Please ignore this patch, sent by mistake to wrong mailing list. Thanks, Tejas > -----Original Message----- > From: Upadhyay, Tejas <tejas.upadhyay@xxxxxxxxx> > Sent: Wednesday, October 9, 2024 3:37 PM > To: intel-gfx@xxxxxxxxxxxxxxxxxxxxx > Cc: Upadhyay, Tejas <tejas.upadhyay@xxxxxxxxx> > Subject: [PATCH] drm/xe/mmap: Add mmap support for PCI memory barrier > > In order to avoid having userspace to use MI_MEM_FENCE, we are adding a > mechanism for userspace to generate a PCI memory barrier with low > overhead (avoiding IOCTL call). > > This is implemented by memory-mapping a page as uncached that is backed > by MMIO on the dGPU and thus allowing userspace to do memory write to > the page without invoking an IOCTL. > We are selecting the MMIO so that it is not accessible from the PCI bus so that > the MMIO writes themselves are ignored, but the PCI memory barrier will still > take action as the MMIO filtering will happen after the memory barrier effect. > > When we detect special defined offset in mmap(), We are mapping 4K page > which contains the last of page of doorbell MMIO range to userspace for > same purpose. > > Note: Test coverage for this is added by IGT > https://patchwork.freedesktop.org/patch/618931/ here. > > Signed-off-by: Tejas Upadhyay <tejas.upadhyay@xxxxxxxxx> > --- > drivers/gpu/drm/xe/xe_device.c | 68 > +++++++++++++++++++++++++++++++++- > 1 file changed, 67 insertions(+), 1 deletion(-) > > diff --git a/drivers/gpu/drm/xe/xe_device.c b/drivers/gpu/drm/xe/xe_device.c > index cd241a8e1838..c97a4d1f0a98 100644 > --- a/drivers/gpu/drm/xe/xe_device.c > +++ b/drivers/gpu/drm/xe/xe_device.c > @@ -239,12 +239,78 @@ static long xe_drm_compat_ioctl(struct file *file, > unsigned int cmd, unsigned lo #define xe_drm_compat_ioctl NULL #endif > > +static void barrier_open(struct vm_area_struct *vma) { > + drm_dev_get(vma->vm_private_data); > +} > + > +static void barrier_close(struct vm_area_struct *vma) { > + drm_dev_put(vma->vm_private_data); > +} > + > +static const struct vm_operations_struct vm_ops_barrier = { > + .open = barrier_open, > + .close = barrier_close, > +}; > + > +static int xe_pci_barrier_mmap(struct file *filp, > + struct vm_area_struct *vma) > +{ > + struct drm_file *priv = filp->private_data; > + struct drm_device *dev = priv->minor->dev; > + unsigned long pfn; > + pgprot_t prot; > + > + if (vma->vm_end - vma->vm_start > PAGE_SIZE) > + return -EINVAL; > + > + if (is_cow_mapping(vma->vm_flags)) > + return -EINVAL; > + > + if (vma->vm_flags & (VM_READ | VM_EXEC)) > + return -EINVAL; > + > + vm_flags_clear(vma, VM_MAYREAD | VM_MAYEXEC); > + vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | > VM_DONTDUMP | VM_IO); > + > + prot = vm_get_page_prot(vma->vm_flags); #define > LAST_DB_PAGE_OFFSET > +0x7ff001 > + pfn = PHYS_PFN(pci_resource_start(to_pci_dev(dev->dev), 0) + > + LAST_DB_PAGE_OFFSET); > + if (vmf_insert_pfn_prot(vma, vma->vm_start, pfn, > + pgprot_noncached(prot)) != > VM_FAULT_NOPAGE) > + return -EFAULT; > + > + vma->vm_ops = &vm_ops_barrier; > + vma->vm_private_data = dev; > + drm_dev_get(vma->vm_private_data); > + return 0; > +} > + > +static int xe_mmap(struct file *filp, struct vm_area_struct *vma) { > + struct drm_file *priv = filp->private_data; > + struct drm_device *dev = priv->minor->dev; > + > + if (drm_dev_is_unplugged(dev)) > + return -ENODEV; > + > +#define XE_PCI_BARRIER_MMAP_OFFSET (0x50 << PAGE_SHIFT) > + switch (vma->vm_pgoff) { > + case XE_PCI_BARRIER_MMAP_OFFSET >> PAGE_SHIFT: > + return xe_pci_barrier_mmap(filp, vma); > + } > + > + return drm_gem_mmap(filp, vma); > +} > + > static const struct file_operations xe_driver_fops = { > .owner = THIS_MODULE, > .open = drm_open, > .release = drm_release_noglobal, > .unlocked_ioctl = xe_drm_ioctl, > - .mmap = drm_gem_mmap, > + .mmap = xe_mmap, > .poll = drm_poll, > .read = drm_read, > .compat_ioctl = xe_drm_compat_ioctl, > -- > 2.34.1
