| Message ID | 1534665071-7976-3-git-send-email-thunder.leizhen@huawei.com |
|---|---|
| State | New |
| Headers | show |
| Series | bugfix and optimization about CMD_SYNC | expand |
On 19/08/2018 08:51, 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 ac6d6df..f3a56e1 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 */ > @@ -786,6 +787,11 @@ void arm_smmu_cmdq_build_sync_msi_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent) > cmd[1] = ent->sync.msiaddr & CMDQ_SYNC_1_MSIADDR_MASK; > } > > +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) > { > @@ -906,6 +912,8 @@ 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"); > @@ -958,9 +966,17 @@ 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_sync_msi_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_sync_msi_cmd(cmd, &ent); > + arm_smmu_cmdq_insert_cmd(smmu, cmd); > + } > spin_unlock_irqrestore(&smmu->cmdq.lock, flags); I find something like this adds support for combining CMD_SYNC commands for regular polling mode: @@ -569,6 +569,7 @@ struct arm_smmu_device { int combined_irq; u32 sync_nr; u8 prev_cmd_opcode; + int prev_cmd_sync_res; unsigned long ias; /* IPA */ unsigned long oas; /* PA */ @@ -985,17 +986,33 @@ static int __arm_smmu_cmdq_issue_sync_msi(struct arm_smmu_device *smmu) static int __arm_smmu_cmdq_issue_sync(struct arm_smmu_device *smmu) { - u64 cmd[CMDQ_ENT_DWORDS]; + static u64 cmd[CMDQ_ENT_DWORDS] = { + _FIELD_PREP(CMDQ_0_OP, CMDQ_OP_CMD_SYNC) | + _FIELD_PREP(CMDQ_SYNC_0_CS, CMDQ_SYNC_0_CS_SEV) | + _FIELD_PREP(CMDQ_SYNC_0_MSH, ARM_SMMU_SH_ISH) | + _FIELD_PREP(CMDQ_SYNC_0_MSIATTR, ARM_SMMU_MEMATTR_OIWB) + }; unsigned long flags; bool wfe = !!(smmu->features & ARM_SMMU_FEAT_SEV); - struct arm_smmu_cmdq_ent ent = { .opcode = CMDQ_OP_CMD_SYNC }; - int ret; + int ret = 0; - arm_smmu_cmdq_build_cmd(cmd, &ent); spin_lock_irqsave(&smmu->cmdq.lock, flags); - arm_smmu_cmdq_insert_cmd(smmu, cmd); - ret = queue_poll_cons(&smmu->cmdq.q, true, wfe); + if (smmu->prev_cmd_opcode != CMDQ_OP_CMD_SYNC || + smmu->prev_cmd_sync_res != 0) { + arm_smmu_cmdq_insert_cmd(smmu, cmd); + smmu->prev_cmd_sync_res = ret = + queue_poll_cons(&smmu->cmdq.q, true, wfe); + } I tested iperf on a 1G network link and was seeing 6-10% CMD_SYNC commands combined. I would really need to test this on a faster connection to see any throughout difference. From the above figures, I think leizhen was seeing 25% combine rate, right? As for this code, it could be neatened... Cheers, John > > return __arm_smmu_sync_poll_msi(smmu, ent.sync.msidata); > -- > 1.8.3 > > > > . >
On 2018/8/30 19:18, John Garry wrote: > On 19/08/2018 08:51, Zhen Lei wrote: >> spin_unlock_irqrestore(&smmu->cmdq.lock, flags); > > I find something like this adds support for combining CMD_SYNC commands for regular polling mode: > > @@ -569,6 +569,7 @@ struct arm_smmu_device { > int combined_irq; > u32 sync_nr; > u8 prev_cmd_opcode; > + int prev_cmd_sync_res; > > unsigned long ias; /* IPA */ > unsigned long oas; /* PA */ > @@ -985,17 +986,33 @@ static int __arm_smmu_cmdq_issue_sync_msi(struct arm_smmu_device *smmu) > > static int __arm_smmu_cmdq_issue_sync(struct arm_smmu_device *smmu) > { > - u64 cmd[CMDQ_ENT_DWORDS]; > + static u64 cmd[CMDQ_ENT_DWORDS] = { > + _FIELD_PREP(CMDQ_0_OP, CMDQ_OP_CMD_SYNC) | > + _FIELD_PREP(CMDQ_SYNC_0_CS, CMDQ_SYNC_0_CS_SEV) | > + _FIELD_PREP(CMDQ_SYNC_0_MSH, ARM_SMMU_SH_ISH) | > + _FIELD_PREP(CMDQ_SYNC_0_MSIATTR, ARM_SMMU_MEMATTR_OIWB) > + }; > unsigned long flags; > bool wfe = !!(smmu->features & ARM_SMMU_FEAT_SEV); > - struct arm_smmu_cmdq_ent ent = { .opcode = CMDQ_OP_CMD_SYNC }; > - int ret; > + int ret = 0; > > - arm_smmu_cmdq_build_cmd(cmd, &ent); > > spin_lock_irqsave(&smmu->cmdq.lock, flags); > - arm_smmu_cmdq_insert_cmd(smmu, cmd); > - ret = queue_poll_cons(&smmu->cmdq.q, true, wfe); > + if (smmu->prev_cmd_opcode != CMDQ_OP_CMD_SYNC || > + smmu->prev_cmd_sync_res != 0) { > + arm_smmu_cmdq_insert_cmd(smmu, cmd); > + smmu->prev_cmd_sync_res = ret = > + queue_poll_cons(&smmu->cmdq.q, true, wfe); > + } > > I tested iperf on a 1G network link and was seeing 6-10% CMD_SYNC commands combined. I would really need to test this on a faster connection to see any throughout difference. > > From the above figures, I think leizhen was seeing 25% combine rate, right? Yes. In my test case, the size of unmap are almost one page, that means 1 TLBI follows 1 SYNC, so the probability that two CMD_SYNCs next to each other will be greater. > > As for this code, it could be neatened... > > Cheers, > John > >> >> return __arm_smmu_sync_poll_msi(smmu, ent.sync.msidata); >> -- >> 1.8.3 >> >> >> >> . >> > > > > . > -- Thanks! BestRegards
diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c index ac6d6df..f3a56e1 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 */ @@ -786,6 +787,11 @@ void arm_smmu_cmdq_build_sync_msi_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent) cmd[1] = ent->sync.msiaddr & CMDQ_SYNC_1_MSIADDR_MASK; } +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) { @@ -906,6 +912,8 @@ 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"); @@ -958,9 +966,17 @@ 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_sync_msi_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_sync_msi_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