Intel Arc A770: Kernel panic on kldload i915kms.ko

[kenrap]

Describe the bug
From using a drm-kmod build from efd9167, the i915kms driver kernel panics when using an (Acer Predator BiFrost) Intel Arc A770 graphics card. And the kernel panic still persists even when using an Intel onboard GPU with the same graphics card installed.

FreeBSD version

FreeBSD freebsd 15.0-CURRENT FreeBSD 15.0-CURRENT #0 main-n271909-28294dc92476: Fri Aug 30 08:28:17 PDT 2024     root@freebsd:/usr/obj/usr/src/amd64.amd64/sys/GENERIC-NODEBUG amd64 1500023 1500023

DRM KMOD version
My own "custom derived" graphics/drm-66-kmod port with GH_TAGNAME pointing to efd9167.

Also git-clone(1) from linux-firmware and copied all of the i915/dg2_* firmware bins to /boot/modules and renamed them appropriately to match the filename style there.

To Reproduce
Boot into the system with either i915kms using kld_list inside /etc/rc.conf or kldload it manually.

Additional context
core.txt.0 dump

[kenrap]

My bad, I didn't see about the "Intel DG2 GUC/HUC support" not being implemented yet from PR #283.

Closing therefore.

I'm reopening this to have it serve as a milestone issue.

I figured why have it be closed anyway since the issue is still valid and my reporting could be useful?

[wulf7]

dg2_dmc_ver2_08.bin: could not load binary firmware /boot/firmware/dg2_dmc_ver2_08.bin either
i915/dg2_dmc_ver2_08.bin: could not load binary firmware /boot/firmware/i915/dg2_dmc_ver2_08.bin either
i915_dg2_dmc_ver2_08.bin: could not load binary firmware /boot/firmware/i915_dg2_dmc_ver2_08.bin either
i915_dg2_dmc_ver2_08_bin: could not load binary firmware /boot/firmware/i915_dg2_dmc_ver2_08_bin either
i915_dg2_dmc_ver2_08_bin: could not load binary firmware /boot/firmware/i915_dg2_dmc_ver2_08_bin either
drmn0: could not load firmware image 'i915/dg2_dmc_ver2_08.bin'
drmn0: [drm] Failed to load DMC firmware i915/dg2_dmc_ver2_08.bin. Disabling runtime power management.
drmn0: [drm] Run pkg install gpu-firmware-kmod to install it

You may start with adding dg2_dmc_ver2_08.bin to firmwares

[wulf7]

But I doubt that it will help

[kenrap]

I'll try that out and report back.

I also updated my bug description to be more specific.

[kenrap]

But I doubt that it will help

And you're right, it didn't.

After trying a couple more ideas, I spent a good amount of time re-learning how to create a new core dump of the kernel panic with the firmware(s) loaded. Sorry for the delay.

core.txt.1 dump

[wulf7]

After taking a look at the code around faulted line, I have got an impression that it can happen due to missing vmap_pfn() implementation.
It is rather easy to check. Just with replacing of return NULL; line in i915_gem_object_map_pfn() function located in drivers/gpu/drm/i915/gem/i915_gem_pages.c file of drm-kmod with panic("oops"); or return ERR_PTR(-ENOSUP);

[kenrap]

It seems you're correct about the missing vmap_pfn() implementation. It triggered an "oops" panic by using the panic("oops"); based on your instructions.

What I did was created this custom patch to put into my /usr/ports/graphics/drm-66-kmod/files:

--- drivers/gpu/drm/i915/gem/i915_gem_pages.c.orig
+++ drivers/gpu/drm/i915/gem/i915_gem_pages.c
@@ -329,7 +329,7 @@ static void *i915_gem_object_map_pfn(struct drm_i915_gem_object *obj,
 {
 #ifdef __FreeBSD__
        // BSDFIXME: Need vmap_pfn() implementation.
-       return NULL;
+       panic("oops");
 #else
        resource_size_t iomap = obj->mm.region->iomap.base -
                obj->mm.region->region.start;

And rebuilt and reinstalled the package of my derived port. Then I did my usual testing and grabbed a new core dump which shows the "oops" panic. Yay! \o/

core.txt.2

[wulf7]

You may try following patches (only compile-tested). It is just quick conversion of vmap() implementation to vmap_pfn() through replacement struct page with page frame number
FreeBSD:

Incomplete patch deleted. See patch three messages below

and drm-kmod:

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_pages.c b/drivers/gpu/drm/i915/gem/i915_gem_pages.c
index 931e7f46733..0ba955611df 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_pages.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_pages.c
@@ -327,10 +327,6 @@ static void *i915_gem_object_map_page(struct drm_i915_gem_object *obj,
 static void *i915_gem_object_map_pfn(struct drm_i915_gem_object *obj,
 				     enum i915_map_type type)
 {
-#ifdef __FreeBSD__
-	// BSDFIXME: Need vmap_pfn() implementation.
-	return NULL;
-#else
 	resource_size_t iomap = obj->mm.region->iomap.base -
 		obj->mm.region->region.start;
 	unsigned long n_pfn = obj->base.size >> PAGE_SHIFT;
@@ -356,7 +352,6 @@ static void *i915_gem_object_map_pfn(struct drm_i915_gem_object *obj,
 		kvfree(pfns);
 
 	return vaddr ?: ERR_PTR(-ENOMEM);
-#endif
 }
 
 /* get, pin, and map the pages of the object into kernel space */

[kenrap]

@wulf7 core.txt.0

[kenrap]

I guess I should note that after doing a lot of experimentation, there was one time I was "lucky" enough at random to get i915kms.ko to load with my Intel Arc but I was noticing this warning at the time:

drmn0: [drm] Reducing the compressed framebuffer size. This may lead to less power savings than a non-reduced-size. Try to increase stolen memory size if available in BIOS.

But that only happened once and it wasn't reproducible.

Also, I don't know if I need to get another coredump by using the following in /boot/loader.conf:

hw.i915kms.enable_guc="2"
compat.linuxkpi.i915_disable_power_well="0"

because I changed the first tunable to 1 and commented out the second one. I needed both as above to allow myself to load i915kms on my integrated Intel GPU and get back into X11.

But right now, I need some rest, because I was doing on like +20 kernel-panic reboots of experiments.

[wulf7]

Next version of FreeBSD patch:

diff --git a/sys/compat/linuxkpi/common/include/linux/vmalloc.h b/sys/compat/linuxkpi/common/include/linux/vmalloc.h
index 00650a2df9b..30f7e0e6297 100644
--- a/sys/compat/linuxkpi/common/include/linux/vmalloc.h
+++ b/sys/compat/linuxkpi/common/include/linux/vmalloc.h
@@ -35,8 +35,11 @@
 #define	VM_MAP		0x0000
 #define	PAGE_KERNEL	0x0000
 
+#define	vmap_pfn(...)	lkpi_vmap_pfn(__VA_ARGS__)
+
 void *vmap(struct page **pages, unsigned int count, unsigned long flags,
     int prot);
+void *lkpi_vmap_pfn(unsigned long *pfns, unsigned int count, int prot);
 void vunmap(void *addr);
 
 #endif	/* _LINUXKPI_LINUX_VMALLOC_H_ */
diff --git a/sys/compat/linuxkpi/common/src/linux_compat.c b/sys/compat/linuxkpi/common/src/linux_compat.c
index 81d24603d1d..bce3af61516 100644
--- a/sys/compat/linuxkpi/common/src/linux_compat.c
+++ b/sys/compat/linuxkpi/common/src/linux_compat.c
@@ -60,6 +60,9 @@
 #include <vm/vm_page.h>
 #include <vm/vm_pager.h>
 
+#include <vm/uma.h>
+#include <vm/uma_int.h>
+
 #include <machine/stdarg.h>
 
 #if defined(__i386__) || defined(__amd64__)
@@ -1804,6 +1807,24 @@ vmmap_remove(void *addr)
 	return (vmmap);
 }
 
+int
+is_vmalloc_addr(const void *addr)
+{
+	struct vmmap *vmmap;
+	uintptr_t p = (uintptr_t)addr;
+
+	mtx_lock(&vmmaplock);
+	LIST_FOREACH(vmmap, &vmmaphead[VM_HASH(addr)], vm_next)
+		if (p >= trunc_page(vmmap->vm_addr) &&
+		    p < round_page((char *)vmmap->vm_addr + vmmap->vm_size))
+			break;
+	mtx_unlock(&vmmaplock);
+	if (vmmap != NULL)
+		return(1);
+
+	return (vtoslab((vm_offset_t)addr & ~UMA_SLAB_MASK) != NULL);
+}
+
 #if defined(__i386__) || defined(__amd64__) || defined(__powerpc__) || defined(__aarch64__) || defined(__riscv)
 void *
 _ioremap_attr(vm_paddr_t phys_addr, unsigned long size, int attr)
@@ -1849,6 +1870,58 @@ vmap(struct page **pages, unsigned int count, unsigned long flags, int prot)
 	return ((void *)off);
 }
 
+#ifdef __amd64__
+static void
+_lkpi_pmap_qenter_pfn(vm_offset_t sva, vm_pindex_t *pi, int count,
+    vm_memattr_t mode)
+{
+	pt_entry_t *endpte, oldpte, pa, *pte;
+	vm_pindex_t p;
+	int cache_bits;
+	pt_entry_t pg_g;
+
+	pg_g = pti ? 0 : X86_PG_G;
+	oldpte = 0;
+	pte = vtopte(sva);
+	endpte = pte + count;
+	cache_bits = pmap_cache_bits(kernel_pmap, mode, false);
+	while (pte < endpte) {
+		p = *pi++;
+		pa = IDX_TO_OFF(p) | cache_bits;
+		if ((*pte & (PG_FRAME | X86_PG_PTE_CACHE)) != pa) {
+			oldpte |= *pte;
+			pte_store(pte, pa | pg_g | pg_nx | X86_PG_A |
+			    X86_PG_M | X86_PG_RW | X86_PG_V);
+		}
+		pte++;
+	}
+	if (__predict_false((oldpte & X86_PG_V) != 0))
+		pmap_invalidate_range(kernel_pmap, sva, sva + count *
+		    PAGE_SIZE);
+}
+#endif
+
+void *
+lkpi_vmap_pfn(unsigned long *pfns, unsigned int count, int prot)
+{
+#ifdef __amd64__
+	vm_offset_t off;
+	size_t size;
+
+	size = count * PAGE_SIZE;
+	off = kva_alloc(size);
+	if (off == 0)
+		return (NULL);
+	vmmap_add((void *)off, size);
+	_lkpi_pmap_qenter_pfn(off, pfns, count, pgprot2cachemode(prot));
+
+	return ((void *)off);
+#else
+	panic("vmap_pfn is not implemented");
+	return (NULL);
+#endif
+}
+
 void
 vunmap(void *addr)
 {
diff --git a/sys/compat/linuxkpi/common/src/linux_page.c b/sys/compat/linuxkpi/common/src/linux_page.c
index 25243382f9e..1ed8b99cdf3 100644
--- a/sys/compat/linuxkpi/common/src/linux_page.c
+++ b/sys/compat/linuxkpi/common/src/linux_page.c
@@ -53,9 +53,6 @@
 #include <vm/vm_reserv.h>
 #include <vm/vm_extern.h>
 
-#include <vm/uma.h>
-#include <vm/uma_int.h>
-
 #include <linux/gfp.h>
 #include <linux/mm.h>
 #include <linux/preempt.h>
@@ -287,12 +284,6 @@ lkpi_get_user_pages(unsigned long start, unsigned long nr_pages,
 	    !!(gup_flags & FOLL_WRITE), pages));
 }
 
-int
-is_vmalloc_addr(const void *addr)
-{
-	return (vtoslab((vm_offset_t)addr & ~UMA_SLAB_MASK) != NULL);
-}
-
 vm_fault_t
 lkpi_vmf_insert_pfn_prot_locked(struct vm_area_struct *vma, unsigned long addr,
     unsigned long pfn, pgprot_t prot)

[kenrap]

Will get on it, thanks! 😎

[kenrap]

@wulf7 core.txt.1