OSDN Git Service

drm/i915/execlists: Process one CSB update at a time
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 28 Jun 2018 20:12:06 +0000 (21:12 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 28 Jun 2018 21:55:04 +0000 (22:55 +0100)
In the next patch, we will process the CSB events directly from the
submission path, rather than only after a CS interrupt. Hence, we will
no longer have the need for a loop until the has-interrupt bit is clear,
and in the meantime can remove that small optimisation.

v2: Tvrtko pointed out it was safer to unconditionally kick the tasklet
after each irq, when assuming that the tasklet is called for each irq.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180628201211.13837-4-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/intel_lrc.c

index 02fe0d3..c95cb27 100644 (file)
@@ -1497,9 +1497,10 @@ gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir)
        bool tasklet = false;
 
        if (iir & GT_CONTEXT_SWITCH_INTERRUPT) {
-               if (READ_ONCE(engine->execlists.active))
-                       tasklet = !test_and_set_bit(ENGINE_IRQ_EXECLIST,
-                                                   &engine->irq_posted);
+               if (READ_ONCE(engine->execlists.active)) {
+                       set_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
+                       tasklet = true;
+               }
        }
 
        if (iir & GT_RENDER_USER_INTERRUPT) {
index 0cee34e..c8d3d41 100644 (file)
@@ -955,166 +955,162 @@ static void process_csb(struct intel_engine_cs *engine)
        struct intel_engine_execlists * const execlists = &engine->execlists;
        struct execlist_port *port = execlists->port;
        struct drm_i915_private *i915 = engine->i915;
+
+       /* The HWSP contains a (cacheable) mirror of the CSB */
+       const u32 *buf =
+               &engine->status_page.page_addr[I915_HWS_CSB_BUF0_INDEX];
+       unsigned int head, tail;
        bool fw = false;
 
-       do {
-               /* The HWSP contains a (cacheable) mirror of the CSB */
-               const u32 *buf =
-                       &engine->status_page.page_addr[I915_HWS_CSB_BUF0_INDEX];
-               unsigned int head, tail;
-
-               /* Clear before reading to catch new interrupts */
-               clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
-               smp_mb__after_atomic();
-
-               if (unlikely(execlists->csb_use_mmio)) {
-                       if (!fw) {
-                               intel_uncore_forcewake_get(i915, execlists->fw_domains);
-                               fw = true;
-                       }
+       /* Clear before reading to catch new interrupts */
+       clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
+       smp_mb__after_atomic();
 
-                       buf = (u32 * __force)
-                               (i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_BUF_LO(engine, 0)));
+       if (unlikely(execlists->csb_use_mmio)) {
+               intel_uncore_forcewake_get(i915, execlists->fw_domains);
+               fw = true;
 
-                       head = readl(i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)));
-                       tail = GEN8_CSB_WRITE_PTR(head);
-                       head = GEN8_CSB_READ_PTR(head);
-                       execlists->csb_head = head;
-               } else {
-                       const int write_idx =
-                               intel_hws_csb_write_index(i915) -
-                               I915_HWS_CSB_BUF0_INDEX;
+               buf = (u32 * __force)
+                       (i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_BUF_LO(engine, 0)));
 
-                       head = execlists->csb_head;
-                       tail = READ_ONCE(buf[write_idx]);
-                       rmb(); /* Hopefully paired with a wmb() in HW */
-               }
-               GEM_TRACE("%s cs-irq head=%d [%d%s], tail=%d [%d%s]\n",
-                         engine->name,
-                         head, GEN8_CSB_READ_PTR(readl(i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)))), fw ? "" : "?",
-                         tail, GEN8_CSB_WRITE_PTR(readl(i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)))), fw ? "" : "?");
+               head = readl(i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)));
+               tail = GEN8_CSB_WRITE_PTR(head);
+               head = GEN8_CSB_READ_PTR(head);
+               execlists->csb_head = head;
+       } else {
+               const int write_idx =
+                       intel_hws_csb_write_index(i915) -
+                       I915_HWS_CSB_BUF0_INDEX;
 
-               while (head != tail) {
-                       struct i915_request *rq;
-                       unsigned int status;
-                       unsigned int count;
+               head = execlists->csb_head;
+               tail = READ_ONCE(buf[write_idx]);
+               rmb(); /* Hopefully paired with a wmb() in HW */
+       }
+       GEM_TRACE("%s cs-irq head=%d [%d%s], tail=%d [%d%s]\n",
+                 engine->name,
+                 head, GEN8_CSB_READ_PTR(readl(i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)))), fw ? "" : "?",
+                 tail, GEN8_CSB_WRITE_PTR(readl(i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)))), fw ? "" : "?");
 
-                       if (++head == GEN8_CSB_ENTRIES)
-                               head = 0;
+       while (head != tail) {
+               struct i915_request *rq;
+               unsigned int status;
+               unsigned int count;
 
-                       /*
-                        * We are flying near dragons again.
-                        *
-                        * We hold a reference to the request in execlist_port[]
-                        * but no more than that. We are operating in softirq
-                        * context and so cannot hold any mutex or sleep. That
-                        * prevents us stopping the requests we are processing
-                        * in port[] from being retired simultaneously (the
-                        * breadcrumb will be complete before we see the
-                        * context-switch). As we only hold the reference to the
-                        * request, any pointer chasing underneath the request
-                        * is subject to a potential use-after-free. Thus we
-                        * store all of the bookkeeping within port[] as
-                        * required, and avoid using unguarded pointers beneath
-                        * request itself. The same applies to the atomic
-                        * status notifier.
-                        */
+               if (++head == GEN8_CSB_ENTRIES)
+                       head = 0;
 
-                       status = READ_ONCE(buf[2 * head]); /* maybe mmio! */
-                       GEM_TRACE("%s csb[%d]: status=0x%08x:0x%08x, active=0x%x\n",
-                                 engine->name, head,
-                                 status, buf[2*head + 1],
-                                 execlists->active);
-
-                       if (status & (GEN8_CTX_STATUS_IDLE_ACTIVE |
-                                     GEN8_CTX_STATUS_PREEMPTED))
-                               execlists_set_active(execlists,
-                                                    EXECLISTS_ACTIVE_HWACK);
-                       if (status & GEN8_CTX_STATUS_ACTIVE_IDLE)
-                               execlists_clear_active(execlists,
-                                                      EXECLISTS_ACTIVE_HWACK);
-
-                       if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK))
-                               continue;
+               /*
+                * We are flying near dragons again.
+                *
+                * We hold a reference to the request in execlist_port[]
+                * but no more than that. We are operating in softirq
+                * context and so cannot hold any mutex or sleep. That
+                * prevents us stopping the requests we are processing
+                * in port[] from being retired simultaneously (the
+                * breadcrumb will be complete before we see the
+                * context-switch). As we only hold the reference to the
+                * request, any pointer chasing underneath the request
+                * is subject to a potential use-after-free. Thus we
+                * store all of the bookkeeping within port[] as
+                * required, and avoid using unguarded pointers beneath
+                * request itself. The same applies to the atomic
+                * status notifier.
+                */
 
-                       /* We should never get a COMPLETED | IDLE_ACTIVE! */
-                       GEM_BUG_ON(status & GEN8_CTX_STATUS_IDLE_ACTIVE);
+               status = READ_ONCE(buf[2 * head]); /* maybe mmio! */
+               GEM_TRACE("%s csb[%d]: status=0x%08x:0x%08x, active=0x%x\n",
+                         engine->name, head,
+                         status, buf[2*head + 1],
+                         execlists->active);
+
+               if (status & (GEN8_CTX_STATUS_IDLE_ACTIVE |
+                             GEN8_CTX_STATUS_PREEMPTED))
+                       execlists_set_active(execlists,
+                                            EXECLISTS_ACTIVE_HWACK);
+               if (status & GEN8_CTX_STATUS_ACTIVE_IDLE)
+                       execlists_clear_active(execlists,
+                                              EXECLISTS_ACTIVE_HWACK);
+
+               if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK))
+                       continue;
 
-                       if (status & GEN8_CTX_STATUS_COMPLETE &&
-                           buf[2*head + 1] == execlists->preempt_complete_status) {
-                               GEM_TRACE("%s preempt-idle\n", engine->name);
-                               complete_preempt_context(execlists);
-                               continue;
-                       }
+               /* We should never get a COMPLETED | IDLE_ACTIVE! */
+               GEM_BUG_ON(status & GEN8_CTX_STATUS_IDLE_ACTIVE);
 
-                       if (status & GEN8_CTX_STATUS_PREEMPTED &&
-                           execlists_is_active(execlists,
-                                               EXECLISTS_ACTIVE_PREEMPT))
-                               continue;
+               if (status & GEN8_CTX_STATUS_COMPLETE &&
+                   buf[2*head + 1] == execlists->preempt_complete_status) {
+                       GEM_TRACE("%s preempt-idle\n", engine->name);
+                       complete_preempt_context(execlists);
+                       continue;
+               }
 
-                       GEM_BUG_ON(!execlists_is_active(execlists,
-                                                       EXECLISTS_ACTIVE_USER));
+               if (status & GEN8_CTX_STATUS_PREEMPTED &&
+                   execlists_is_active(execlists,
+                                       EXECLISTS_ACTIVE_PREEMPT))
+                       continue;
 
-                       rq = port_unpack(port, &count);
-                       GEM_TRACE("%s out[0]: ctx=%d.%d, global=%d (fence %llx:%d) (current %d), prio=%d\n",
-                                 engine->name,
-                                 port->context_id, count,
-                                 rq ? rq->global_seqno : 0,
-                                 rq ? rq->fence.context : 0,
-                                 rq ? rq->fence.seqno : 0,
-                                 intel_engine_get_seqno(engine),
-                                 rq ? rq_prio(rq) : 0);
+               GEM_BUG_ON(!execlists_is_active(execlists,
+                                               EXECLISTS_ACTIVE_USER));
 
-                       /* Check the context/desc id for this event matches */
-                       GEM_DEBUG_BUG_ON(buf[2 * head + 1] != port->context_id);
+               rq = port_unpack(port, &count);
+               GEM_TRACE("%s out[0]: ctx=%d.%d, global=%d (fence %llx:%d) (current %d), prio=%d\n",
+                         engine->name,
+                         port->context_id, count,
+                         rq ? rq->global_seqno : 0,
+                         rq ? rq->fence.context : 0,
+                         rq ? rq->fence.seqno : 0,
+                         intel_engine_get_seqno(engine),
+                         rq ? rq_prio(rq) : 0);
+
+               /* Check the context/desc id for this event matches */
+               GEM_DEBUG_BUG_ON(buf[2 * head + 1] != port->context_id);
+
+               GEM_BUG_ON(count == 0);
+               if (--count == 0) {
+                       /*
+                        * On the final event corresponding to the
+                        * submission of this context, we expect either
+                        * an element-switch event or a completion
+                        * event (and on completion, the active-idle
+                        * marker). No more preemptions, lite-restore
+                        * or otherwise.
+                        */
+                       GEM_BUG_ON(status & GEN8_CTX_STATUS_PREEMPTED);
+                       GEM_BUG_ON(port_isset(&port[1]) &&
+                                  !(status & GEN8_CTX_STATUS_ELEMENT_SWITCH));
+                       GEM_BUG_ON(!port_isset(&port[1]) &&
+                                  !(status & GEN8_CTX_STATUS_ACTIVE_IDLE));
 
-                       GEM_BUG_ON(count == 0);
-                       if (--count == 0) {
-                               /*
-                                * On the final event corresponding to the
-                                * submission of this context, we expect either
-                                * an element-switch event or a completion
-                                * event (and on completion, the active-idle
-                                * marker). No more preemptions, lite-restore
-                                * or otherwise.
-                                */
-                               GEM_BUG_ON(status & GEN8_CTX_STATUS_PREEMPTED);
-                               GEM_BUG_ON(port_isset(&port[1]) &&
-                                          !(status & GEN8_CTX_STATUS_ELEMENT_SWITCH));
-                               GEM_BUG_ON(!port_isset(&port[1]) &&
-                                          !(status & GEN8_CTX_STATUS_ACTIVE_IDLE));
+                       /*
+                        * We rely on the hardware being strongly
+                        * ordered, that the breadcrumb write is
+                        * coherent (visible from the CPU) before the
+                        * user interrupt and CSB is processed.
+                        */
+                       GEM_BUG_ON(!i915_request_completed(rq));
 
-                               /*
-                                * We rely on the hardware being strongly
-                                * ordered, that the breadcrumb write is
-                                * coherent (visible from the CPU) before the
-                                * user interrupt and CSB is processed.
-                                */
-                               GEM_BUG_ON(!i915_request_completed(rq));
-
-                               execlists_context_schedule_out(rq,
-                                                              INTEL_CONTEXT_SCHEDULE_OUT);
-                               i915_request_put(rq);
-
-                               GEM_TRACE("%s completed ctx=%d\n",
-                                         engine->name, port->context_id);
-
-                               port = execlists_port_complete(execlists, port);
-                               if (port_isset(port))
-                                       execlists_user_begin(execlists, port);
-                               else
-                                       execlists_user_end(execlists);
-                       } else {
-                               port_set(port, port_pack(rq, count));
-                       }
-               }
+                       execlists_context_schedule_out(rq,
+                                                      INTEL_CONTEXT_SCHEDULE_OUT);
+                       i915_request_put(rq);
 
-               if (head != execlists->csb_head) {
-                       execlists->csb_head = head;
-                       writel(_MASKED_FIELD(GEN8_CSB_READ_PTR_MASK, head << 8),
-                              i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)));
+                       GEM_TRACE("%s completed ctx=%d\n",
+                                 engine->name, port->context_id);
+
+                       port = execlists_port_complete(execlists, port);
+                       if (port_isset(port))
+                               execlists_user_begin(execlists, port);
+                       else
+                               execlists_user_end(execlists);
+               } else {
+                       port_set(port, port_pack(rq, count));
                }
-       } while (test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted));
+       }
+
+       if (head != execlists->csb_head) {
+               execlists->csb_head = head;
+               writel(_MASKED_FIELD(GEN8_CSB_READ_PTR_MASK, head << 8),
+                      i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)));
+       }
 
        if (unlikely(fw))
                intel_uncore_forcewake_put(i915, execlists->fw_domains);