Commit bf890040 authored by Lucas De Marchi's avatar Lucas De Marchi

drm/i915/guc: Use a single pass to calculate regset

The ADS initialitazion was using 2 passes to calculate the regset sent
to GuC to initialize each engine: the first pass to just have the final
object size and the second to set each register in place in the final
gem object.

However in order to maintain an ordered set of registers to pass to guc,
each register needs to be added and moved in the final array. The second
phase may actually happen in IO memory rather than system memory and
accessing IO memory by simply dereferencing the pointer doesn't work on
all architectures. Other places of the ADS initializaition were
converted to use the iosys_map API, but here there may be a lot more
accesses to IO memory. So, instead of following that same approach,
convert the regset initialization to calculate the final array in 1
pass and in the second pass that array is just copied to its final
location, updating the pointers for each engine written to the ADS blob.

One important thing is that struct temp_regset now have
different semantics: `registers` continues to track the registers of a
single engine, however the other fields are updated together, according
to the newly added `storage`, which tracks the memory allocated for
all the registers. So rename some of these fields and add a
__mmio_reg_add(): this function (possibly) allocates memory and operates
on the storage pointer while guc_mmio_reg_add() continues to manage the
registers pointer.

On a Tiger Lake system using enable_guc=3, the following log message is
now seen:

	[  187.334310] i915 0000:00:02.0: [drm:intel_guc_ads_create [i915]] Used 4 KB for temporary ADS regset

This change has also been tested on an ARM64 host with DG2 and other
discrete graphics cards.

v2 (Daniele):
  - Fix leaking tempset on error path
  - Add comments on struct temp_regset to document the meaning of each
    field

Cc: Matt Roper <matthew.d.roper@intel.com>
Cc: John Harrison <John.C.Harrison@Intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Signed-off-by: default avatarLucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: default avatarDaniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220208070141.2095177-3-lucas.demarchi@intel.com
parent f4044ca1
...@@ -150,6 +150,13 @@ struct intel_guc { ...@@ -150,6 +150,13 @@ struct intel_guc {
struct __guc_ads_blob *ads_blob; struct __guc_ads_blob *ads_blob;
/** @ads_regset_size: size of the save/restore regsets in the ADS */ /** @ads_regset_size: size of the save/restore regsets in the ADS */
u32 ads_regset_size; u32 ads_regset_size;
/**
* @ads_regset_count: number of save/restore registers in the ADS for
* each engine
*/
u32 ads_regset_count[I915_NUM_ENGINES];
/** @ads_regset: save/restore regsets in the ADS */
struct guc_mmio_reg *ads_regset;
/** @ads_golden_ctxt_size: size of the golden contexts in the ADS */ /** @ads_golden_ctxt_size: size of the golden contexts in the ADS */
u32 ads_golden_ctxt_size; u32 ads_golden_ctxt_size;
/** @ads_engine_usage_size: size of engine usage in the ADS */ /** @ads_engine_usage_size: size of engine usage in the ADS */
......
...@@ -206,14 +206,18 @@ static void guc_mapping_table_init(struct intel_gt *gt, ...@@ -206,14 +206,18 @@ static void guc_mapping_table_init(struct intel_gt *gt,
/* /*
* The save/restore register list must be pre-calculated to a temporary * The save/restore register list must be pre-calculated to a temporary
* buffer of driver defined size before it can be generated in place * buffer before it can be copied inside the ADS.
* inside the ADS.
*/ */
#define MAX_MMIO_REGS 128 /* Arbitrary size, increase as needed */
struct temp_regset { struct temp_regset {
/*
* ptr to the section of the storage for the engine currently being
* worked on
*/
struct guc_mmio_reg *registers; struct guc_mmio_reg *registers;
u32 used; /* ptr to the base of the allocated storage for all engines */
u32 size; struct guc_mmio_reg *storage;
u32 storage_used;
u32 storage_max;
}; };
static int guc_mmio_reg_cmp(const void *a, const void *b) static int guc_mmio_reg_cmp(const void *a, const void *b)
...@@ -224,18 +228,44 @@ static int guc_mmio_reg_cmp(const void *a, const void *b) ...@@ -224,18 +228,44 @@ static int guc_mmio_reg_cmp(const void *a, const void *b)
return (int)ra->offset - (int)rb->offset; return (int)ra->offset - (int)rb->offset;
} }
static struct guc_mmio_reg * __must_check
__mmio_reg_add(struct temp_regset *regset, struct guc_mmio_reg *reg)
{
u32 pos = regset->storage_used;
struct guc_mmio_reg *slot;
if (pos >= regset->storage_max) {
size_t size = ALIGN((pos + 1) * sizeof(*slot), PAGE_SIZE);
struct guc_mmio_reg *r = krealloc(regset->storage,
size, GFP_KERNEL);
if (!r) {
WARN_ONCE(1, "Incomplete regset list: can't add register (%d)\n",
-ENOMEM);
return ERR_PTR(-ENOMEM);
}
regset->registers = r + (regset->registers - regset->storage);
regset->storage = r;
regset->storage_max = size / sizeof(*slot);
}
slot = &regset->storage[pos];
regset->storage_used++;
*slot = *reg;
return slot;
}
static long __must_check guc_mmio_reg_add(struct temp_regset *regset, static long __must_check guc_mmio_reg_add(struct temp_regset *regset,
u32 offset, u32 flags) u32 offset, u32 flags)
{ {
u32 count = regset->used; u32 count = regset->storage_used - (regset->registers - regset->storage);
struct guc_mmio_reg reg = { struct guc_mmio_reg reg = {
.offset = offset, .offset = offset,
.flags = flags, .flags = flags,
}; };
struct guc_mmio_reg *slot; struct guc_mmio_reg *slot;
GEM_BUG_ON(count >= regset->size);
/* /*
* The mmio list is built using separate lists within the driver. * The mmio list is built using separate lists within the driver.
* It's possible that at some point we may attempt to add the same * It's possible that at some point we may attempt to add the same
...@@ -246,9 +276,9 @@ static long __must_check guc_mmio_reg_add(struct temp_regset *regset, ...@@ -246,9 +276,9 @@ static long __must_check guc_mmio_reg_add(struct temp_regset *regset,
sizeof(reg), guc_mmio_reg_cmp)) sizeof(reg), guc_mmio_reg_cmp))
return 0; return 0;
slot = &regset->registers[count]; slot = __mmio_reg_add(regset, &reg);
regset->used++; if (IS_ERR(slot))
*slot = reg; return PTR_ERR(slot);
while (slot-- > regset->registers) { while (slot-- > regset->registers) {
GEM_BUG_ON(slot[0].offset == slot[1].offset); GEM_BUG_ON(slot[0].offset == slot[1].offset);
...@@ -275,7 +305,11 @@ static int guc_mmio_regset_init(struct temp_regset *regset, ...@@ -275,7 +305,11 @@ static int guc_mmio_regset_init(struct temp_regset *regset,
unsigned int i; unsigned int i;
int ret = 0; int ret = 0;
regset->used = 0; /*
* Each engine's registers point to a new start relative to
* storage
*/
regset->registers = regset->storage + regset->storage_used;
ret |= GUC_MMIO_REG_ADD(regset, RING_MODE_GEN7(base), true); ret |= GUC_MMIO_REG_ADD(regset, RING_MODE_GEN7(base), true);
ret |= GUC_MMIO_REG_ADD(regset, RING_HWS_PGA(base), false); ret |= GUC_MMIO_REG_ADD(regset, RING_HWS_PGA(base), false);
...@@ -297,34 +331,36 @@ static int guc_mmio_regset_init(struct temp_regset *regset, ...@@ -297,34 +331,36 @@ static int guc_mmio_regset_init(struct temp_regset *regset,
return ret ? -1 : 0; return ret ? -1 : 0;
} }
static int guc_mmio_reg_state_query(struct intel_guc *guc) static long guc_mmio_reg_state_create(struct intel_guc *guc)
{ {
struct intel_gt *gt = guc_to_gt(guc); struct intel_gt *gt = guc_to_gt(guc);
struct intel_engine_cs *engine; struct intel_engine_cs *engine;
enum intel_engine_id id; enum intel_engine_id id;
struct temp_regset temp_set; struct temp_regset temp_set = {};
u32 total; long total = 0;
long ret;
/*
* Need to actually build the list in order to filter out
* duplicates and other such data dependent constructions.
*/
temp_set.size = MAX_MMIO_REGS;
temp_set.registers = kmalloc_array(temp_set.size,
sizeof(*temp_set.registers),
GFP_KERNEL);
if (!temp_set.registers)
return -ENOMEM;
total = 0;
for_each_engine(engine, gt, id) { for_each_engine(engine, gt, id) {
guc_mmio_regset_init(&temp_set, engine); u32 used = temp_set.storage_used;
total += temp_set.used;
ret = guc_mmio_regset_init(&temp_set, engine);
if (ret < 0)
goto fail_regset_init;
guc->ads_regset_count[id] = temp_set.storage_used - used;
total += guc->ads_regset_count[id];
} }
kfree(temp_set.registers); guc->ads_regset = temp_set.storage;
drm_dbg(&guc_to_gt(guc)->i915->drm, "Used %zu KB for temporary ADS regset\n",
(temp_set.storage_max * sizeof(struct guc_mmio_reg)) >> 10);
return total * sizeof(struct guc_mmio_reg); return total * sizeof(struct guc_mmio_reg);
fail_regset_init:
kfree(temp_set.storage);
return ret;
} }
static void guc_mmio_reg_state_init(struct intel_guc *guc, static void guc_mmio_reg_state_init(struct intel_guc *guc,
...@@ -332,40 +368,38 @@ static void guc_mmio_reg_state_init(struct intel_guc *guc, ...@@ -332,40 +368,38 @@ static void guc_mmio_reg_state_init(struct intel_guc *guc,
{ {
struct intel_gt *gt = guc_to_gt(guc); struct intel_gt *gt = guc_to_gt(guc);
struct intel_engine_cs *engine; struct intel_engine_cs *engine;
struct guc_mmio_reg *ads_registers;
enum intel_engine_id id; enum intel_engine_id id;
struct temp_regset temp_set;
struct guc_mmio_reg_set *ads_reg_set;
u32 addr_ggtt, offset; u32 addr_ggtt, offset;
u8 guc_class;
offset = guc_ads_regset_offset(guc); offset = guc_ads_regset_offset(guc);
addr_ggtt = intel_guc_ggtt_offset(guc, guc->ads_vma) + offset; addr_ggtt = intel_guc_ggtt_offset(guc, guc->ads_vma) + offset;
temp_set.registers = (struct guc_mmio_reg *)(((u8 *)blob) + offset); ads_registers = (struct guc_mmio_reg *)(((u8 *)blob) + offset);
temp_set.size = guc->ads_regset_size / sizeof(temp_set.registers[0]);
memcpy(ads_registers, guc->ads_regset, guc->ads_regset_size);
for_each_engine(engine, gt, id) { for_each_engine(engine, gt, id) {
u32 count = guc->ads_regset_count[id];
struct guc_mmio_reg_set *ads_reg_set;
u8 guc_class;
/* Class index is checked in class converter */ /* Class index is checked in class converter */
GEM_BUG_ON(engine->instance >= GUC_MAX_INSTANCES_PER_CLASS); GEM_BUG_ON(engine->instance >= GUC_MAX_INSTANCES_PER_CLASS);
guc_class = engine_class_to_guc_class(engine->class); guc_class = engine_class_to_guc_class(engine->class);
ads_reg_set = &blob->ads.reg_state_list[guc_class][engine->instance]; ads_reg_set = &blob->ads.reg_state_list[guc_class][engine->instance];
guc_mmio_regset_init(&temp_set, engine); if (!count) {
if (!temp_set.used) {
ads_reg_set->address = 0; ads_reg_set->address = 0;
ads_reg_set->count = 0; ads_reg_set->count = 0;
continue; continue;
} }
ads_reg_set->address = addr_ggtt; ads_reg_set->address = addr_ggtt;
ads_reg_set->count = temp_set.used; ads_reg_set->count = count;
temp_set.size -= temp_set.used; addr_ggtt += count * sizeof(struct guc_mmio_reg);
temp_set.registers += temp_set.used;
addr_ggtt += temp_set.used * sizeof(struct guc_mmio_reg);
} }
GEM_BUG_ON(temp_set.size);
} }
static void fill_engine_enable_masks(struct intel_gt *gt, static void fill_engine_enable_masks(struct intel_gt *gt,
...@@ -607,8 +641,11 @@ int intel_guc_ads_create(struct intel_guc *guc) ...@@ -607,8 +641,11 @@ int intel_guc_ads_create(struct intel_guc *guc)
GEM_BUG_ON(guc->ads_vma); GEM_BUG_ON(guc->ads_vma);
/* Need to calculate the reg state size dynamically: */ /*
ret = guc_mmio_reg_state_query(guc); * Create reg state size dynamically on system memory to be copied to
* the final ads blob on gt init/reset
*/
ret = guc_mmio_reg_state_create(guc);
if (ret < 0) if (ret < 0)
return ret; return ret;
guc->ads_regset_size = ret; guc->ads_regset_size = ret;
...@@ -648,6 +685,7 @@ void intel_guc_ads_destroy(struct intel_guc *guc) ...@@ -648,6 +685,7 @@ void intel_guc_ads_destroy(struct intel_guc *guc)
{ {
i915_vma_unpin_and_release(&guc->ads_vma, I915_VMA_RELEASE_MAP); i915_vma_unpin_and_release(&guc->ads_vma, I915_VMA_RELEASE_MAP);
guc->ads_blob = NULL; guc->ads_blob = NULL;
kfree(guc->ads_regset);
} }
static void guc_ads_private_data_reset(struct intel_guc *guc) static void guc_ads_private_data_reset(struct intel_guc *guc)
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment