+ /*
+ * We can try to avoid the cmpxchg() loop by simply incrementing the
+ * lock counter. When held in exclusive state, the lock counter is set
+ * to INT_MIN so these increments won't hurt as the value will remain
+ * negative.
+ */
+ if (atomic_fetch_inc_relaxed(&cmdq->lock) >= 0)
+ return;
+
+ do {
+ val = atomic_cond_read_relaxed(&cmdq->lock, VAL >= 0);
+ } while (atomic_cmpxchg_relaxed(&cmdq->lock, val, val + 1) != val);
+}
+
+static void arm_smmu_cmdq_shared_unlock(struct arm_smmu_cmdq *cmdq)
+{
+ (void)atomic_dec_return_release(&cmdq->lock);
+}
+
+static bool arm_smmu_cmdq_shared_tryunlock(struct arm_smmu_cmdq *cmdq)
+{
+ if (atomic_read(&cmdq->lock) == 1)
+ return false;
+
+ arm_smmu_cmdq_shared_unlock(cmdq);
+ return true;
+}
+
+#define arm_smmu_cmdq_exclusive_trylock_irqsave(cmdq, flags) \
+({ \
+ bool __ret; \
+ local_irq_save(flags); \
+ __ret = !atomic_cmpxchg_relaxed(&cmdq->lock, 0, INT_MIN); \
+ if (!__ret) \
+ local_irq_restore(flags); \
+ __ret; \
+})
+
+#define arm_smmu_cmdq_exclusive_unlock_irqrestore(cmdq, flags) \
+({ \
+ atomic_set_release(&cmdq->lock, 0); \
+ local_irq_restore(flags); \
+})
+
+
+/*
+ * Command queue insertion.
+ * This is made fiddly by our attempts to achieve some sort of scalability
+ * since there is one queue shared amongst all of the CPUs in the system. If
+ * you like mixed-size concurrency, dependency ordering and relaxed atomics,
+ * then you'll *love* this monstrosity.
+ *
+ * The basic idea is to split the queue up into ranges of commands that are
+ * owned by a given CPU; the owner may not have written all of the commands
+ * itself, but is responsible for advancing the hardware prod pointer when
+ * the time comes. The algorithm is roughly:
+ *
+ * 1. Allocate some space in the queue. At this point we also discover
+ * whether the head of the queue is currently owned by another CPU,
+ * or whether we are the owner.
+ *
+ * 2. Write our commands into our allocated slots in the queue.
+ *
+ * 3. Mark our slots as valid in arm_smmu_cmdq.valid_map.
+ *
+ * 4. If we are an owner:
+ * a. Wait for the previous owner to finish.
+ * b. Mark the queue head as unowned, which tells us the range
+ * that we are responsible for publishing.
+ * c. Wait for all commands in our owned range to become valid.
+ * d. Advance the hardware prod pointer.
+ * e. Tell the next owner we've finished.
+ *
+ * 5. If we are inserting a CMD_SYNC (we may or may not have been an
+ * owner), then we need to stick around until it has completed:
+ * a. If we have MSIs, the SMMU can write back into the CMD_SYNC
+ * to clear the first 4 bytes.
+ * b. Otherwise, we spin waiting for the hardware cons pointer to
+ * advance past our command.
+ *
+ * The devil is in the details, particularly the use of locking for handling
+ * SYNC completion and freeing up space in the queue before we think that it is
+ * full.
+ */
+static void __arm_smmu_cmdq_poll_set_valid_map(struct arm_smmu_cmdq *cmdq,
+ u32 sprod, u32 eprod, bool set)
+{
+ u32 swidx, sbidx, ewidx, ebidx;
+ struct arm_smmu_ll_queue llq = {
+ .max_n_shift = cmdq->q.llq.max_n_shift,
+ .prod = sprod,
+ };
projects / linux.git / commitdiff