base;
         void __iomem *base;
         int irq;
 
@@ -189,6 +190,12 @@ struct dmm {
         struct list_head alloc_head;
 
         const struct dmm_platform_data *plat_data;
+
+        bool dmm_workaround;
+        spinlock_t wa_lock;
+        u32 *wa_dma_data;
+        dma_addr_t wa_dma_handle;
+        struct dma_chan *wa_dma_chan;
 };
 
 #endif
diff --git a/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c b/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c
index 624d2023dd6b..252f5ebb1acc 100644
--- a/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c
+++ b/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c
@@ -18,6 +18,7 @@
 #include <linux/completion.h>
 #include <linux/delay.h>
 #include <linux/dma-mapping.h>
+#include <linux/dmaengine.h>
 #include <linux/errno.h>
 #include <linux/init.h>
 #include <linux/interrupt.h>
@@ -79,14 +80,138 @@ static const u32 reg[][4] = {
                         DMM_PAT_DESCR__2, DMM_PAT_DESCR__3},
 };
 
+static int dmm_dma_copy(struct dmm *dmm, dma_addr_t src, dma_addr_t dst)
+{
+        struct dma_device *dma_dev = dmm->wa_dma_chan->device;
+        struct dma_async_tx_descriptor *tx;
+        enum dma_status status;
+        dma_cookie_t cookie;
+
+        tx = dma_dev->device_prep_dma_memcpy(dmm->wa_dma_chan, dst, src, 4, 0);
+        if (!tx) {
+                dev_err(dmm->dev, "Failed to prepare DMA memcpy\n");
+                return -EIO;
+        }
+
+        cookie = tx->tx_submit(tx);
+        if (dma_submit_error(cookie)) {
+                dev_err(dmm->dev, "Failed to do DMA tx_submit\n");
+                return -EIO;
+        }
+
+        dma_async_issue_pending(dmm->wa_dma_chan);
+        status = dma_sync_wait(dmm->wa_dma_chan, cookie);
+        if (status != DMA_COMPLETE)
+                dev_err(dmm->dev, "i878 wa DMA copy failure\n");
+
+        dmaengine_terminate_all(dmm->wa_dma_chan);
+        return 0;
+}
+
+static u32 dmm_read_wa(struct dmm *dmm, u32 reg)
+{
+        dma_addr_t src, dst;
+        int r;
+
+        src = dmm->phys_base + reg;
+        dst = dmm->wa_dma_handle;
+
+        r = dmm_dma_copy(dmm, src, dst);
+        if (r) {
+                dev_err(dmm->dev, "sDMA read transfer timeout\n");
+                return readl(dmm->base + reg);
+        }
+
+        /*
+         * As per i878 workaround, the DMA is used to access the DMM registers.
+         * Make sure that the readl is not moved by the compiler or the CPU
+         * earlier than the DMA finished writing the value to memory.
+         */
+        rmb();
+        return readl(dmm->wa_dma_data);
+}
+
+static void dmm_write_wa(struct dmm *dmm, u32 val, u32 reg)
+{
+        dma_addr_t src, dst;
+        int r;
+
+        writel(val, dmm->wa_dma_data);
+        /*
+         * As per i878 workaround, the DMA is used to access the DMM registers.
+         * Make sure that the writel is not moved by the compiler or the CPU, so
+         * the data will be in place before we start the DMA to do the actual
+         * register write.
+         */
+        wmb();
+
+        src = dmm->wa_dma_handle;
+        dst = dmm->phys_base + reg;
+
+        r = dmm_dma_copy(dmm, src, dst);
+        if (r) {
+                dev_err(dmm->dev, "sDMA write transfer timeout\n");
+                writel(val, dmm->base + reg);
+        }
+}
+
 static u32 dmm_read(struct dmm *dmm, u32 reg)
 {
-        return readl(dmm->base + reg);
+        if (dmm->dmm_workaround) {
+                u32 v;
+                unsigned long flags;
+
+                spin_lock_irqsave(&dmm->wa_lock, flags);
+                v = dmm_read_wa(dmm, reg);
+                spin_unlock_irqrestore(&dmm->wa_lock, flags);
+
+                return v;
+        } else {
+                return readl(dmm->base + reg);
+        }
 }
 
 static void dmm_write(struct dmm *dmm, u32 val, u32 reg)
 {
-        writel(val, dmm->base + reg);
+        if (dmm->dmm_workaround) {
+                unsigned long flags;
+
+                spin_lock_irqsave(&dmm->wa_lock, flags);
+                dmm_write_wa(dmm, val, reg);
+                spin_unlock_irqrestore(&dmm->wa_lock, flags);
+        } else {
+                writel(val, dmm->base + reg);
+        }
+}
+
+static int dmm_workaround_init(struct dmm *dmm)
+{
+        dma_cap_mask_t mask;
+
+        spin_lock_init(&dmm->wa_lock);
+
+        dmm->wa_dma_data = dma_alloc_coherent(dmm->dev,  sizeof(u32),
+                                              &dmm->wa_dma_handle, GFP_KERNEL);
+        if (!dmm->wa_dma_data)
+                return -ENOMEM;
+
+        dma_cap_zero(mask);
+        dma_cap_set(DMA_MEMCPY, mask);
+
+        dmm->wa_dma_chan = dma_request_channel(mask, NULL, NULL);
+        if (!dmm->wa_dma_chan) {
+                dma_free_coherent(dmm->dev, 4, dmm->wa_dma_data, dmm->wa_dma_handle);
+                return -ENODEV;
+        }
+
+        return 0;
+}
+
+static void dmm_workaround_uninit(struct dmm *dmm)
+{
+        dma_release_channel(dmm->wa_dma_chan);
+
+        dma_free_coherent(dmm->dev, 4, dmm->wa_dma_data, dmm->wa_dma_handle);
 }
 
 /* simple allocator to grab next 16 byte aligned memory from txn */
@@ -640,6 +765,9 @@ static int omap_dmm_remove(struct platform_device *dev)
                 if (omap_dmm->dummy_page)
                         __free_page(omap_dmm->dummy_page);
 
+                if (omap_dmm->dmm_workaround)
+                        dmm_workaround_uninit(omap_dmm);
+
                 iounmap(omap_dmm->base);
                 kfree(omap_dmm);
                 omap_dmm = NULL;
@@ -685,6 +813,7 @@ static int omap_dmm_probe(struct platform_device *dev)
                 goto fail;
         }
 
+        omap_dmm->phys_base = mem->start;
         omap_dmm->base = ioremap(mem->start, SZ_2K);
 
         if (!omap_dmm->base) {
@@ -700,6 +829,22 @@ static int omap_dmm_probe(struct platform_device *dev)
 
         omap_dmm->dev = &dev->dev;
 
+        if (of_machine_is_compatible("ti,dra7")) {
+                /*
+                 * DRA7 Errata i878 says that MPU should not be used to access
+                 * RAM and DMM at the same time. As it's not possible to prevent
+                 * MPU accessing RAM, we need to access DMM via a proxy.
+                 */
+                if (!dmm_workaround_init(omap_dmm)) {
+                        omap_dmm->dmm_workaround = true;
+                        dev_info(&dev->dev,
+                                "workaround for errata i878 in use\n");
+                } else {
+                        dev_warn(&dev->dev,
+                                 "failed to initialize work-around for i878\n");
+                }
+        }
+
         hwinfo = dmm_read(omap_dmm, DMM_PAT_HWINFO);
         omap_dmm->num_engines = (hwinfo >> 24) & 0x1F;
         omap_dmm->num_lut = (hwinfo >> 16) & 0x1F;