Hi Ben,

My bad, I will correct that. I'll have a better look at the three patches you attached tomorrow.

Cheers,

Pierre

On 01:42 PM - Dec 11 2014, Ben Skeggs wrote:

On Wed, Dec 10, 2014 at 5:53 PM, Pierre Moreau pierre.morrow@free.fr wrote:

...

(This is a v3 of patch "drm/nouveau/fb/nv50: Add PFB writes")

This fix a GPU lockup on 9400M (NVAC) when using acceleration, see https://bugs.freedesktop.org/show_bug.cgi?id=27501

v2:

• Move code to subdev/fb/nv50.c as suggested by Roy Spliet;
• Remove arbitrary writes to 100c18/100c24
• Replace write to 100c1c of arbitrary value by the address of a scratch page as proposed by Ilia Mirkin;
• Remove enabling of bits 16 and 0 as they don't yield in any changes.

v3:

• Move code to subdev/fb/nvaa.c as suggested by Ilia Mirkin.

The following changes were made thanks to information provided by Robert Morell from NVidia:

• Allocate a dma page for use by the pollers;
• Re-enable pollers at bits 16 and 0;
• Set pollers address to a proper value.

Hey Pierre,

This patch is incorrect. As Robert pointed out in an older thread, the registers don't take the physical address of a page, but a somewhat odd "negative offset from the end of carveout".

See this example (referring to Robert's last email):

220.926392 read32 #3 +0x00100e10 -> 0x00070000 220.926406 read32 #3 +0x00100e14 -> 0x00010000

carveout_base = 0x70000000 carveout_size = 0x10000000

-- snip --

223.300495 read32 #3 +0x00100c14 -> 0x00000000 223.300521 read32 #3 +0x00100c18 -> 0x00000000 223.300547 write32 #3 +0x00100c18 <- 0x000027ff

base = (0x70000000 + 0x10000000) - ((0x27ff + 1) << 5) base = 0x7ffb0000 << in carveout

223.300573 read32 #3 +0x00100c14 -> 0x00000000 223.300599 write32 #3 +0x00100c14 <- 0x00000001 223.300625 read32 #3 +0x00100c1c -> 0x00000002 223.300651 write32 #3 +0x00100c1c <- 0x000027fe

base = (0x70000000 + 0x10000000) - ((0x27fe + 1) << 5) base = 0x7ffb0020 << in carveout

223.300677 read32 #3 +0x00100c14 -> 0x00000001 223.300702 write32 #3 +0x00100c14 <- 0x00000003 223.300728 read32 #3 +0x00100c24 -> 0x00000004 223.300754 write32 #3 +0x00100c24 <- 0x000027fd

base = (0x70000000 + 0x10000000) - ((0x27ff + 1) << 5) base = 0x7ffb0040 << in carveout

223.300780 read32 #3 +0x00100c14 -> 0x00000003 223.300806 write32 #3 +0x00100c14 <- 0x00010003

It's a bit tricky at that point in the nouveau init process to allocate device memory, but, we already reserve 1MiB of memory at the end of "VRAM" (carveout / stolen memory) so we can make use of that.

I've attached 3 patches. The first two of them, I'll probably merge as-is.

The third patch is an example of how this should probably be done, in case you want to continue tackling this some more :)

Cheers, Ben.

...

Signed-off-by: Pierre Moreau pierre.morrow@free.fr

drm/core/subdev/fb/nvaa.h | 1 + nvkm/subdev/fb/nv50.h | 4 +++ nvkm/subdev/fb/nvaa.c | 74 ++++++++++++++++++++++++++++++++++++++++++++--- nvkm/subdev/fb/nvaa.h | 19 ++++++++++++ 4 files changed, 94 insertions(+), 4 deletions(-) create mode 120000 drm/core/subdev/fb/nvaa.h create mode 100644 nvkm/subdev/fb/nvaa.h

diff --git a/drm/core/subdev/fb/nvaa.h b/drm/core/subdev/fb/nvaa.h new file mode 120000 index 0000000..b450e8c --- /dev/null +++ b/drm/core/subdev/fb/nvaa.h @@ -0,0 +1 @@ +../../../../nvkm/subdev/fb/nvaa.h \ No newline at end of file diff --git a/nvkm/subdev/fb/nv50.h b/nvkm/subdev/fb/nv50.h index c5e5a88..0b20975 100644 --- a/nvkm/subdev/fb/nv50.h +++ b/nvkm/subdev/fb/nv50.h @@ -9,6 +9,10 @@ struct nv50_fb_priv { dma_addr_t r100c08; };

+#define nv50_fb_create(p,e,c,d,o) \

•   nv50_fb_ctor((p), (e), (c), (d), sizeof(**o),                          \
  

•                   (struct nouveau_object **)o)
  

int nv50_fb_ctor(struct nouveau_object *, struct nouveau_object *, struct nouveau_oclass *, void *, u32, struct nouveau_object **); diff --git a/nvkm/subdev/fb/nvaa.c b/nvkm/subdev/fb/nvaa.c index cba8e68..b70ab2f 100644 --- a/nvkm/subdev/fb/nvaa.c +++ b/nvkm/subdev/fb/nvaa.c @@ -22,15 +22,81 @@

• Authors: Ben Skeggs

*/

-#include "nv50.h" +#include "nvaa.h"

+int +nvaa_fb_ctor(struct nouveau_object *parent, struct nouveau_object *engine,

•        struct nouveau_oclass *oclass, void *data, u32 size,
  

•        struct nouveau_object **pobject)
  

+{

•   struct nouveau_device *device = nv_device(parent);
  

•   struct nvaa_fb_priv *priv;
  

•   int ret;
  

•   ret = nv50_fb_create(parent, engine, oclass, data, &priv);
  

•   *pobject = nv_object(priv);
  

•   if (ret)
  

•           return ret;
  

•   priv = (struct nvaa_fb_priv *)(*pobject);
  

•   priv->r100c18_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
  

•   if (priv->r100c18_page) {
  

•           priv->r100c18 = dma_map_page(nv_device_base(device),
  

•                                        priv->r100c18_page, 0, PAGE_SIZE,
  

•                                        DMA_BIDIRECTIONAL);
  

•           if (dma_mapping_error(nv_device_base(device), priv->r100c18))
  

•                   return -EFAULT;
  

•   } else {
  

•           nv_warn(priv, "failed 0x100c18 page alloc\n");
  

•   }
  

•   return 0;
  

+}

+void +nvaa_fb_dtor(struct nouveau_object *object) +{

•   struct nouveau_device *device = nv_device(object);
  

•   struct nvaa_fb_priv *priv = (void *)object;
  

•   if (priv->r100c18_page) {
  

•           dma_unmap_page(nv_device_base(device), priv->r100c18, PAGE_SIZE,
  

•                          DMA_BIDIRECTIONAL);
  

•           __free_page(priv->r100c18_page);
  

•   }
  

•   nv50_fb_dtor(object);
  

+}

+int +nvaa_fb_init(struct nouveau_object *object) +{

•   struct nvaa_fb_priv *priv = (void *)object;
  

•   int ret;
  

•   ret = nv50_fb_init(object);
  

•   if (ret)
  

•           return ret;
  

•   /* Enable NISO poller for various clients and set their associated
  

•    * read address, only for MCP77/78 and MCP79/7A. (fd#25701)
  

•    */
  

•   nv_wr32(priv, 0x100c18, priv->r100c18 >> 8);
  

•   nv_mask(priv, 0x100c14, 0x00000000, 0x00000001);
  

•   nv_wr32(priv, 0x100c1c, (priv->r100c18 >> 8) + 1);
  

•   nv_mask(priv, 0x100c14, 0x00000000, 0x00000002);
  

•   nv_wr32(priv, 0x100c24, (priv->r100c18 >> 8) + 2);
  

•   nv_mask(priv, 0x100c14, 0x00000000, 0x00010000);
  

•   return 0;
  

+}

struct nouveau_oclass * nvaa_fb_oclass = &(struct nv50_fb_impl) { .base.base.handle = NV_SUBDEV(FB, 0xaa), .base.base.ofuncs = &(struct nouveau_ofuncs) {

•           .ctor = nv50_fb_ctor,
  

•           .dtor = nv50_fb_dtor,
  

•           .init = nv50_fb_init,
  

•           .ctor = nvaa_fb_ctor,
  

•           .dtor = nvaa_fb_dtor,
  

•           .init = nvaa_fb_init,
            .fini = _nouveau_fb_fini,
    },
    .base.memtype = nv50_fb_memtype_valid,
  

diff --git a/nvkm/subdev/fb/nvaa.h b/nvkm/subdev/fb/nvaa.h new file mode 100644 index 0000000..84e1eca --- /dev/null +++ b/nvkm/subdev/fb/nvaa.h @@ -0,0 +1,19 @@ +#ifndef __NVKM_FB_NVAA_H__ +#define __NVKM_FB_NVAA_H__

+#include "nv50.h"

+struct nvaa_fb_priv {

•   struct nv50_fb_priv base;
  

•   struct page *r100c18_page;
  

•   dma_addr_t r100c18;
  

+};

+int nvaa_fb_ctor(struct nouveau_object *, struct nouveau_object *,

•             struct nouveau_oclass *, void *, u32,
  

•             struct nouveau_object **);
  

+void nvaa_fb_dtor(struct nouveau_object *); +int nvaa_fb_init(struct nouveau_object *);

+#endif

2.1.3

---

Nouveau mailing list Nouveau@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/nouveau

From 12b3ad50288ceaba3feae5a11772014a3a334953 Mon Sep 17 00:00:00 2001 From: Ben Skeggs bskeggs@redhat.com Date: Thu, 11 Dec 2014 13:19:31 +1000 Subject: [PATCH 1/3] fb/ram/mcp77: subclass nouveau_ram

Signed-off-by: Ben Skeggs bskeggs@redhat.com

nvkm/subdev/fb/ramnvaa.c | 24 ++++++++++++++---------- 1 file changed, 14 insertions(+), 10 deletions(-)

diff --git a/nvkm/subdev/fb/ramnvaa.c b/nvkm/subdev/fb/ramnvaa.c index 00f2ca7..8ee3d37 100644 --- a/nvkm/subdev/fb/ramnvaa.c +++ b/nvkm/subdev/fb/ramnvaa.c @@ -24,6 +24,10 @@

#include "nv50.h"

+struct nvaa_ram_priv {

• struct nouveau_ram base;

+};

static int nvaa_ram_ctor(struct nouveau_object *parent, struct nouveau_object *engine, struct nouveau_oclass *oclass, void *data, u32 datasize, @@ -32,26 +36,26 @@ nvaa_ram_ctor(struct nouveau_object *parent, struct nouveau_object *engine, const u32 rsvd_head = ( 256 * 1024) >> 12; /* vga memory */ const u32 rsvd_tail = (1024 * 1024) >> 12; /* vbios etc */ struct nouveau_fb *pfb = nouveau_fb(parent);

• struct nouveau_ram *ram;

• struct nvaa_ram_priv *priv; int ret;

• ret = nouveau_ram_create(parent, engine, oclass, &ram);
• *pobject = nv_object(ram);

• ret = nouveau_ram_create(parent, engine, oclass, &priv);
• *pobject = nv_object(priv); if (ret) return ret;

• ram->size = nv_rd32(pfb, 0x10020c);
• ram->size = (ram->size & 0xffffff00) | ((ram->size & 0x000000ff) << 32);

• priv->base.size = nv_rd32(pfb, 0x10020c);
• priv->base.size = (priv->base.size & 0xffffff00) | ((priv->base.size & 0x000000ff) << 32);

• ret = nouveau_mm_init(&pfb->vram, rsvd_head, (ram->size >> 12) -

• ret = nouveau_mm_init(&pfb->vram, rsvd_head, (priv->base.size >> 12) - (rsvd_head + rsvd_tail), 1); if (ret) return ret;

• ram->type = NV_MEM_TYPE_STOLEN;
• ram->stolen = (u64)nv_rd32(pfb, 0x100e10) << 12;
• ram->get = nv50_ram_get;
• ram->put = nv50_ram_put;

• priv->base.type = NV_MEM_TYPE_STOLEN;
• priv->base.stolen = (u64)nv_rd32(pfb, 0x100e10) << 12;
• priv->base.get = nv50_ram_get;
• priv->base.put = nv50_ram_put; return 0;

}

-- 2.2.0

From 85ca5098d65f31362024bfd96c0b04b571d10cfb Mon Sep 17 00:00:00 2001 From: Ben Skeggs bskeggs@redhat.com Date: Thu, 11 Dec 2014 13:21:24 +1000 Subject: [PATCH 2/3] fb/ram/mcp77: use carveout reg to determine size

Signed-off-by: Ben Skeggs bskeggs@redhat.com

nvkm/subdev/fb/ramnvaa.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/nvkm/subdev/fb/ramnvaa.c b/nvkm/subdev/fb/ramnvaa.c index 8ee3d37..f56ee55 100644 --- a/nvkm/subdev/fb/ramnvaa.c +++ b/nvkm/subdev/fb/ramnvaa.c @@ -44,16 +44,15 @@ nvaa_ram_ctor(struct nouveau_object *parent, struct nouveau_object *engine, if (ret) return ret;

• priv->base.size = nv_rd32(pfb, 0x10020c);
• priv->base.size = (priv->base.size & 0xffffff00) | ((priv->base.size & 0x000000ff) << 32);

• priv->base.type = NV_MEM_TYPE_STOLEN;

• priv->base.stolen = (u64)nv_rd32(pfb, 0x100e10) << 12;

• priv->base.size = (u64)nv_rd32(pfb, 0x100e14) << 12;

  ret = nouveau_mm_init(&pfb->vram, rsvd_head, (priv->base.size >> 12) - (rsvd_head + rsvd_tail), 1); if (ret) return ret;

• priv->base.type = NV_MEM_TYPE_STOLEN;
• priv->base.stolen = (u64)nv_rd32(pfb, 0x100e10) << 12; priv->base.get = nv50_ram_get; priv->base.put = nv50_ram_put; return 0;

-- 2.2.0

From 3b2a91eb59d308bed98637d4bcdfe29a2c9434ad Mon Sep 17 00:00:00 2001 From: Ben Skeggs bskeggs@redhat.com Date: Thu, 11 Dec 2014 13:39:18 +1000 Subject: [PATCH 3/3] EXAMPLEfb/ram/mcp77: niso poller setup

---

nvkm/subdev/fb/ramnvaa.c | 36 ++++++++++++++++++++++++++++++++---- 1 file changed, 32 insertions(+), 4 deletions(-)

diff --git a/nvkm/subdev/fb/ramnvaa.c b/nvkm/subdev/fb/ramnvaa.c index f56ee55..9cac747 100644 --- a/nvkm/subdev/fb/ramnvaa.c +++ b/nvkm/subdev/fb/ramnvaa.c @@ -26,6 +26,7 @@

struct nvaa_ram_priv { struct nouveau_ram base;

• u64 poller_base;

};

static int @@ -33,8 +34,8 @@ nvaa_ram_ctor(struct nouveau_object *parent, struct nouveau_object *engine, struct nouveau_oclass *oclass, void *data, u32 datasize, struct nouveau_object **pobject) {

• const u32 rsvd_head = ( 256 * 1024) >> 12; /* vga memory */
• const u32 rsvd_tail = (1024 * 1024) >> 12; /* vbios etc */

• u32 rsvd_head = ( 256 * 1024); /* vga memory */
• u32 rsvd_tail = (1024 * 1024); /* vbios etc */ struct nouveau_fb *pfb = nouveau_fb(parent); struct nvaa_ram_priv *priv; int ret;

@@ -47,9 +48,16 @@ nvaa_ram_ctor(struct nouveau_object *parent, struct nouveau_object *engine, priv->base.type = NV_MEM_TYPE_STOLEN; priv->base.stolen = (u64)nv_rd32(pfb, 0x100e10) << 12; priv->base.size = (u64)nv_rd32(pfb, 0x100e14) << 12;

• // 0x10000000

• ret = nouveau_mm_init(&pfb->vram, rsvd_head, (priv->base.size >> 12) -

• 	      (rsvd_head + rsvd_tail), 1);
  

• rsvd_tail += 0x1000;
• priv->poller_base = priv->base.size - rsvd_tail;
• // 0x10000000 - 0x101000
• // 0x0feff000 << offset from start of carveout
• ret = nouveau_mm_init(&pfb->vram, rsvd_head >> 12,

• 	      (priv->base.size >> 12) -
  

• 	      (rsvd_head + rsvd_tail) >> 12, 1);
  

  if (ret) return ret;

@@ -58,6 +66,26 @@ nvaa_ram_ctor(struct nouveau_object *parent, struct nouveau_object *engine, return 0; }

+static int +nvaa_ram_init(struct nouveau_object *object) +{

• struct nvaa_ram_priv *priv = (void *)object;
• int ret;
• ret = nouveau_ram_init(&priv->base);
• if (ret)

• return ret;
  

• dniso = ((priv->base.size - (priv->poller_base + 0x00)) >> 5) - 1;
• // 0x807f
• hostnb = ((priv->base.size - (priv->poller_base + 0x20)) >> 5) - 1;
• // 0x807e
• flush = ((priv->base.size - (priv->poller_base + 0x40)) >> 5) - 1;
• // 0x807d
• return 0;

+}

struct nouveau_oclass nvaa_ram_oclass = { .ofuncs = &(struct nouveau_ofuncs) { -- 2.2.0