[Intel-gfx] [PATCH 5/7] drm/i915: Implement GGTT update method with MI_UPDATE_GTT
Piotr Piórkowski
piotr.piorkowski at intel.com
Fri Sep 15 16:06:02 UTC 2023
Zeng, Oak <oak.zeng at intel.com> wrote on pią [2023-wrz-15 17:56:56 +0200]:
>
>
> Thanks,
> Oak
>
> > -----Original Message-----
> > From: Das, Nirmoy <nirmoy.das at intel.com>
> > Sent: Friday, September 15, 2023 4:34 AM
> > To: intel-gfx at lists.freedesktop.org
> > Cc: Zeng, Oak <oak.zeng at intel.com>; chris.p.wilson at linux.intel.com; Piorkowski,
> > Piotr <piotr.piorkowski at intel.com>; Shyti, Andi <andi.shyti at intel.com>; Mun,
> > Gwan-gyeong <gwan-gyeong.mun at intel.com>; Roper, Matthew D
> > <matthew.d.roper at intel.com>; Das, Nirmoy <nirmoy.das at intel.com>
> > Subject: [PATCH 5/7] drm/i915: Implement GGTT update method with
> > MI_UPDATE_GTT
> >
> > Implement GGTT update method with blitter command, MI_UPDATE_GTT
> > and install those handlers if a platform requires that.
> >
> > v2: Make sure we hold the GT wakeref and Blitter engine wakeref before
> > we call mutex_lock/intel_context_enter below. When GT/engine are not
> > awake, the intel_context_enter calls into some runtime pm function which
> > can end up with kmalloc/fs_reclaim. But trigger fs_reclaim holding a
> > mutex lock is not allowed because shrinker can also try to hold the same
> > mutex lock. It is a circular lock. So hold the GT/blitter engine wakeref
> > before calling mutex_lock, to fix the circular lock.
> >
> > Signed-off-by: Nirmoy Das <nirmoy.das at intel.com>
> > Signed-off-by: Oak Zeng <oak.zeng at intel.com>
> > ---
> > drivers/gpu/drm/i915/gt/intel_ggtt.c | 235 +++++++++++++++++++++++++++
> > 1 file changed, 235 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c
> > b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> > index dd0ed941441a..b94de7cebfce 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c
> > +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> > @@ -15,18 +15,23 @@
> > #include "display/intel_display.h"
> > #include "gem/i915_gem_lmem.h"
> >
> > +#include "intel_context.h"
> > #include "intel_ggtt_gmch.h"
> > +#include "intel_gpu_commands.h"
> > #include "intel_gt.h"
> > #include "intel_gt_regs.h"
> > #include "intel_pci_config.h"
> > +#include "intel_ring.h"
> > #include "i915_drv.h"
> > #include "i915_pci.h"
> > +#include "i915_request.h"
> > #include "i915_scatterlist.h"
> > #include "i915_utils.h"
> > #include "i915_vgpu.h"
> >
> > #include "intel_gtt.h"
> > #include "gen8_ppgtt.h"
> > +#include "intel_engine_pm.h"
> >
> > static void i915_ggtt_color_adjust(const struct drm_mm_node *node,
> > unsigned long color,
> > @@ -252,6 +257,145 @@ u64 gen8_ggtt_pte_encode(dma_addr_t addr,
> > return pte;
> > }
> >
> > +static bool should_update_ggtt_with_bind(struct i915_ggtt *ggtt)
> > +{
> > + struct intel_gt *gt = ggtt->vm.gt;
> > +
> > + return intel_gt_is_bind_context_ready(gt);
> > +}
> > +
> > +static struct intel_context *gen8_ggtt_bind_get_ce(struct i915_ggtt *ggtt)
> > +{
> > + struct intel_context *ce;
> > + struct intel_gt *gt = ggtt->vm.gt;
> > +
> > + if (intel_gt_is_wedged(gt))
> > + return NULL;
> > +
> > + ce = gt->engine[BCS0]->bind_context;
> > + GEM_BUG_ON(!ce);
> > +
> > + /*
> > + * If the GT is not awake already at this stage then fallback
> > + * to pci based GGTT update otherwise __intel_wakeref_get_first()
> > + * would conflict with fs_reclaim trying to allocate memory while
> > + * doing rpm_resume().
> > + */
> > + if (!intel_gt_pm_get_if_awake(gt))
> > + return NULL;
> > +
> > + intel_engine_pm_get(ce->engine);
> > +
> > + return ce;
> > +}
> > +
> > +static void gen8_ggtt_bind_put_ce(struct intel_context *ce)
> > +{
> > + intel_engine_pm_put(ce->engine);
> > + intel_gt_pm_put(ce->engine->gt);
> > +}
> > +
> > +static bool gen8_ggtt_bind_ptes(struct i915_ggtt *ggtt, u32 offset,
> > + struct sg_table *pages, u32 num_entries,
> > + const gen8_pte_t pte)
> > +{
> > + struct i915_sched_attr attr = {};
> > + struct intel_gt *gt = ggtt->vm.gt;
> > + const gen8_pte_t scratch_pte = ggtt->vm.scratch[0]->encode;
> > + struct sgt_iter iter;
> > + struct i915_request *rq;
> > + struct intel_context *ce;
> > + u32 *cs;
> > +
> > + if (!num_entries)
> > + return true;
> > +
> > + ce = gen8_ggtt_bind_get_ce(ggtt);
> > + if (!ce)
> > + return false;
> > +
> > + if (pages)
> > + iter = __sgt_iter(pages->sgl, true);
> > +
> > + while (num_entries) {
> > + int count = 0;
> > + dma_addr_t addr;
> > + /*
> > + * MI_UPDATE_GTT can update 512 entries in a single command
> > but
> > + * that end up with engine reset, 511 works.
> > + */
> > + u32 n_ptes = min_t(u32, 511, num_entries);
> > +
> > + if (mutex_lock_interruptible(&ce->timeline->mutex))
> > + goto put_ce;
> > +
> > + intel_context_enter(ce);
> > + rq = __i915_request_create(ce, GFP_NOWAIT | GFP_ATOMIC);
> > + intel_context_exit(ce);
> > + if (IS_ERR(rq)) {
> > + GT_TRACE(gt, "Failed to get bind request\n");
> > + mutex_unlock(&ce->timeline->mutex);
> > + goto put_ce;
> > + }
> > +
> > + cs = intel_ring_begin(rq, 2 * n_ptes + 2);
> > + if (IS_ERR(cs)) {
> > + GT_TRACE(gt, "Failed to ring space for GGTT bind\n");
> > + i915_request_set_error_once(rq, PTR_ERR(cs));
> > + /* once a request is created, it must be queued */
> > + goto queue_err_rq;
> > + }
> > +
> > + *cs++ = MI_UPDATE_GTT | (2 * n_ptes);
> > + *cs++ = offset << 12;
> > +
> > + if (pages) {
> > + for_each_sgt_daddr_next(addr, iter) {
> > + if (count == n_ptes)
> > + break;
> > + *cs++ = lower_32_bits(pte | addr);
> > + *cs++ = upper_32_bits(pte | addr);
> > + count++;
> > + }
> > + /* fill remaining with scratch pte, if any */
> > + if (count < n_ptes) {
> > + memset64((u64 *)cs, scratch_pte,
> > + n_ptes - count);
> > + cs += (n_ptes - count) * 2;
> > + }
>
> Should we return error instead of silently fill pte with scratch? Or maybe even gem_bug_on on this case? The caller didn't allocate enough pages, means something wrong happened...
We should avoid GEM_BUG_ON, however, maybe it's a good idea for this function to return an int
instead of a bool, because there are a lot of places here that return an err. In that case,
ret > 0 could mean the number of PTEs we updated correctly and ret < 0 would be an error.
Piotr
>
> Oak
>
> > + } else {
> > + memset64((u64 *)cs, pte, n_ptes);
> > + cs += n_ptes * 2;
> > + }
> > +
> > + intel_ring_advance(rq, cs);
> > +queue_err_rq:
> > + i915_request_get(rq);
> > + __i915_request_commit(rq);
> > + __i915_request_queue(rq, &attr);
> > +
> > + mutex_unlock(&ce->timeline->mutex);
> > + /* This will break if the request is complete or after engine reset
> > */
> > + i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
> > + if (rq->fence.error)
> > + goto err_rq;
> > +
> > + i915_request_put(rq);
> > +
> > + num_entries -= n_ptes;
> > + offset += n_ptes;
> > + }
> > +
> > + gen8_ggtt_bind_put_ce(ce);
> > + return true;
> > +
> > +err_rq:
> > + i915_request_put(rq);
> > +put_ce:
> > + gen8_ggtt_bind_put_ce(ce);
> > + return false;
> > +}
> > +
> > static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
> > {
> > writeq(pte, addr);
> > @@ -272,6 +416,21 @@ static void gen8_ggtt_insert_page(struct
> > i915_address_space *vm,
> > ggtt->invalidate(ggtt);
> > }
> >
> > +static void gen8_ggtt_insert_page_bind(struct i915_address_space *vm,
> > + dma_addr_t addr, u64 offset,
> > + unsigned int pat_index, u32 flags)
> > +{
> > + struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
> > + gen8_pte_t pte;
> > +
> > + pte = ggtt->vm.pte_encode(addr, pat_index, flags);
> > + if (should_update_ggtt_with_bind(i915_vm_to_ggtt(vm)) &&
> > + gen8_ggtt_bind_ptes(ggtt, offset, NULL, 1, pte))
> > + return ggtt->invalidate(ggtt);
> > +
> > + gen8_ggtt_insert_page(vm, addr, offset, pat_index, flags);
> > +}
> > +
> > static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
> > struct i915_vma_resource *vma_res,
> > unsigned int pat_index,
> > @@ -311,6 +470,50 @@ static void gen8_ggtt_insert_entries(struct
> > i915_address_space *vm,
> > ggtt->invalidate(ggtt);
> > }
> >
> > +static bool __gen8_ggtt_insert_entries_bind(struct i915_address_space *vm,
> > + struct i915_vma_resource *vma_res,
> > + unsigned int pat_index, u32 flags)
> > +{
> > + struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
> > + gen8_pte_t scratch_pte = vm->scratch[0]->encode;
> > + gen8_pte_t pte_encode;
> > + u64 start, end;
> > +
> > + pte_encode = ggtt->vm.pte_encode(0, pat_index, flags);
> > + start = (vma_res->start - vma_res->guard) / I915_GTT_PAGE_SIZE;
> > + end = start + vma_res->guard / I915_GTT_PAGE_SIZE;
> > + if (!gen8_ggtt_bind_ptes(ggtt, start, NULL, end - start, scratch_pte))
> > + goto err;
> > +
> > + start = end;
> > + end += (vma_res->node_size + vma_res->guard) / I915_GTT_PAGE_SIZE;
> > + if (!gen8_ggtt_bind_ptes(ggtt, start, vma_res->bi.pages,
> > + vma_res->node_size / I915_GTT_PAGE_SIZE, pte_encode))
> > + goto err;
> > +
> > + start += vma_res->node_size / I915_GTT_PAGE_SIZE;
> > + if (!gen8_ggtt_bind_ptes(ggtt, start, NULL, end - start, scratch_pte))
> > + goto err;
> > +
> > + return true;
> > +
> > +err:
> > + return false;
> > +}
> > +
> > +static void gen8_ggtt_insert_entries_bind(struct i915_address_space *vm,
> > + struct i915_vma_resource *vma_res,
> > + unsigned int pat_index, u32 flags)
> > +{
> > + struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
> > +
> > + if (should_update_ggtt_with_bind(i915_vm_to_ggtt(vm)) &&
> > + __gen8_ggtt_insert_entries_bind(vm, vma_res, pat_index, flags))
> > + return ggtt->invalidate(ggtt);
> > +
> > + gen8_ggtt_insert_entries(vm, vma_res, pat_index, flags);
> > +}
> > +
> > static void gen8_ggtt_clear_range(struct i915_address_space *vm,
> > u64 start, u64 length)
> > {
> > @@ -332,6 +535,27 @@ static void gen8_ggtt_clear_range(struct
> > i915_address_space *vm,
> > gen8_set_pte(>t_base[i], scratch_pte);
> > }
> >
> > +static void gen8_ggtt_scratch_range_bind(struct i915_address_space *vm,
> > + u64 start, u64 length)
> > +{
> > + struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
> > + unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
> > + unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
> > + const gen8_pte_t scratch_pte = vm->scratch[0]->encode;
> > + const int max_entries = ggtt_total_entries(ggtt) - first_entry;
> > +
> > + if (WARN(num_entries > max_entries,
> > + "First entry = %d; Num entries = %d (max=%d)\n",
> > + first_entry, num_entries, max_entries))
> > + num_entries = max_entries;
> > +
> > + if (should_update_ggtt_with_bind(ggtt) && gen8_ggtt_bind_ptes(ggtt,
> > first_entry,
> > + NULL, num_entries, scratch_pte))
> > + return ggtt->invalidate(ggtt);
> > +
> > + gen8_ggtt_clear_range(vm, start, length);
> > +}
> > +
> > static void gen6_ggtt_insert_page(struct i915_address_space *vm,
> > dma_addr_t addr,
> > u64 offset,
> > @@ -997,6 +1221,17 @@ static int gen8_gmch_probe(struct i915_ggtt *ggtt)
> > I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
> > }
> >
> > + if (i915_ggtt_require_binder(i915)) {
> > + ggtt->vm.scratch_range = gen8_ggtt_scratch_range_bind;
> > + ggtt->vm.insert_page = gen8_ggtt_insert_page_bind;
> > + ggtt->vm.insert_entries = gen8_ggtt_insert_entries_bind;
> > + /*
> > + * On GPU is hung, we might bind VMAs for error capture.
> > + * Fallback to CPU GGTT updates in that case.
> > + */
> > + ggtt->vm.raw_insert_page = gen8_ggtt_insert_page;
> > + }
> > +
> > if (intel_uc_wants_guc(&ggtt->vm.gt->uc))
> > ggtt->invalidate = guc_ggtt_invalidate;
> > else
> > --
> > 2.41.0
>
--
More information about the Intel-gfx
mailing list