NVIDIA 450.57/450.66 driver patch for Linux 5.9-rc1

nvidia-fix-linux-5.9.patch

Tentative patch for NVIDIA driver compatibility for Linux 5.9-rc1

NOTE: This allows the module to build and mostly work, but the nvidia-uvm
      module won't load due to the new patch from Christoph Hellwig that
      prevents modules to simultaneously link to GPL and proprietary symbols
      (torvalds/linux commit 0fd9cc6b0c72245375520ffc8d97ce5857b63b94)
      It seems to be non-trivial to fix this issue "for real" (workarounds not included)

---

diff --git a/common/inc/nv-mm.h b/common/inc/nv-mm.h
index b04ff67..fc9f67e 100644
--- a/common/inc/nv-mm.h
+++ b/common/inc/nv-mm.h
@@ -23,6 +23,7 @@
 #ifndef __NV_MM_H__
 #define __NV_MM_H__
 
+#include <linux/version.h>
 #include "conftest.h"
 
 #if !defined(NV_VM_FAULT_T_IS_PRESENT)
@@ -139,7 +140,16 @@ typedef int vm_fault_t;
             if (force)
                 flags |= FOLL_FORCE;
 
-        #if defined(NV_GET_USER_PAGES_REMOTE_HAS_LOCKED_ARG)
+/* torvalds/linux commit 64019a2e467a288a16b65ab55ddcbf58c1b00187
+   "After the cleanup of page fault accounting, gup does not need to pass
+   task_struct around any more.  Remove that parameter in the whole gup
+   stack." */
+        #if (LINUX_VERSION_CODE >= KERNEL_VERSION(5, 9, 0))
+
+               return get_user_pages_remote(mm, start, nr_pages, flags,
+                                            pages, vmas, NULL);
+
+        #elif defined(NV_GET_USER_PAGES_REMOTE_HAS_LOCKED_ARG)
 
                return get_user_pages_remote(tsk, mm, start, nr_pages, flags,
                                             pages, vmas, NULL);
diff --git a/nvidia-drm/nvidia-drm-drv.c b/nvidia-drm/nvidia-drm-drv.c
index c6ba478..fd6333e 100644
--- a/nvidia-drm/nvidia-drm-drv.c
+++ b/nvidia-drm/nvidia-drm-drv.c
@@ -20,6 +20,7 @@
  * DEALINGS IN THE SOFTWARE.
  */
 
+#include <linux/version.h>
 #include "nvidia-drm-conftest.h" /* NV_DRM_AVAILABLE and NV_DRM_DRM_GEM_H_PRESENT */
 
 #include "nvidia-drm-priv.h"
@@ -473,16 +474,36 @@ static void nv_drm_unload(struct drm_device *dev)
 
 #if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
 
+/* torvalds/linux commit 907f53200f982cd377e75abbc1b18ed624d7ae66
+   "The function always returns zero (success). Ideally we'll remove it all
+   together - although that's requires a little more work.
+
+   For now, we can drop the return type and simplify the drm core code
+   surrounding it." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
 static int nv_drm_master_set(struct drm_device *dev,
+#else
+static void nv_drm_master_set(struct drm_device *dev,
+#endif
                              struct drm_file *file_priv, bool from_open)
 {
     struct nv_drm_device *nv_dev = to_nv_device(dev);
 
+/* torvalds/linux commit 907f53200f982cd377e75abbc1b18ed624d7ae66
+   "The function always returns zero (success). Ideally we'll remove it all
+   together - although that's requires a little more work.
+
+   For now, we can drop the return type and simplify the drm core code
+   surrounding it." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
     if (!nvKms->grabOwnership(nv_dev->pDevice)) {
         return -EINVAL;
     }
 
     return 0;
+#else
+    nvKms->grabOwnership(nv_dev->pDevice);
+#endif
 }
 
 #if defined(NV_DRM_MASTER_DROP_HAS_FROM_RELEASE_ARG)
@@ -719,7 +740,16 @@ static struct drm_driver nv_drm_driver = {
 #endif
                                DRIVER_GEM  | DRIVER_RENDER,
 
+/* torvalds/linux commit 0a19b068acc47d05212f03e494381926dc0381e2
+   Should work since NVIDIA does not use struct_mutex?
+   "- More struct_mutex removal for !legacy drivers:
+   - Remove gem_free_object()
+   - Removal of drm_gem_object_put_unlocked()." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
     .gem_free_object        = nv_drm_gem_free,
+#else
+    .gem_free_object_unlocked = nv_drm_gem_free,
+#endif
 
     .ioctls                 = nv_drm_ioctls,
     .num_ioctls             = ARRAY_SIZE(nv_drm_ioctls),
diff --git a/nvidia-drm/nvidia-drm-gem.c b/nvidia-drm/nvidia-drm-gem.c
index ad4bca4..e63d352 100644
--- a/nvidia-drm/nvidia-drm-gem.c
+++ b/nvidia-drm/nvidia-drm-gem.c
@@ -20,6 +20,7 @@
  * DEALINGS IN THE SOFTWARE.
  */
 
+#include <linux/version.h>
 #include "nvidia-drm-conftest.h"
 
 #if defined(NV_DRM_AVAILABLE)
@@ -46,10 +47,17 @@
 
 void nv_drm_gem_free(struct drm_gem_object *gem)
 {
-    struct drm_device *dev = gem->dev;
     struct nv_drm_gem_object *nv_gem = to_nv_gem_object(gem);
 
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
+/* torvalds/linux commit 0a19b068acc47d05212f03e494381926dc0381e2
+   Should work since NVIDIA does not use struct_mutex?
+   "- More struct_mutex removal for !legacy drivers:
+   - Remove gem_free_object()
+   - Removal of drm_gem_object_put_unlocked()." */
+    struct drm_device *dev = gem->dev;
     WARN_ON(!mutex_is_locked(&dev->struct_mutex));
+#endif
 
     /* Cleanup core gem object */
 
diff --git a/nvidia-drm/nvidia-drm-gem.h b/nvidia-drm/nvidia-drm-gem.h
index 4e7ffd3..b5dd705 100644
--- a/nvidia-drm/nvidia-drm-gem.h
+++ b/nvidia-drm/nvidia-drm-gem.h
@@ -23,6 +23,7 @@
 #ifndef __NVIDIA_DRM_GEM_H__
 #define __NVIDIA_DRM_GEM_H__
 
+#include <linux/version.h>
 #include "nvidia-drm-conftest.h"
 
 #if defined(NV_DRM_AVAILABLE)
@@ -96,7 +97,16 @@ static inline void
 nv_drm_gem_object_unreference_unlocked(struct nv_drm_gem_object *nv_gem)
 {
 #if defined(NV_DRM_GEM_OBJECT_GET_PRESENT)
+/* torvalds/linux commit 0a19b068acc47d05212f03e494381926dc0381e2
+   Should work since NVIDIA does not use struct_mutex?
+   "- More struct_mutex removal for !legacy drivers:
+   - Remove gem_free_object()
+   - Removal of drm_gem_object_put_unlocked()." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
     drm_gem_object_put_unlocked(&nv_gem->base);
+#else
+    drm_gem_object_put(&nv_gem->base);
+#endif
 #else
     drm_gem_object_unreference_unlocked(&nv_gem->base);
 #endif
diff --git a/nvidia-drm/nvidia-drm-utils.c b/nvidia-drm/nvidia-drm-utils.c
index da485c6..a5d7a2a 100644
--- a/nvidia-drm/nvidia-drm-utils.c
+++ b/nvidia-drm/nvidia-drm-utils.c
@@ -20,6 +20,7 @@
  * DEALINGS IN THE SOFTWARE.
  */
 
+#include <linux/version.h>
 #include "nvidia-drm-conftest.h" /* NV_DRM_ATOMIC_MODESET_AVAILABLE */
 
 #if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
@@ -103,7 +104,13 @@ void
 nvkms_display_mode_to_drm_mode(const struct NvKmsKapiDisplayMode *displayMode,
                                struct drm_display_mode *mode)
 {
+/* torvalds/linux commit 0425662fdf05665235e768e2fbcb4ced12432b43
+   "Get rid of mode->vrefresh and just calculate it on demand. Saves
+   a bit of space and avoids the cached value getting out of sync
+   with reality." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
     mode->vrefresh    = (displayMode->timings.refreshRate + 500) / 1000; /* In Hz */
+#endif
 
     mode->clock       = (displayMode->timings.pixelClockHz + 500) / 1000; /* In Hz */
 
@@ -159,7 +166,15 @@ nvkms_display_mode_to_drm_mode(const struct NvKmsKapiDisplayMode *displayMode,
 void drm_mode_to_nvkms_display_mode(const struct drm_display_mode *src,
                                     struct NvKmsKapiDisplayMode *dst)
 {
+/* torvalds/linux commit 0425662fdf05665235e768e2fbcb4ced12432b43
+   "Get rid of mode->vrefresh and just calculate it on demand. Saves
+   a bit of space and avoids the cached value getting out of sync
+   with reality." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
     dst->timings.refreshRate  = src->vrefresh * 1000;
+#else
+    dst->timings.refreshRate  = drm_mode_vrefresh(src) * 1000;
+#endif
 
     dst->timings.pixelClockHz = src->clock * 1000; /* In Hz */
 
diff --git a/nvidia-uvm/uvm8_tools.c b/nvidia-uvm/uvm8_tools.c
index 620e1cc..1aee896 100644
--- a/nvidia-uvm/uvm8_tools.c
+++ b/nvidia-uvm/uvm8_tools.c
@@ -20,6 +20,7 @@
     DEALINGS IN THE SOFTWARE.
 
 *******************************************************************************/
+#include <linux/version.h>
 #include "uvm_common.h"
 #include "uvm_ioctl.h"
 #include "uvm8_gpu.h"
@@ -192,7 +193,15 @@ static uvm_tools_event_tracker_t *tools_event_tracker(struct file *filp)
 {
     long event_tracker = atomic_long_read((atomic_long_t *)&filp->private_data);
 
+/* torvalds/linux commit 145ff1ec090dce9beb5a9590b5dc288e7bb2e65d
+   "Removal of the tremendously unpopular read_barrier_depends()
+   barrier, which is a NOP on all architectures apart from Alpha, in
+   favour of allowing architectures to override READ_ONCE() and do
+   whatever dance they need to do to ensure address dependencies
+   provide LOAD -> LOAD/STORE ordering." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
     smp_read_barrier_depends();
+#endif
     return (uvm_tools_event_tracker_t *)event_tracker;
 }
 
diff --git a/nvidia-uvm/uvm8_va_range.h b/nvidia-uvm/uvm8_va_range.h
index 69cb0a0..edc21cc 100644
--- a/nvidia-uvm/uvm8_va_range.h
+++ b/nvidia-uvm/uvm8_va_range.h
@@ -24,6 +24,7 @@
 #ifndef __UVM8_VA_RANGE_H__
 #define __UVM8_VA_RANGE_H__
 
+#include <linux/version.h>
 #include "uvm_linux.h"
 #include "nv-kref.h"
 #include "uvm_common.h"
@@ -755,7 +756,15 @@ static uvm_va_block_t *uvm_va_range_block(uvm_va_range_t *va_range, size_t index
     // make sure that any initialization of this block by the creating thread is
     // visible to later accesses in this thread, which requires a data
     // dependency barrier.
+/* torvalds/linux commit 145ff1ec090dce9beb5a9590b5dc288e7bb2e65d
+   "Removal of the tremendously unpopular read_barrier_depends()
+   barrier, which is a NOP on all architectures apart from Alpha, in
+   favour of allowing architectures to override READ_ONCE() and do
+   whatever dance they need to do to ensure address dependencies
+   provide LOAD -> LOAD/STORE ordering." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
     smp_read_barrier_depends();
+#endif
     return block;
 }
 
diff --git a/nvidia/nv-pci.c b/nvidia/nv-pci.c
index 9a8a715..1e08a7c 100644
--- a/nvidia/nv-pci.c
+++ b/nvidia/nv-pci.c
@@ -8,6 +8,7 @@
  * _NVRM_COPYRIGHT_END_
  */
 
+#include <linux/version.h>
 #include "nv-pci-table.h"
 #include "nv-pci.h"
 #include "nv-ibmnpu.h"
@@ -424,7 +425,15 @@ nv_pci_probe
 
 #if defined(CONFIG_VGA_ARB) && !defined(NVCPU_PPC64LE)
 #if defined(VGA_DEFAULT_DEVICE)
+/* torvalds/linux commit f369bc3f9096f5d355e8b80540bc30ac9a602912
+   Does this make sense? What happened when vga_tryget failed (ret -EBUSY)? Should be interruptible?
+   "vgaarb: mark vga_tryget static
+   This symbols isn't used anywhere outside of vgaarb.c." */
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(5, 9, 0))
     vga_tryget(VGA_DEFAULT_DEVICE, VGA_RSRC_LEGACY_MASK);
+#else
+    vga_get(VGA_DEFAULT_DEVICE, VGA_RSRC_LEGACY_MASK, 0);
+#endif
 #endif
     vga_set_legacy_decoding(pci_dev, VGA_RSRC_NONE);
 #endif

PS: This is obsolete since NVIDIA has released new drivers that work on 5.9 (except the nvidia-uvm problem, which remains).