diff options
Diffstat (limited to 'drivers/gpu/drm/nouveau/nouveau_dma.c')
-rw-r--r-- | drivers/gpu/drm/nouveau/nouveau_dma.c | 198 |
1 files changed, 170 insertions, 28 deletions
diff --git a/drivers/gpu/drm/nouveau/nouveau_dma.c b/drivers/gpu/drm/nouveau/nouveau_dma.c index 703553687b20..c8482a108a78 100644 --- a/drivers/gpu/drm/nouveau/nouveau_dma.c +++ b/drivers/gpu/drm/nouveau/nouveau_dma.c @@ -29,12 +29,37 @@ #include "nouveau_drv.h" #include "nouveau_dma.h" +void +nouveau_dma_pre_init(struct nouveau_channel *chan) +{ + struct drm_nouveau_private *dev_priv = chan->dev->dev_private; + struct nouveau_bo *pushbuf = chan->pushbuf_bo; + + if (dev_priv->card_type == NV_50) { + const int ib_size = pushbuf->bo.mem.size / 2; + + chan->dma.ib_base = (pushbuf->bo.mem.size - ib_size) >> 2; + chan->dma.ib_max = (ib_size / 8) - 1; + chan->dma.ib_put = 0; + chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put; + + chan->dma.max = (pushbuf->bo.mem.size - ib_size) >> 2; + } else { + chan->dma.max = (pushbuf->bo.mem.size >> 2) - 2; + } + + chan->dma.put = 0; + chan->dma.cur = chan->dma.put; + chan->dma.free = chan->dma.max - chan->dma.cur; +} + int nouveau_dma_init(struct nouveau_channel *chan) { struct drm_device *dev = chan->dev; struct drm_nouveau_private *dev_priv = dev->dev_private; struct nouveau_gpuobj *m2mf = NULL; + struct nouveau_gpuobj *nvsw = NULL; int ret, i; /* Create NV_MEMORY_TO_MEMORY_FORMAT for buffer moves */ @@ -47,6 +72,15 @@ nouveau_dma_init(struct nouveau_channel *chan) if (ret) return ret; + /* Create an NV_SW object for various sync purposes */ + ret = nouveau_gpuobj_sw_new(chan, NV_SW, &nvsw); + if (ret) + return ret; + + ret = nouveau_gpuobj_ref_add(dev, chan, NvSw, nvsw, NULL); + if (ret) + return ret; + /* NV_MEMORY_TO_MEMORY_FORMAT requires a notifier object */ ret = nouveau_notifier_alloc(chan, NvNotify0, 32, &chan->m2mf_ntfy); if (ret) @@ -64,12 +98,6 @@ nouveau_dma_init(struct nouveau_channel *chan) return ret; } - /* Initialise DMA vars */ - chan->dma.max = (chan->pushbuf_bo->bo.mem.size >> 2) - 2; - chan->dma.put = 0; - chan->dma.cur = chan->dma.put; - chan->dma.free = chan->dma.max - chan->dma.cur; - /* Insert NOPS for NOUVEAU_DMA_SKIPS */ ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS); if (ret) @@ -87,6 +115,13 @@ nouveau_dma_init(struct nouveau_channel *chan) BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_NOTIFY, 1); OUT_RING(chan, NvNotify0); + /* Initialise NV_SW */ + ret = RING_SPACE(chan, 2); + if (ret) + return ret; + BEGIN_RING(chan, NvSubSw, 0, 1); + OUT_RING(chan, NvSw); + /* Sit back and pray the channel works.. */ FIRE_RING(chan); @@ -106,37 +141,66 @@ OUT_RINGp(struct nouveau_channel *chan, const void *data, unsigned nr_dwords) chan->dma.cur += nr_dwords; } -static inline bool -READ_GET(struct nouveau_channel *chan, uint32_t *get) +/* Fetch and adjust GPU GET pointer + * + * Returns: + * value >= 0, the adjusted GET pointer + * -EINVAL if GET pointer currently outside main push buffer + * -EBUSY if timeout exceeded + */ +static inline int +READ_GET(struct nouveau_channel *chan, uint32_t *prev_get, uint32_t *timeout) { uint32_t val; val = nvchan_rd32(chan, chan->user_get); - if (val < chan->pushbuf_base || - val >= chan->pushbuf_base + chan->pushbuf_bo->bo.mem.size) { - /* meaningless to dma_wait() except to know whether the - * GPU has stalled or not - */ - *get = val; - return false; + + /* reset counter as long as GET is still advancing, this is + * to avoid misdetecting a GPU lockup if the GPU happens to + * just be processing an operation that takes a long time + */ + if (val != *prev_get) { + *prev_get = val; + *timeout = 0; + } + + if ((++*timeout & 0xff) == 0) { + DRM_UDELAY(1); + if (*timeout > 100000) + return -EBUSY; } - *get = (val - chan->pushbuf_base) >> 2; - return true; + if (val < chan->pushbuf_base || + val > chan->pushbuf_base + (chan->dma.max << 2)) + return -EINVAL; + + return (val - chan->pushbuf_base) >> 2; } -int -nouveau_dma_wait(struct nouveau_channel *chan, int size) +void +nv50_dma_push(struct nouveau_channel *chan, struct nouveau_bo *bo, + int delta, int length) { - uint32_t get, prev_get = 0, cnt = 0; - bool get_valid; + struct nouveau_bo *pb = chan->pushbuf_bo; + uint64_t offset = bo->bo.offset + delta; + int ip = (chan->dma.ib_put * 2) + chan->dma.ib_base; - while (chan->dma.free < size) { - /* reset counter as long as GET is still advancing, this is - * to avoid misdetecting a GPU lockup if the GPU happens to - * just be processing an operation that takes a long time - */ - get_valid = READ_GET(chan, &get); + BUG_ON(chan->dma.ib_free < 1); + nouveau_bo_wr32(pb, ip++, lower_32_bits(offset)); + nouveau_bo_wr32(pb, ip++, upper_32_bits(offset) | length << 8); + + chan->dma.ib_put = (chan->dma.ib_put + 1) & chan->dma.ib_max; + nvchan_wr32(chan, 0x8c, chan->dma.ib_put); + chan->dma.ib_free--; +} + +static int +nv50_dma_push_wait(struct nouveau_channel *chan, int count) +{ + uint32_t cnt = 0, prev_get = 0; + + while (chan->dma.ib_free < count) { + uint32_t get = nvchan_rd32(chan, 0x88); if (get != prev_get) { prev_get = get; cnt = 0; @@ -148,6 +212,71 @@ nouveau_dma_wait(struct nouveau_channel *chan, int size) return -EBUSY; } + chan->dma.ib_free = get - chan->dma.ib_put; + if (chan->dma.ib_free <= 0) + chan->dma.ib_free += chan->dma.ib_max + 1; + } + + return 0; +} + +static int +nv50_dma_wait(struct nouveau_channel *chan, int slots, int count) +{ + uint32_t cnt = 0, prev_get = 0; + int ret; + + ret = nv50_dma_push_wait(chan, slots + 1); + if (unlikely(ret)) + return ret; + + while (chan->dma.free < count) { + int get = READ_GET(chan, &prev_get, &cnt); + if (unlikely(get < 0)) { + if (get == -EINVAL) + continue; + + return get; + } + + if (get <= chan->dma.cur) { + chan->dma.free = chan->dma.max - chan->dma.cur; + if (chan->dma.free >= count) + break; + + FIRE_RING(chan); + do { + get = READ_GET(chan, &prev_get, &cnt); + if (unlikely(get < 0)) { + if (get == -EINVAL) + continue; + return get; + } + } while (get == 0); + chan->dma.cur = 0; + chan->dma.put = 0; + } + + chan->dma.free = get - chan->dma.cur - 1; + } + + return 0; +} + +int +nouveau_dma_wait(struct nouveau_channel *chan, int slots, int size) +{ + uint32_t prev_get = 0, cnt = 0; + int get; + + if (chan->dma.ib_max) + return nv50_dma_wait(chan, slots, size); + + while (chan->dma.free < size) { + get = READ_GET(chan, &prev_get, &cnt); + if (unlikely(get == -EBUSY)) + return -EBUSY; + /* loop until we have a usable GET pointer. the value * we read from the GPU may be outside the main ring if * PFIFO is processing a buffer called from the main ring, @@ -157,7 +286,7 @@ nouveau_dma_wait(struct nouveau_channel *chan, int size) * from the SKIPS area, so the code below doesn't have to deal * with some fun corner cases. */ - if (!get_valid || get < NOUVEAU_DMA_SKIPS) + if (unlikely(get == -EINVAL) || get < NOUVEAU_DMA_SKIPS) continue; if (get <= chan->dma.cur) { @@ -183,6 +312,19 @@ nouveau_dma_wait(struct nouveau_channel *chan, int size) * after processing the currently pending commands. */ OUT_RING(chan, chan->pushbuf_base | 0x20000000); + + /* wait for GET to depart from the skips area. + * prevents writing GET==PUT and causing a race + * condition that causes us to think the GPU is + * idle when it's not. + */ + do { + get = READ_GET(chan, &prev_get, &cnt); + if (unlikely(get == -EBUSY)) + return -EBUSY; + if (unlikely(get == -EINVAL)) + continue; + } while (get <= NOUVEAU_DMA_SKIPS); WRITE_PUT(NOUVEAU_DMA_SKIPS); /* we're now submitting commands at the start of |