On Tue, 2021-06-08 at 17:18 +0100, Matthew Auld wrote:
> On Tue, 8 Jun 2021 at 10:29, Thomas Hellström
> <thomas.hellstrom@xxxxxxxxxxxxxxx> wrote:
> >
> > From: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> >
> > If we pipeline the PTE updates and then do the copy of those pages
> > within a single unpreemptible command packet, we can submit the
> > copies
> > and leave them to be scheduled without having to synchronously wait
> > under a global lock. In order to manage migration, we need to
> > preallocate the page tables (and keep them pinned and available for
> > use
> > at any time), causing a bottleneck for migrations as all clients
> > must
> > contend on the limited resources. By inlining the ppGTT updates and
> > performing the blit atomically, each client only owns the PTE while
> > in
> > use, and so we can reschedule individual operations however we see
> > fit.
> > And most importantly, we do not need to take a global lock on the
> > shared
> > vm, and wait until the operation is complete before releasing the
> > lock
> > for others to claim the PTE for themselves.
> >
> > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > Co-developed-by: Thomas Hellström
> > <thomas.hellstrom@xxxxxxxxxxxxxxx>
> > Signed-off-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
> > ---
> > drivers/gpu/drm/i915/Makefile | 1 +
> > drivers/gpu/drm/i915/gt/intel_engine.h | 1 +
> > drivers/gpu/drm/i915/gt/intel_gpu_commands.h | 2 +
> > drivers/gpu/drm/i915/gt/intel_migrate.c | 543
> > ++++++++++++++++++
> > drivers/gpu/drm/i915/gt/intel_migrate.h | 45 ++
> > drivers/gpu/drm/i915/gt/intel_migrate_types.h | 15 +
> > drivers/gpu/drm/i915/gt/intel_ring.h | 1 +
> > drivers/gpu/drm/i915/gt/selftest_migrate.c | 291 ++++++++++
> > .../drm/i915/selftests/i915_live_selftests.h | 1 +
> > 9 files changed, 900 insertions(+)
> > create mode 100644 drivers/gpu/drm/i915/gt/intel_migrate.c
> > create mode 100644 drivers/gpu/drm/i915/gt/intel_migrate.h
> > create mode 100644 drivers/gpu/drm/i915/gt/intel_migrate_types.h
> > create mode 100644 drivers/gpu/drm/i915/gt/selftest_migrate.c
> >
> > diff --git a/drivers/gpu/drm/i915/Makefile
> > b/drivers/gpu/drm/i915/Makefile
> > index ea8ee4b3e018..9f18902be626 100644
> > --- a/drivers/gpu/drm/i915/Makefile
> > +++ b/drivers/gpu/drm/i915/Makefile
> > @@ -109,6 +109,7 @@ gt-y += \
> > gt/intel_gtt.o \
> > gt/intel_llc.o \
> > gt/intel_lrc.o \
> > + gt/intel_migrate.o \
> > gt/intel_mocs.o \
> > gt/intel_ppgtt.o \
> > gt/intel_rc6.o \
> > diff --git a/drivers/gpu/drm/i915/gt/intel_engine.h
> > b/drivers/gpu/drm/i915/gt/intel_engine.h
> > index 0862c42b4cac..949965680c37 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_engine.h
> > +++ b/drivers/gpu/drm/i915/gt/intel_engine.h
> > @@ -188,6 +188,7 @@ intel_write_status_page(struct intel_engine_cs
> > *engine, int reg, u32 value)
> > #define I915_GEM_HWS_PREEMPT_ADDR (I915_GEM_HWS_PREEMPT *
> > sizeof(u32))
> > #define I915_GEM_HWS_SEQNO 0x40
> > #define I915_GEM_HWS_SEQNO_ADDR (I915_GEM_HWS_SEQNO
> > * sizeof(u32))
> > +#define I915_GEM_HWS_MIGRATE (0x42 * sizeof(u32))
> > #define I915_GEM_HWS_SCRATCH 0x80
> >
> > #define I915_HWS_CSB_BUF0_INDEX 0x10
> > diff --git a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
> > b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
> > index 2694dbb9967e..1c3af0fc0456 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
> > +++ b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
> > @@ -123,8 +123,10 @@
> > #define MI_SEMAPHORE_SAD_NEQ_SDD (5 << 12)
> > #define MI_SEMAPHORE_TOKEN_MASK REG_GENMASK(9, 5)
> > #define MI_SEMAPHORE_TOKEN_SHIFT 5
> > +#define MI_STORE_DATA_IMM MI_INSTR(0x20, 0)
> > #define MI_STORE_DWORD_IMM MI_INSTR(0x20, 1)
> > #define MI_STORE_DWORD_IMM_GEN4 MI_INSTR(0x20, 2)
> > +#define MI_STORE_QWORD_IMM_GEN8 (MI_INSTR(0x20, 3) | REG_BIT(21))
> > #define MI_MEM_VIRTUAL (1 << 22) /* 945,g33,965 */
> > #define MI_USE_GGTT (1 << 22) /* g4x+ */
> > #define MI_STORE_DWORD_INDEX MI_INSTR(0x21, 1)
> > diff --git a/drivers/gpu/drm/i915/gt/intel_migrate.c
> > b/drivers/gpu/drm/i915/gt/intel_migrate.c
> > new file mode 100644
> > index 000000000000..1f60f8ee36f8
> > --- /dev/null
> > +++ b/drivers/gpu/drm/i915/gt/intel_migrate.c
> > @@ -0,0 +1,543 @@
> > +// SPDX-License-Identifier: MIT
> > +/*
> > + * Copyright © 2020 Intel Corporation
> > + */
> > +
> > +#include "i915_drv.h"
> > +#include "intel_context.h"
> > +#include "intel_gpu_commands.h"
> > +#include "intel_gt.h"
> > +#include "intel_gtt.h"
> > +#include "intel_migrate.h"
> > +#include "intel_ring.h"
> > +
> > +struct insert_pte_data {
> > + u64 offset;
> > + bool is_lmem;
> > +};
> > +
> > +#define CHUNK_SZ SZ_8M /* ~1ms at 8GiB/s preemption delay */
> > +
> > +static bool engine_supports_migration(struct intel_engine_cs
> > *engine)
> > +{
> > + if (!engine)
> > + return false;
> > +
> > + /*
> > + * We need the ability to prevent aribtration
> > (MI_ARB_ON_OFF),
> > + * the ability to write PTE using inline data
> > (MI_STORE_DATA)
> > + * and of course the ability to do the block transfer
> > (blits).
> > + */
> > + GEM_BUG_ON(engine->class != COPY_ENGINE_CLASS);
> > +
> > + return true;
> > +}
> > +
> > +static void insert_pte(struct i915_address_space *vm,
> > + struct i915_page_table *pt,
> > + void *data)
> > +{
> > + struct insert_pte_data *d = data;
> > +
> > + vm->insert_page(vm, px_dma(pt), d->offset, I915_CACHE_NONE,
> > + d->is_lmem ? PTE_LM : 0);
> > + d->offset += PAGE_SIZE;
> > +}
> > +
> > +static struct i915_address_space *migrate_vm(struct intel_gt *gt)
> > +{
> > + struct i915_vm_pt_stash stash = {};
> > + struct i915_ppgtt *vm;
> > + int err;
> > + int i;
> > +
> > + /*
> > + * We construct a very special VM for use by all migration
> > contexts,
> > + * it is kept pinned so that it can be used at any time. As
> > we need
> > + * to pre-allocate the page directories for the migration
> > VM, this
> > + * limits us to only using a small number of prepared vma.
> > + *
> > + * To be able to pipeline and reschedule migration
> > operations while
> > + * avoiding unnecessary contention on the vm itself, the
> > PTE updates
> > + * are inline with the blits. All the blits use the same
> > fixed
> > + * addresses, with the backing store redirection being
> > updated on the
> > + * fly. Only 2 implicit vma are used for all migration
> > operations.
> > + *
> > + * We lay the ppGTT out as:
> > + *
> > + * [0, CHUNK_SZ) -> first object
> > + * [CHUNK_SZ, 2 * CHUNK_SZ) -> second object
> > + * [2 * CHUNK_SZ, 2 * CHUNK_SZ + 2 * CHUNK_SZ >> 9] ->
> > PTE
> > + *
> > + * By exposing the dma addresses of the page directories
> > themselves
> > + * within the ppGTT, we are then able to rewrite the PTE
> > prior to use.
> > + * But the PTE update and subsequent migration operation
> > must be atomic,
> > + * i.e. within the same non-preemptible window so that we
> > do not switch
> > + * to another migration context that overwrites the PTE.
> > + */
> > +
> > + vm = i915_ppgtt_create(gt);
> > + if (IS_ERR(vm))
> > + return ERR_CAST(vm);
> > +
> > + if (!vm->vm.allocate_va_range || !vm->vm.foreach) {
> > + err = -ENODEV;
> > + goto err_vm;
> > + }
> > +
> > + /*
> > + * Each engine instance is assigned its own chunk in the
> > VM, so
> > + * that we can run multiple instances concurrently
> > + */
> > + for (i = 0; i < ARRAY_SIZE(gt-
> > >engine_class[COPY_ENGINE_CLASS]); i++) {
> > + struct intel_engine_cs *engine;
> > + u64 base = (u64)i << 32;
> > + struct insert_pte_data d = {};
> > + struct i915_gem_ww_ctx ww;
> > + u64 sz;
> > +
> > + engine = gt->engine_class[COPY_ENGINE_CLASS][i];
> > + if (!engine_supports_migration(engine))
> > + continue;
> > +
> > + /*
> > + * We copy in 8MiB chunks. Each PDE covers 2MiB, so
> > we need
> > + * 4x2 page directories for source/destination.
> > + */
> > + sz = 2 * CHUNK_SZ;
> > + d.offset = base + sz;
> > +
> > + /*
> > + * We need another page directory setup so that we
> > can write
> > + * the 8x512 PTE in each chunk.
> > + */
> > + sz += (sz >> 12) * sizeof(u64);
> > +
> > + err = i915_vm_alloc_pt_stash(&vm->vm, &stash, sz);
> > + if (err)
> > + goto err_vm;
> > +
> > + for_i915_gem_ww(&ww, err, true) {
> > + err = i915_vm_lock_objects(&vm->vm, &ww);
> > + if (err)
> > + continue;
> > + err = i915_vm_map_pt_stash(&vm->vm,
> > &stash);
> > + if (err)
> > + continue;
> > +
> > + vm->vm.allocate_va_range(&vm->vm, &stash,
> > base, base + sz);
> > + }
> > + i915_vm_free_pt_stash(&vm->vm, &stash);
> > + if (err)
> > + goto err_vm;
> > +
> > + /* Now allow the GPU to rewrite the PTE via its own
> > ppGTT */
> > + d.is_lmem = i915_gem_object_is_lmem(vm-
> > >vm.scratch[0]);
> > + vm->vm.foreach(&vm->vm, base, base + sz,
> > insert_pte, &d);
>
> Will this play nice with Xe HP where we have min page size
> restrictions for the GTT + lmem? The page-table allocations sidestep
> such restrictions since they were previously never inserted into the
> GTT. Maybe we need the flat-ppGTT at some point? Perhaps add a TODO
> or
> something?
I've previously discussed that with Chris, and apparently it should be
as easy as adjusting the sz >> 12 above. But yeah, Chris also brought
up the flat-ppGTT idea as something we should look into.
Mieanwhile, we should add a TODO.
/Thomas
