| Message ID | 1533815333-17320-3-git-send-email-thunder.leizhen@huawei.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | bugfix and optimization about CMD_SYNC | expand |
On 09/08/18 12:48, Zhen Lei wrote: > More than two CMD_SYNCs maybe adjacent in the command queue, and the first > one has done what others want to do. Drop the redundant CMD_SYNCs can > improve IO performance especially under the pressure scene. > > I did the statistics in my test environment, the number of CMD_SYNCs can > be reduced about 1/3. See below: > CMD_SYNCs reduced: 19542181 > CMD_SYNCs total: 58098548 (include reduced) > CMDs total: 116197099 (TLBI:SYNC about 1:1) > > Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> > --- > drivers/iommu/arm-smmu-v3.c | 22 +++++++++++++++++++--- > 1 file changed, 19 insertions(+), 3 deletions(-) > > diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c > index d17a9a7..b96d2d2 100644 > --- a/drivers/iommu/arm-smmu-v3.c > +++ b/drivers/iommu/arm-smmu-v3.c > @@ -567,6 +567,7 @@ struct arm_smmu_device { > int gerr_irq; > int combined_irq; > u32 sync_nr; > + u8 prev_cmd_opcode; > > unsigned long ias; /* IPA */ > unsigned long oas; /* PA */ > @@ -775,6 +776,11 @@ static int queue_remove_raw(struct arm_smmu_queue *q, u64 *ent) > return 0; > } > > +static inline u8 arm_smmu_cmd_opcode_get(u64 *cmd) > +{ > + return cmd[0] & CMDQ_0_OP; > +} > + > /* High-level queue accessors */ > static int arm_smmu_cmdq_build_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent) > { > @@ -900,6 +906,8 @@ static void arm_smmu_cmdq_insert_cmd(struct arm_smmu_device *smmu, u64 *cmd) > struct arm_smmu_queue *q = &smmu->cmdq.q; > bool wfe = !!(smmu->features & ARM_SMMU_FEAT_SEV); > > + smmu->prev_cmd_opcode = arm_smmu_cmd_opcode_get(cmd); > + > while (queue_insert_raw(q, cmd) == -ENOSPC) { > if (queue_poll_cons(q, false, wfe)) > dev_err_ratelimited(smmu->dev, "CMDQ timeout\n"); > @@ -952,9 +960,17 @@ static int __arm_smmu_cmdq_issue_sync_msi(struct arm_smmu_device *smmu) > }; > > spin_lock_irqsave(&smmu->cmdq.lock, flags); > - ent.sync.msidata = ++smmu->sync_nr; > - arm_smmu_cmdq_build_cmd(cmd, &ent); > - arm_smmu_cmdq_insert_cmd(smmu, cmd); > + if (smmu->prev_cmd_opcode == CMDQ_OP_CMD_SYNC) { > + /* > + * Previous command is CMD_SYNC also, there is no need to add > + * one more. Just poll it. > + */ > + ent.sync.msidata = smmu->sync_nr; Aha! at the time I had pondered how to make multiple callers wait on a previous sync instead of issuing another back-to-back, but it seemed complicated precisely *because* of the counter being updated outside the lock. If only I'd realised... :) Now I just need to figure out if we can do the same for the polling case. Robin. > + } else { > + ent.sync.msidata = ++smmu->sync_nr; > + arm_smmu_cmdq_build_cmd(cmd, &ent); > + arm_smmu_cmdq_insert_cmd(smmu, cmd); > + } > spin_unlock_irqrestore(&smmu->cmdq.lock, flags); > > return __arm_smmu_sync_poll_msi(smmu, ent.sync.msidata); > -- > 1.8.3 > >
diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c index d17a9a7..b96d2d2 100644 --- a/drivers/iommu/arm-smmu-v3.c +++ b/drivers/iommu/arm-smmu-v3.c @@ -567,6 +567,7 @@ struct arm_smmu_device { int gerr_irq; int combined_irq; u32 sync_nr; + u8 prev_cmd_opcode; unsigned long ias; /* IPA */ unsigned long oas; /* PA */ @@ -775,6 +776,11 @@ static int queue_remove_raw(struct arm_smmu_queue *q, u64 *ent) return 0; } +static inline u8 arm_smmu_cmd_opcode_get(u64 *cmd) +{ + return cmd[0] & CMDQ_0_OP; +} + /* High-level queue accessors */ static int arm_smmu_cmdq_build_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent) { @@ -900,6 +906,8 @@ static void arm_smmu_cmdq_insert_cmd(struct arm_smmu_device *smmu, u64 *cmd) struct arm_smmu_queue *q = &smmu->cmdq.q; bool wfe = !!(smmu->features & ARM_SMMU_FEAT_SEV); + smmu->prev_cmd_opcode = arm_smmu_cmd_opcode_get(cmd); + while (queue_insert_raw(q, cmd) == -ENOSPC) { if (queue_poll_cons(q, false, wfe)) dev_err_ratelimited(smmu->dev, "CMDQ timeout\n"); @@ -952,9 +960,17 @@ static int __arm_smmu_cmdq_issue_sync_msi(struct arm_smmu_device *smmu) }; spin_lock_irqsave(&smmu->cmdq.lock, flags); - ent.sync.msidata = ++smmu->sync_nr; - arm_smmu_cmdq_build_cmd(cmd, &ent); - arm_smmu_cmdq_insert_cmd(smmu, cmd); + if (smmu->prev_cmd_opcode == CMDQ_OP_CMD_SYNC) { + /* + * Previous command is CMD_SYNC also, there is no need to add + * one more. Just poll it. + */ + ent.sync.msidata = smmu->sync_nr; + } else { + ent.sync.msidata = ++smmu->sync_nr; + arm_smmu_cmdq_build_cmd(cmd, &ent); + arm_smmu_cmdq_insert_cmd(smmu, cmd); + } spin_unlock_irqrestore(&smmu->cmdq.lock, flags); return __arm_smmu_sync_poll_msi(smmu, ent.sync.msidata);
More than two CMD_SYNCs maybe adjacent in the command queue, and the first one has done what others want to do. Drop the redundant CMD_SYNCs can improve IO performance especially under the pressure scene. I did the statistics in my test environment, the number of CMD_SYNCs can be reduced about 1/3. See below: CMD_SYNCs reduced: 19542181 CMD_SYNCs total: 58098548 (include reduced) CMDs total: 116197099 (TLBI:SYNC about 1:1) Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> --- drivers/iommu/arm-smmu-v3.c | 22 +++++++++++++++++++--- 1 file changed, 19 insertions(+), 3 deletions(-) -- 1.8.3