On Tue, Sep 20, 2022 at 12:49:40PM +0530, Aravind Iddamsetty wrote:
As an integrated GPU, MTL does not have local memory and
HAS_LMEM() returns false. However the platform's stolen memory
is presented via BAR2 (i.e., the BAR we traditionally consider
to be the LMEM BAR) and should be managed by the driver the same
way that local memory is on dgpu platforms (which includes
setting the "lmem" bit on page table entries). We use the term
"local stolen memory" to refer to this model.
Cc: Matt Roper <matthew.d.roper@xxxxxxxxx>
Cc: Lucas De Marchi <lucas.demarchi@xxxxxxxxx>
Signed-off-by: CQ Tang <cq.tang@xxxxxxxxx>
Signed-off-by: Aravind Iddamsetty <aravind.iddamsetty@xxxxxxxxx>
Original-author: CQ Tang
---
drivers/gpu/drm/i915/gem/i915_gem_stolen.c | 113 +++++++++++++++++----
drivers/gpu/drm/i915/gt/intel_ggtt.c | 2 +-
drivers/gpu/drm/i915/i915_drv.h | 3 +
3 files changed, 100 insertions(+), 18 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_stolen.c b/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
index acc561c0f0aa..bad5250fb764 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
@@ -77,6 +77,19 @@ void i915_gem_stolen_remove_node(struct drm_i915_private *i915,
mutex_unlock(&i915->mm.stolen_lock);
}
+static bool is_dsm_invalid(struct drm_i915_private *i915, struct resource *dsm)
+{
+ if (!HAS_BAR2_SMEM_STOLEN(i915)) {
I called a similar function as is_dsm_valid() in
https://patchwork.freedesktop.org/series/108620/
sounds weird with "invalid" and the double negation on return early
style.
+ if (dsm->start == 0)
+ return true;
+ }
+
+ if (dsm->end <= dsm->start)
+ return true;
+
+ return false;
+}
+
static int i915_adjust_stolen(struct drm_i915_private *i915,
struct resource *dsm)
{
@@ -84,7 +97,7 @@ static int i915_adjust_stolen(struct drm_i915_private *i915,
struct intel_uncore *uncore = ggtt->vm.gt->uncore;
struct resource *r;
- if (dsm->start == 0 || dsm->end <= dsm->start)
+ if (is_dsm_invalid(i915, dsm))
return -EINVAL;
/*
@@ -136,7 +149,7 @@ static int i915_adjust_stolen(struct drm_i915_private *i915,
* overlaps with the non-stolen system memory range, since lmem is local
* to the gpu.
*/
- if (HAS_LMEM(i915))
+ if (HAS_LMEM(i915) || HAS_BAR2_SMEM_STOLEN(i915))
comment above makes no sense when you add this. For this specific case
it's still system memory, reserved by the BIOS and that we access by
mapping the lmembar
return 0;
/*
@@ -371,8 +384,6 @@ static void icl_get_stolen_reserved(struct drm_i915_private *i915,
drm_dbg(&i915->drm, "GEN6_STOLEN_RESERVED = 0x%016llx\n", reg_val);
- *base = reg_val & GEN11_STOLEN_RESERVED_ADDR_MASK;
-
switch (reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK) {
case GEN8_STOLEN_RESERVED_1M:
*size = 1024 * 1024;
@@ -390,6 +401,12 @@ static void icl_get_stolen_reserved(struct drm_i915_private *i915,
*size = 8 * 1024 * 1024;
MISSING_CASE(reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK);
}
+
+ if ((GRAPHICS_VER_FULL(i915) >= IP_VER(12, 70)) && !IS_DGFX(i915))
+ /* the base is initialized to stolen top so subtract size to get base */
+ *base -= *size;
that doesn't necessarily holds true. According to the spec the wopcm
base is 1MB aligned so even if it is "at the top", it may not mean it is at the
very very top that we can just subtract the size.
+ else
+ *base = reg_val & GEN11_STOLEN_RESERVED_ADDR_MASK;
}
static int i915_gem_init_stolen(struct intel_memory_region *mem)
@@ -423,8 +440,7 @@ static int i915_gem_init_stolen(struct intel_memory_region *mem)
if (i915_adjust_stolen(i915, &i915->dsm))
return 0;
- GEM_BUG_ON(i915->dsm.start == 0);
- GEM_BUG_ON(i915->dsm.end <= i915->dsm.start);
+ GEM_BUG_ON(is_dsm_invalid(i915, &i915->dsm));
stolen_top = i915->dsm.end + 1;
reserved_base = stolen_top;
@@ -796,6 +812,46 @@ static const struct intel_memory_region_ops i915_region_stolen_lmem_ops = {
.init_object = _i915_gem_object_stolen_init,
};
+static int get_mtl_gms_size(struct intel_uncore *uncore)
+{
+ u16 ggc, gms;
+
+ ggc = intel_uncore_read16(uncore, _MMIO(0x108040));
??
+
+ /* check GGMS, should be fixed 0x3 (8MB) */
+ if ((ggc & 0xc0) != 0xc0)
+ return -EIO;
+
+ /* return valid GMS value, -EIO if invalid */
+ gms = ggc >> 8;
+ switch (gms) {
+ case 0x0 ... 0x10:
+ return gms * 32;
+ case 0x20:
+ return 1024;
+ case 0x30:
+ return 1536;
+ case 0x40:
+ return 2048;
+ case 0xf0 ... 0xfe:
+ return (gms - 0xf0 + 1) * 4;
+ default:
+ return -EIO;
+ }
+}
+
+static inline bool lmembar_is_igpu_stolen(struct drm_i915_private *i915)
doesn't deserve an inline. lmembar_is_igpu_stolen() doesn't make much
sense as the lmembar is never igpu stolen.... you probably mean
something else here
+{
+ u32 regions = RUNTIME_INFO(i915)->memory_regions;
+
+ if (regions & REGION_LMEM)
+ return false;
+
+ drm_WARN_ON(&i915->drm, (regions & REGION_STOLEN_LMEM) == 0);
+
+ return true;
+}
+
struct intel_memory_region *
i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
u16 instance)
@@ -806,19 +862,16 @@ i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
struct intel_memory_region *mem;
resource_size_t io_start, io_size;
resource_size_t min_page_size;
+ int ret;
if (WARN_ON_ONCE(instance))
return ERR_PTR(-ENODEV);
- if (!i915_pci_resource_valid(pdev, GEN12_LMEM_BAR))
+ if (!i915_pci_resource_valid(pdev, GFXMEM_BAR))
at least for MTL, Bspec 63830 still calls this lmembar. So the rename
for me is a net loss
return ERR_PTR(-ENXIO);
- /* Use DSM base address instead for stolen memory */
- dsm_base = intel_uncore_read64(uncore, GEN12_DSMBASE);
- if (IS_DG1(uncore->i915)) {
- lmem_size = pci_resource_len(pdev, GEN12_LMEM_BAR);
- if (WARN_ON(lmem_size < dsm_base))
- return ERR_PTR(-ENODEV);
+ if (lmembar_is_igpu_stolen(i915) || IS_DG1(i915)) {
+ lmem_size = pci_resource_len(pdev, GFXMEM_BAR);
this looks confusing, but apparently correct. For DG1 the stolen is
on top of lmem. For MTL, it's on the end of lmembar (256M). This works
because on DG1 aperture == lmem size.
} else {
resource_size_t lmem_range;
@@ -827,13 +880,39 @@ i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
lmem_size *= SZ_1G;
}
- dsm_size = lmem_size - dsm_base;
- if (pci_resource_len(pdev, GEN12_LMEM_BAR) < lmem_size) {
+ if (HAS_BAR2_SMEM_STOLEN(i915)) {
+ /*
+ * MTL dsm size is in GGC register, not the bar size.
it's not exclusive to MTL. it has been there for ages and it was never
the BAR size like this comment says. Or at least it doesn't match the
else condition that is using the GEN12_DSMBASE register
+ * also MTL uses offset to DSMBASE in ptes, so i915
+ * uses dsm_base = 0 to setup stolen region.
+ */
+ ret = get_mtl_gms_size(uncore);
+ if (ret < 0) {
+ drm_err(&i915->drm, "invalid MTL GGC register setting\n");
+ return ERR_PTR(ret);
+ }
+
+ dsm_base = 0;
if we stop handling part of the values in the registers as relative to
the mapping and rather handle them as we read from the registers
(physical addresses), the size calculations should still match and we
shouldn't need all the if/else dance. If we pass the right io_start we
can then make them relative to the mapping by subtracting it, or if we
don't want GTT to be in the mapping we subtract it.
That makes me wonder if choosing the i915_gem_stolen_lmem_setup() for
all of this is even the right choice given we are actually talking about
system memory that is mapped through the lmembar.
