drm/panthor: Fix IO-page mmap() for 32-bit userspace on 64-bit kernel
authorBoris Brezillon <boris.brezillon@collabora.com>
Tue, 26 Mar 2024 11:12:03 +0000 (12:12 +0100)
committerBoris Brezillon <boris.brezillon@collabora.com>
Tue, 2 Apr 2024 07:33:46 +0000 (09:33 +0200)
When mapping an IO region, the pseudo-file offset is dependent on the
userspace architecture. panthor_device_mmio_offset() abstracts that
away for us by turning a userspace MMIO offset into its kernel
equivalent, but we were not updating vm_area_struct::vm_pgoff
accordingly, leading us to attach the MMIO region to the wrong file
offset.

This has implications when we start mixing 64 bit and 32 bit apps, but
that's only really a problem when we start having more that 2^43 bytes of
memory allocated, which is very unlikely to happen.

What's more problematic is the fact this turns our
unmap_mapping_range(DRM_PANTHOR_USER_MMIO_OFFSET) calls, which are
supposed to kill the MMIO mapping when entering suspend, into NOPs.
Which means we either keep the dummy flush_id mapping active at all
times, or we risk a BUS_FAULT if the MMIO region was mapped, and the
GPU is suspended after that.

Solve that by patching vm_pgoff early in panthor_mmap(). With
this in place, we no longer need the panthor_device_mmio_offset()
helper.

v3:
- No changes

v2:
- Kill panthor_device_mmio_offset()

Fixes: 5fe909cae118 ("drm/panthor: Add the device logical block")
Reported-by: Adrián Larumbe <adrian.larumbe@collabora.com>
Reported-by: Lukas F. Hartmann <lukas@mntmn.com>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/10835
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Reviewed-by: Liviu Dudau <liviu.dudau@arm.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240326111205.510019-1-boris.brezillon@collabora.com
drivers/gpu/drm/panthor/panthor_device.c
drivers/gpu/drm/panthor/panthor_device.h
drivers/gpu/drm/panthor/panthor_drv.c

index 8b00f342d1b16569e45d338854b5a9484ea4ce27..75276cbeba208da8156ef8b31eec247bd855342e 100644 (file)
@@ -348,7 +348,7 @@ static vm_fault_t panthor_mmio_vm_fault(struct vm_fault *vmf)
 {
        struct vm_area_struct *vma = vmf->vma;
        struct panthor_device *ptdev = vma->vm_private_data;
-       u64 id = (u64)vma->vm_pgoff << PAGE_SHIFT;
+       u64 offset = (u64)vma->vm_pgoff << PAGE_SHIFT;
        unsigned long pfn;
        pgprot_t pgprot;
        vm_fault_t ret;
@@ -361,7 +361,7 @@ static vm_fault_t panthor_mmio_vm_fault(struct vm_fault *vmf)
        mutex_lock(&ptdev->pm.mmio_lock);
        active = atomic_read(&ptdev->pm.state) == PANTHOR_DEVICE_PM_STATE_ACTIVE;
 
-       switch (panthor_device_mmio_offset(id)) {
+       switch (offset) {
        case DRM_PANTHOR_USER_FLUSH_ID_MMIO_OFFSET:
                if (active)
                        pfn = __phys_to_pfn(ptdev->phys_addr + CSF_GPU_LATEST_FLUSH_ID);
@@ -392,9 +392,9 @@ static const struct vm_operations_struct panthor_mmio_vm_ops = {
 
 int panthor_device_mmap_io(struct panthor_device *ptdev, struct vm_area_struct *vma)
 {
-       u64 id = (u64)vma->vm_pgoff << PAGE_SHIFT;
+       u64 offset = (u64)vma->vm_pgoff << PAGE_SHIFT;
 
-       switch (panthor_device_mmio_offset(id)) {
+       switch (offset) {
        case DRM_PANTHOR_USER_FLUSH_ID_MMIO_OFFSET:
                if (vma->vm_end - vma->vm_start != PAGE_SIZE ||
                    (vma->vm_flags & (VM_WRITE | VM_EXEC)))
index c84c27dcc92c7e33502e8147415b00cb8a298c58..b68e8821d219cc5fef257354b239749c3b10ea46 100644 (file)
@@ -365,30 +365,6 @@ static int panthor_request_ ## __name ## _irq(struct panthor_device *ptdev,                        \
                                         pirq);                                                 \
 }
 
-/**
- * panthor_device_mmio_offset() - Turn a user MMIO offset into a kernel one
- * @offset: Offset to convert.
- *
- * With 32-bit systems being limited by the 32-bit representation of mmap2's
- * pgoffset field, we need to make the MMIO offset arch specific. This function
- * converts a user MMIO offset into something the kernel driver understands.
- *
- * If the kernel and userspace architecture match, the offset is unchanged. If
- * the kernel is 64-bit and userspace is 32-bit, the offset is adjusted to match
- * 64-bit offsets. 32-bit kernel with 64-bit userspace is impossible.
- *
- * Return: Adjusted offset.
- */
-static inline u64 panthor_device_mmio_offset(u64 offset)
-{
-#ifdef CONFIG_ARM64
-       if (test_tsk_thread_flag(current, TIF_32BIT))
-               offset += DRM_PANTHOR_USER_MMIO_OFFSET_64BIT - DRM_PANTHOR_USER_MMIO_OFFSET_32BIT;
-#endif
-
-       return offset;
-}
-
 extern struct workqueue_struct *panthor_cleanup_wq;
 
 #endif
index 11b3ccd58f85ffc5226667b525a0a69ec4fc7593..730dd0c69cb813c0c7f0825474baa3fd8507a7a9 100644 (file)
@@ -1327,7 +1327,22 @@ static int panthor_mmap(struct file *filp, struct vm_area_struct *vma)
        if (!drm_dev_enter(file->minor->dev, &cookie))
                return -ENODEV;
 
-       if (panthor_device_mmio_offset(offset) >= DRM_PANTHOR_USER_MMIO_OFFSET)
+#ifdef CONFIG_ARM64
+       /*
+        * With 32-bit systems being limited by the 32-bit representation of
+        * mmap2's pgoffset field, we need to make the MMIO offset arch
+        * specific. This converts a user MMIO offset into something the kernel
+        * driver understands.
+        */
+       if (test_tsk_thread_flag(current, TIF_32BIT) &&
+           offset >= DRM_PANTHOR_USER_MMIO_OFFSET_32BIT) {
+               offset += DRM_PANTHOR_USER_MMIO_OFFSET_64BIT -
+                         DRM_PANTHOR_USER_MMIO_OFFSET_32BIT;
+               vma->vm_pgoff = offset >> PAGE_SHIFT;
+       }
+#endif
+
+       if (offset >= DRM_PANTHOR_USER_MMIO_OFFSET)
                ret = panthor_device_mmap_io(ptdev, vma);
        else
                ret = drm_gem_mmap(filp, vma);