Skip to contentSkip to conversation

QCOM smmu-500 wait-for-safe handling for sdm845

PATCH Superseded Sai Prakash Ranjan 3 participants 14 messages Sep 17, 2019 – Sep 20, 2019

Subsystems (2)

v6 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

v6 cover QCOM smmu-500 wait-for-safe handling for sdm845

Previous version of the patches are at [1]:

QCOM's implementation of smmu-500 on sdm845 adds a hardware logic called wait-for-safe. This logic helps in meeting the invalidation requirements from 'real-time clients', such as display and camera. This wait-for-safe logic ensures that the invalidations happen after getting an ack from these devices.
In this patch-series we are disabling this wait-for-safe logic from the arm-smmu driver's probe as with this enabled the hardware tries to throttle invalidations from 'non-real-time clients', such as USB and UFS.

For detailed information please refer to patch [3/4] in this series. I have included the device tree patch too in this series for someone who would like to test out this. Here's a branch [2] that gets display on MTP SDM845 device.

This patch series is inspired from downstream work to handle under-performance issues on real-time clients on sdm845. In downstream we add separate page table ops to handle TLB maintenance and toggle wait-for-safe in tlb_sync call so that achieve required performance for display and camera [3, 4].

Changes since v5:
* Addressed Robin's comments and removed unwanted header file.

Changes since v4:
* Addressed Stephen's comments.
* Moved QCOM specific implementation to arm-smmu-qcom.c as per Robin's suggestion.

Changes since v3:
* Based on arm-smmu implementation cleanup series [5] by Robin Murphy which is already merged in Will's tree [6].
* Implemented the sdm845 specific reset hook which does arm_smmu_device_reset() followed by making SCM call to disable the wait-for-safe logic.
* Removed depedency for SCM call on any dt flag. We invariably try to disable the wait-for-safe logic on sdm845. The platforms such as mtp845, and db845 that implement handlers for this particular SCM call should be able disable wait-for-safe logic.
Other platforms such as cheza don't enable the wait-for-safe logic at all from their bootloaders. So there's no need to disable the same.
* No change in SCM call patches 1 & 2.

Changes since v2:
* Dropped the patch to add atomic io_read/write scm API.
* Removed support for any separate page table ops to handle wait-for-safe. Currently just disabling this wait-for-safe logic from arm_smmu_device_probe() to achieve performance on USB/UFS on sdm845.
* Added a device tree patch to add smmu option for fw-implemented support for SCM call to take care of SAFE toggling.

Changes since v1:
* Addressed Will and Robin's comments:
- Dropped the patch[4] that forked out __arm_smmu_tlb_inv_range_nosync(), and __arm_smmu_tlb_sync().
- Cleaned up the errata patch further to use downstream polling mechanism for tlb sync.
* No change in SCM call patches - patches 1 to 3.

[1] https://lore.kernel.org/patchwork/cover/1127798/
[2] https://github.com/vivekgautam1/linux/tree/v5.2-rc4/sdm845-display-working
[3] https://source.codeaurora.org/quic/la/kernel/msm-4.9/commit/drivers/iommu/arm-smmu.c?h=CogSystems-msm-49/msm-4.9&id=da765c6c75266b38191b38ef086274943f353ea7
[4] https://source.codeaurora.org/quic/la/kernel/msm-4.9/commit/drivers/iommu/arm-smmu.c?h=CogSystems-msm-49/msm-4.9&id=8696005aaaf745de68f57793c1a534a34345c30a
[5] https://patchwork.kernel.org/patch/11096265/
[6] https://git.kernel.org/pub/scm/linux/kernel/git/will/linux.git/

Vivek Gautam (3):
firmware: qcom_scm-64: Add atomic version of qcom_scm_call firmware/qcom_scm: Add scm call to handle smmu errata
iommu: arm-smmu-impl: Add sdm845 implementation hook

 drivers/firmware/qcom_scm-32.c |   5 ++
 drivers/firmware/qcom_scm-64.c | 152 +++++++++++++++++++++++----------
 drivers/firmware/qcom_scm.c    |   6 ++
 drivers/firmware/qcom_scm.h    |   5 ++
 drivers/iommu/Makefile         |   2 +-
 drivers/iommu/arm-smmu-impl.c  |   6 +-
 drivers/iommu/arm-smmu-qcom.c  |  51 +++++++++++
 drivers/iommu/arm-smmu.h       |   3 +
 include/linux/qcom_scm.h       |   2 +
 9 files changed, 185 insertions(+), 47 deletions(-)
 create mode 100644 drivers/iommu/arm-smmu-qcom.c
Signature · 7 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
#1 lore raw

PATCH v6 1/3 firmware: qcom_scm-64: Add atomic version of qcom_scm_call

From: Vivek Gautam <[email protected]>

There are scnenarios where drivers are required to make a scm call in atomic context, such as in one of the qcom's arm-smmu-500 errata [1].

[1] ("https://source.codeaurora.org/quic/la/kernel/msm-4.9/

      tree/drivers/iommu/arm-smmu.c?h=msm-4.9#n4842")
---
 drivers/firmware/qcom_scm-64.c | 139 ++++++++++++++++++++++-----------
 1 file changed, 95 insertions(+), 44 deletions(-)
drivers/firmware/qcom_scm-64.c +95 -44
index 91d5ad7cf58b..6697bb741b83 100644
@@ -62,32 +62,72 @@ static DEFINE_MUTEX(qcom_scm_lock);
 #define FIRST_EXT_ARG_IDX 3
 #define N_REGISTER_ARGS (MAX_QCOM_SCM_ARGS - N_EXT_QCOM_SCM_ARGS + 1)
 
-/**
- * qcom_scm_call() - Invoke a syscall in the secure world
- * @dev:	device
- * @svc_id:	service identifier
- * @cmd_id:	command identifier
- * @desc:	Descriptor structure containing arguments and return values
- *
- * Sends a command to the SCM and waits for the command to finish processing.
- * This should *only* be called in pre-emptible context.
-*/
-static int qcom_scm_call(struct device *dev, u32 svc_id, u32 cmd_id,
-			 const struct qcom_scm_desc *desc,
-			 struct arm_smccc_res *res)
+static void __qcom_scm_call_do(const struct qcom_scm_desc *desc,
+			       struct arm_smccc_res *res, u32 fn_id,
+			       u64 x5, u32 type)
+{
+	u64 cmd;
+	struct arm_smccc_quirk quirk = { .id = ARM_SMCCC_QUIRK_QCOM_A6 };
+
+	cmd = ARM_SMCCC_CALL_VAL(type, qcom_smccc_convention,
+				 ARM_SMCCC_OWNER_SIP, fn_id);
+
+	quirk.state.a6 = 0;
+
+	do {
+		arm_smccc_smc_quirk(cmd, desc->arginfo, desc->args[0],
+				    desc->args[1], desc->args[2], x5,
+				    quirk.state.a6, 0, res, &quirk);
+
+		if (res->a0 == QCOM_SCM_INTERRUPTED)
+			cmd = res->a0;
+
+	} while (res->a0 == QCOM_SCM_INTERRUPTED);
+}
+
+static void qcom_scm_call_do(const struct qcom_scm_desc *desc,
+			     struct arm_smccc_res *res, u32 fn_id,
+			     u64 x5, bool atomic)
+{
+	int retry_count = 0;
+
+	if (atomic) {
+		__qcom_scm_call_do(desc, res, fn_id, x5, ARM_SMCCC_FAST_CALL);
+		return;
+	}
+
+	do {
+		mutex_lock(&qcom_scm_lock);
+
+		__qcom_scm_call_do(desc, res, fn_id, x5,
+				   ARM_SMCCC_STD_CALL);
+
+		mutex_unlock(&qcom_scm_lock);
+
+		if (res->a0 == QCOM_SCM_V2_EBUSY) {
+			if (retry_count++ > QCOM_SCM_EBUSY_MAX_RETRY)
+				break;
+			msleep(QCOM_SCM_EBUSY_WAIT_MS);
+		}
+	}  while (res->a0 == QCOM_SCM_V2_EBUSY);
+}
+
+static int ___qcom_scm_call(struct device *dev, u32 svc_id, u32 cmd_id,
+			    const struct qcom_scm_desc *desc,
+			    struct arm_smccc_res *res, bool atomic)
 {
 	int arglen = desc->arginfo & 0xf;
-	int retry_count = 0, i;
+	int i;
 	u32 fn_id = QCOM_SCM_FNID(svc_id, cmd_id);
-	u64 cmd, x5 = desc->args[FIRST_EXT_ARG_IDX];
+	u64 x5 = desc->args[FIRST_EXT_ARG_IDX];
 	dma_addr_t args_phys = 0;
 	void *args_virt = NULL;
 	size_t alloc_len;
-	struct arm_smccc_quirk quirk = {.id = ARM_SMCCC_QUIRK_QCOM_A6};
+	gfp_t flag = atomic ? GFP_ATOMIC : GFP_KERNEL;
 
 	if (unlikely(arglen > N_REGISTER_ARGS)) {
 		alloc_len = N_EXT_QCOM_SCM_ARGS * sizeof(u64);
-		args_virt = kzalloc(PAGE_ALIGN(alloc_len), GFP_KERNEL);
+		args_virt = kzalloc(PAGE_ALIGN(alloc_len), flag);
 
 		if (!args_virt)
 			return -ENOMEM;
@@ -117,33 +157,7 @@ static int qcom_scm_call(struct device *dev, u32 svc_id, u32 cmd_id,
 		x5 = args_phys;
 	}
 
-	do {
-		mutex_lock(&qcom_scm_lock);
-
-		cmd = ARM_SMCCC_CALL_VAL(ARM_SMCCC_STD_CALL,
-					 qcom_smccc_convention,
-					 ARM_SMCCC_OWNER_SIP, fn_id);
-
-		quirk.state.a6 = 0;
-
-		do {
-			arm_smccc_smc_quirk(cmd, desc->arginfo, desc->args[0],
-				      desc->args[1], desc->args[2], x5,
-				      quirk.state.a6, 0, res, &quirk);
-
-			if (res->a0 == QCOM_SCM_INTERRUPTED)
-				cmd = res->a0;
-
-		} while (res->a0 == QCOM_SCM_INTERRUPTED);
-
-		mutex_unlock(&qcom_scm_lock);
-
-		if (res->a0 == QCOM_SCM_V2_EBUSY) {
-			if (retry_count++ > QCOM_SCM_EBUSY_MAX_RETRY)
-				break;
-			msleep(QCOM_SCM_EBUSY_WAIT_MS);
-		}
-	}  while (res->a0 == QCOM_SCM_V2_EBUSY);
+	qcom_scm_call_do(desc, res, fn_id, x5, atomic);
 
 	if (args_virt) {
 		dma_unmap_single(dev, args_phys, alloc_len, DMA_TO_DEVICE);
@@ -156,6 +170,43 @@ static int qcom_scm_call(struct device *dev, u32 svc_id, u32 cmd_id,
 	return 0;
 }
 
+/**
+ * qcom_scm_call() - Invoke a syscall in the secure world
+ * @dev:	device
+ * @svc_id:	service identifier
+ * @cmd_id:	command identifier
+ * @desc:	Descriptor structure containing arguments and return values
+ *
+ * Sends a command to the SCM and waits for the command to finish processing.
+ * This should *only* be called in pre-emptible context.
+ */
+static int qcom_scm_call(struct device *dev, u32 svc_id, u32 cmd_id,
+			 const struct qcom_scm_desc *desc,
+			 struct arm_smccc_res *res)
+{
+	might_sleep();
+	return ___qcom_scm_call(dev, svc_id, cmd_id, desc, res, false);
+}
+
+/**
+ * qcom_scm_call_atomic() - atomic variation of qcom_scm_call()
+ * @dev:	device
+ * @svc_id:	service identifier
+ * @cmd_id:	command identifier
+ * @desc:	Descriptor structure containing arguments and return values
+ * @res:	Structure containing results from SMC/HVC call
+ *
+ * Sends a command to the SCM and waits for the command to finish processing.
+ * This should be called in atomic context only.
+ */
+static int qcom_scm_call_atomic(struct device *dev, u32 svc_id, u32 cmd_id,
+				const struct qcom_scm_desc *desc,
+				struct arm_smccc_res *res)
+{
+	cant_sleep();
+	return ___qcom_scm_call(dev, svc_id, cmd_id, desc, res, true);
+}
+
 /**
  * qcom_scm_set_cold_boot_addr() - Set the cold boot address for cpus
  * @entry: Entry point function for the cpus
Signature · 7 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Reviewed-by on v6 1/3
#2 lore raw

Re: [PATCHv6 1/3] firmware: qcom_scm-64: Add atomic version of qcom_scm_call

Quoting Sai Prakash Ranjan (2019-09-17 02:45:02)

Quoted 6 lines

From: Vivek Gautam <[email protected]>

There are scnenarios where drivers are required to make a scm call in atomic context, such as in one of the qcom's arm-smmu-500 errata [1].

Quoted 7 lines

[1] ("https://source.codeaurora.org/quic/la/kernel/msm-4.9/

      tree/drivers/iommu/arm-smmu.c?h=msm-4.9#n4842")

---

#3 lore raw

PATCH v6 2/3 firmware/qcom_scm: Add scm call to handle smmu errata

In reply to Sai Prakash Ranjan

From: Vivek Gautam <[email protected]>

Qcom's smmu-500 needs to toggle wait-for-safe sequence to handle TLB invalidation sync's.
Few firmwares allow doing that through SCM interface. Add API to toggle wait for safe from firmware through a SCM call.

---
 drivers/firmware/qcom_scm-32.c |  5 +++++
 drivers/firmware/qcom_scm-64.c | 13 +++++++++++++
 drivers/firmware/qcom_scm.c    |  6 ++++++
 drivers/firmware/qcom_scm.h    |  5 +++++
 include/linux/qcom_scm.h       |  2 ++
 5 files changed, 31 insertions(+)
drivers/firmware/qcom_scm-32.c +5 -0
index 215061c581e1..bee8729525ec 100644
@@ -614,3 +614,8 @@ int __qcom_scm_io_writel(struct device *dev, phys_addr_t addr, unsigned int val)
 	return qcom_scm_call_atomic2(QCOM_SCM_SVC_IO, QCOM_SCM_IO_WRITE,
 				     addr, val);
 }
+
+int __qcom_scm_qsmmu500_wait_safe_toggle(struct device *dev, bool enable)
+{
+	return -ENODEV;
+}
drivers/firmware/qcom_scm-64.c +13 -0
index 6697bb741b83..2f1b7d1660ff 100644
@@ -553,3 +553,16 @@ int __qcom_scm_io_writel(struct device *dev, phys_addr_t addr, unsigned int val)
 	return qcom_scm_call(dev, QCOM_SCM_SVC_IO, QCOM_SCM_IO_WRITE,
 			     &desc, &res);
 }
+
+int __qcom_scm_qsmmu500_wait_safe_toggle(struct device *dev, bool en)
+{
+	struct qcom_scm_desc desc = {0};
+	struct arm_smccc_res res;
+
+	desc.args[0] = QCOM_SCM_CONFIG_ERRATA1_CLIENT_ALL;
+	desc.args[1] = en;
+	desc.arginfo = QCOM_SCM_ARGS(2);
+
+	return qcom_scm_call_atomic(dev, QCOM_SCM_SVC_SMMU_PROGRAM,
+				    QCOM_SCM_CONFIG_ERRATA1, &desc, &res);
+}
drivers/firmware/qcom_scm.c +6 -0
index 2ddc118dba1b..2b3b7a8c4270 100644
@@ -344,6 +344,12 @@ int qcom_scm_iommu_secure_ptbl_init(u64 addr, u32 size, u32 spare)
 }
 EXPORT_SYMBOL(qcom_scm_iommu_secure_ptbl_init);
 
+int qcom_scm_qsmmu500_wait_safe_toggle(bool en)
+{
+	return __qcom_scm_qsmmu500_wait_safe_toggle(__scm->dev, en);
+}
+EXPORT_SYMBOL(qcom_scm_qsmmu500_wait_safe_toggle);
+
 int qcom_scm_io_readl(phys_addr_t addr, unsigned int *val)
 {
 	return __qcom_scm_io_readl(__scm->dev, addr, val);
drivers/firmware/qcom_scm.h +5 -0
index 99506bd873c0..baee744dbcfe 100644
@@ -91,10 +91,15 @@ extern int __qcom_scm_restore_sec_cfg(struct device *dev, u32 device_id,
 				      u32 spare);
 #define QCOM_SCM_IOMMU_SECURE_PTBL_SIZE	3
 #define QCOM_SCM_IOMMU_SECURE_PTBL_INIT	4
+#define QCOM_SCM_SVC_SMMU_PROGRAM	0x15
+#define QCOM_SCM_CONFIG_ERRATA1		0x3
+#define QCOM_SCM_CONFIG_ERRATA1_CLIENT_ALL	0x2
 extern int __qcom_scm_iommu_secure_ptbl_size(struct device *dev, u32 spare,
 					     size_t *size);
 extern int __qcom_scm_iommu_secure_ptbl_init(struct device *dev, u64 addr,
 					     u32 size, u32 spare);
+extern int __qcom_scm_qsmmu500_wait_safe_toggle(struct device *dev,
+						bool enable);
 #define QCOM_MEM_PROT_ASSIGN_ID	0x16
 extern int  __qcom_scm_assign_mem(struct device *dev,
 				  phys_addr_t mem_region, size_t mem_sz,
include/linux/qcom_scm.h +2 -0
index 3f12cc77fb58..aee3d8580d89 100644
@@ -57,6 +57,7 @@ extern int qcom_scm_set_remote_state(u32 state, u32 id);
 extern int qcom_scm_restore_sec_cfg(u32 device_id, u32 spare);
 extern int qcom_scm_iommu_secure_ptbl_size(u32 spare, size_t *size);
 extern int qcom_scm_iommu_secure_ptbl_init(u64 addr, u32 size, u32 spare);
+extern int qcom_scm_qsmmu500_wait_safe_toggle(bool en);
 extern int qcom_scm_io_readl(phys_addr_t addr, unsigned int *val);
 extern int qcom_scm_io_writel(phys_addr_t addr, unsigned int val);
 #else
@@ -96,6 +97,7 @@ qcom_scm_set_remote_state(u32 state,u32 id) { return -ENODEV; }
 static inline int qcom_scm_restore_sec_cfg(u32 device_id, u32 spare) { return -ENODEV; }
 static inline int qcom_scm_iommu_secure_ptbl_size(u32 spare, size_t *size) { return -ENODEV; }
 static inline int qcom_scm_iommu_secure_ptbl_init(u64 addr, u32 size, u32 spare) { return -ENODEV; }
+static inline int qcom_scm_qsmmu500_wait_safe_toggle(bool en) { return -ENODEV; }
 static inline int qcom_scm_io_readl(phys_addr_t addr, unsigned int *val) { return -ENODEV; }
 static inline int qcom_scm_io_writel(phys_addr_t addr, unsigned int val) { return -ENODEV; }
 #endif
Signature · 7 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Reviewed-by on v6 2/3
#4 lore raw

Re: [PATCHv6 2/3] firmware/qcom_scm: Add scm call to handle smmu errata

Quoting Sai Prakash Ranjan (2019-09-17 02:45:03)

Quoted 8 lines

From: Vivek Gautam <[email protected]>

Qcom's smmu-500 needs to toggle wait-for-safe sequence to handle TLB invalidation sync's.
Few firmwares allow doing that through SCM interface. Add API to toggle wait for safe from firmware through a SCM call.

Quoted 4 lines

---

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#5 lore raw

PATCH v6 3/3 iommu: arm-smmu-impl: Add sdm845 implementation hook

In reply to Sai Prakash Ranjan

From: Vivek Gautam <[email protected]>

Add reset hook for sdm845 based platforms to turn off the wait-for-safe sequence.

Understanding how wait-for-safe logic affects USB and UFS performance on MTP845 and DB845 boards:

Qcom's implementation of arm,mmu-500 adds a WAIT-FOR-SAFE logic to address under-performance issues in real-time clients, such as Display, and Camera.
On receiving an invalidation requests, the SMMU forwards SAFE request to these clients and waits for SAFE ack signal from real-time clients. The SAFE signal from such clients is used to qualify the start of invalidation.
This logic is controlled by chicken bits, one for each - MDP (display), IFE0, and IFE1 (camera), that can be accessed only from secure software on sdm845.

This configuration, however, degrades the performance of non-real time clients, such as USB, and UFS etc. This happens because, with wait-for-safe logic enabled the hardware tries to throttle non-real time clients while waiting for SAFE ack signals from real-time clients.

On mtp845 and db845 devices, with wait-for-safe logic enabled by the bootloaders we see degraded performance of USB and UFS when kernel enables the smmu stage-1 translations for these clients. Turn off this wait-for-safe logic from the kernel gets us back the perf of USB and UFS devices until we re-visit this when we start seeing perf issues on display/camera on upstream supported SDM845 platforms. The bootloaders on these boards implement secure monitor callbacks to handle a specific command - QCOM_SCM_SVC_SMMU_PROGRAM with which the logic can be toggled.

There are other boards such as cheza whose bootloaders don't enable this logic. Such boards don't implement callbacks to handle the specific SCM call so disabling this logic for such boards will be a no-op.

This change is inspired by the downstream change from Patrick Daly to address performance issues with display and camera by handling this wait-for-safe within separte io-pagetable ops to do TLB maintenance. So a big thanks to him for the change and for all the offline discussions.

Without this change the UFS reads are pretty slow:

$ time dd if=/dev/sda of=/dev/zero bs=1048576 count=10 conv=sync
10+0 records in
10+0 records out
10485760 bytes (10.0MB) copied, 22.394903 seconds, 457.2KB/s
real    0m 22.39s
user    0m 0.00s
sys     0m 0.01s

With this change they are back to rock!

$ time dd if=/dev/sda of=/dev/zero bs=1048576 count=300 conv=sync
300+0 records in
300+0 records out
314572800 bytes (300.0MB) copied, 1.030541 seconds, 291.1MB/s
real    0m 1.03s
user    0m 0.00s
sys     0m 0.54s
---
 drivers/iommu/Makefile        |  2 +-
 drivers/iommu/arm-smmu-impl.c |  6 +++--
 drivers/iommu/arm-smmu-qcom.c | 51 +++++++++++++++++++++++++++++++++++
 drivers/iommu/arm-smmu.h      |  3 +++
 4 files changed, 59 insertions(+), 3 deletions(-)
 create mode 100644 drivers/iommu/arm-smmu-qcom.c
drivers/iommu/Makefile +1 -1
index a2729aadd300..2816e49a8c46 100644
@@ -13,7 +13,7 @@ obj-$(CONFIG_MSM_IOMMU) += msm_iommu.o
 obj-$(CONFIG_AMD_IOMMU) += amd_iommu.o amd_iommu_init.o
 obj-$(CONFIG_AMD_IOMMU_DEBUGFS) += amd_iommu_debugfs.o
 obj-$(CONFIG_AMD_IOMMU_V2) += amd_iommu_v2.o
-obj-$(CONFIG_ARM_SMMU) += arm-smmu.o arm-smmu-impl.o
+obj-$(CONFIG_ARM_SMMU) += arm-smmu.o arm-smmu-impl.o arm-smmu-qcom.o
 obj-$(CONFIG_ARM_SMMU_V3) += arm-smmu-v3.o
 obj-$(CONFIG_DMAR_TABLE) += dmar.o
 obj-$(CONFIG_INTEL_IOMMU) += intel-iommu.o intel-pasid.o
drivers/iommu/arm-smmu-impl.c +4 -2
index 3f88cd078dd5..d62da270f430 100644
@@ -9,7 +9,6 @@
 
 #include "arm-smmu.h"
 
-
 static int arm_smmu_gr0_ns(int offset)
 {
 	switch(offset) {
@@ -109,7 +108,7 @@ static struct arm_smmu_device *cavium_smmu_impl_init(struct arm_smmu_device *smm
 #define ARM_MMU500_ACR_S2CRB_TLBEN	(1 << 10)
 #define ARM_MMU500_ACR_SMTNMB_TLBEN	(1 << 8)
 
-static int arm_mmu500_reset(struct arm_smmu_device *smmu)
+int arm_mmu500_reset(struct arm_smmu_device *smmu)
 {
 	u32 reg, major;
 	int i;
@@ -170,5 +169,8 @@ struct arm_smmu_device *arm_smmu_impl_init(struct arm_smmu_device *smmu)
 				  "calxeda,smmu-secure-config-access"))
 		smmu->impl = &calxeda_impl;
 
+	if (of_device_is_compatible(smmu->dev->of_node, "qcom,sdm845-smmu-500"))
+		return qcom_smmu_impl_init(smmu);
+
 	return smmu;
 }
drivers/iommu/arm-smmu-qcom.c new file +51 -0

51 lines not on this page · raw message

drivers/iommu/arm-smmu.h +3 -0
index ac9eac966cf5..4bc9e853c95d 100644
@@ -391,5 +391,8 @@ static inline void arm_smmu_writeq(struct arm_smmu_device *smmu, int page,
 	arm_smmu_writeq((s), ARM_SMMU_CB((s), (n)), (o), (v))
 
 struct arm_smmu_device *arm_smmu_impl_init(struct arm_smmu_device *smmu);
+struct arm_smmu_device *qcom_smmu_impl_init(struct arm_smmu_device *smmu);
+
+int arm_mmu500_reset(struct arm_smmu_device *smmu);
 
 #endif /* _ARM_SMMU_H */
Signature · 7 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
#6 lore raw

Re: [PATCHv6 3/3] iommu: arm-smmu-impl: Add sdm845 implementation hook

Quoted 138 lines · Sai Prakash Ranjan

From: Vivek Gautam <[email protected]>

Add reset hook for sdm845 based platforms to turn off the wait-for-safe sequence.

Understanding how wait-for-safe logic affects USB and UFS performance on MTP845 and DB845 boards:

Qcom's implementation of arm,mmu-500 adds a WAIT-FOR-SAFE logic to address under-performance issues in real-time clients, such as Display, and Camera.
On receiving an invalidation requests, the SMMU forwards SAFE request to these clients and waits for SAFE ack signal from real-time clients. The SAFE signal from such clients is used to qualify the start of invalidation.
This logic is controlled by chicken bits, one for each - MDP (display), IFE0, and IFE1 (camera), that can be accessed only from secure software on sdm845.

This configuration, however, degrades the performance of non-real time clients, such as USB, and UFS etc. This happens because, with wait-for-safe logic enabled the hardware tries to throttle non-real time clients while waiting for SAFE ack signals from real-time clients.

On mtp845 and db845 devices, with wait-for-safe logic enabled by the bootloaders we see degraded performance of USB and UFS when kernel enables the smmu stage-1 translations for these clients. Turn off this wait-for-safe logic from the kernel gets us back the perf of USB and UFS devices until we re-visit this when we start seeing perf issues on display/camera on upstream supported SDM845 platforms. The bootloaders on these boards implement secure monitor callbacks to handle a specific command - QCOM_SCM_SVC_SMMU_PROGRAM with which the logic can be toggled.

There are other boards such as cheza whose bootloaders don't enable this logic. Such boards don't implement callbacks to handle the specific SCM call so disabling this logic for such boards will be a no-op.

This change is inspired by the downstream change from Patrick Daly to address performance issues with display and camera by handling this wait-for-safe within separte io-pagetable ops to do TLB maintenance. So a big thanks to him for the change and for all the offline discussions.

Without this change the UFS reads are pretty slow:

$ time dd if=/dev/sda of=/dev/zero bs=1048576 count=10 conv=sync
10+0 records in
10+0 records out
10485760 bytes (10.0MB) copied, 22.394903 seconds, 457.2KB/s
real    0m 22.39s
user    0m 0.00s
sys     0m 0.01s

With this change they are back to rock!

$ time dd if=/dev/sda of=/dev/zero bs=1048576 count=300 conv=sync
300+0 records in
300+0 records out
314572800 bytes (300.0MB) copied, 1.030541 seconds, 291.1MB/s
real    0m 1.03s
user    0m 0.00s
sys     0m 0.54s
---
 drivers/iommu/Makefile        |  2 +-
 drivers/iommu/arm-smmu-impl.c |  6 +++--
 drivers/iommu/arm-smmu-qcom.c | 51 +++++++++++++++++++++++++++++++++++
 drivers/iommu/arm-smmu.h      |  3 +++
 4 files changed, 59 insertions(+), 3 deletions(-)
 create mode 100644 drivers/iommu/arm-smmu-qcom.c
drivers/iommu/Makefile +1 -1
index a2729aadd300..2816e49a8c46 100644
@@ -13,7 +13,7 @@ obj-$(CONFIG_MSM_IOMMU) += msm_iommu.o
 obj-$(CONFIG_AMD_IOMMU) += amd_iommu.o amd_iommu_init.o
 obj-$(CONFIG_AMD_IOMMU_DEBUGFS) += amd_iommu_debugfs.o
 obj-$(CONFIG_AMD_IOMMU_V2) += amd_iommu_v2.o
-obj-$(CONFIG_ARM_SMMU) += arm-smmu.o arm-smmu-impl.o
+obj-$(CONFIG_ARM_SMMU) += arm-smmu.o arm-smmu-impl.o arm-smmu-qcom.o
 obj-$(CONFIG_ARM_SMMU_V3) += arm-smmu-v3.o
 obj-$(CONFIG_DMAR_TABLE) += dmar.o
 obj-$(CONFIG_INTEL_IOMMU) += intel-iommu.o intel-pasid.o
drivers/iommu/arm-smmu-impl.c +4 -2
index 3f88cd078dd5..d62da270f430 100644
@@ -9,7 +9,6 @@
 
 #include "arm-smmu.h"
 
-
 static int arm_smmu_gr0_ns(int offset)
 {
 	switch(offset) {
@@ -109,7 +108,7 @@ static struct arm_smmu_device *cavium_smmu_impl_init(struct arm_smmu_device *smm
 #define ARM_MMU500_ACR_S2CRB_TLBEN	(1 << 10)
 #define ARM_MMU500_ACR_SMTNMB_TLBEN	(1 << 8)
 
-static int arm_mmu500_reset(struct arm_smmu_device *smmu)
+int arm_mmu500_reset(struct arm_smmu_device *smmu)
 {
 	u32 reg, major;
 	int i;
@@ -170,5 +169,8 @@ struct arm_smmu_device *arm_smmu_impl_init(struct arm_smmu_device *smmu)
 				  "calxeda,smmu-secure-config-access"))
 		smmu->impl = &calxeda_impl;
 
+	if (of_device_is_compatible(smmu->dev->of_node, "qcom,sdm845-smmu-500"))
+		return qcom_smmu_impl_init(smmu);
+
 	return smmu;
 }
drivers/iommu/arm-smmu-qcom.c new file +16 -0
new file mode 100644
index 000000000000..24c071c1d8b0
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (c) 2019, The Linux Foundation. All rights reserved.
+ */
+
+#include <linux/qcom_scm.h>
+
+#include "arm-smmu.h"
+
+struct qcom_smmu {
+	struct arm_smmu_device smmu;
+};
+
+static int qcom_sdm845_smmu500_reset(struct arm_smmu_device *smmu)
+{
+	int ret;

drivers/iommu/arm-smmu-qcom.c line 26

Quoted 10 lines
drivers/iommu/arm-smmu-qcom.c new file +10 -0
+
+	arm_mmu500_reset(smmu);
+
+	/*
+	 * To address performance degradation in non-real time clients,
+	 * such as USB and UFS, turn off wait-for-safe on sdm845 based boards,
+	 * such as MTP and db845, whose firmwares implement secure monitor
+	 * call handlers to turn on/off the wait-for-safe logic.
+	 */
+	ret = qcom_scm_qsmmu500_wait_safe_toggle(0);

In the transition to this new design we lost the ability to enable/disable the safe toggle per board, which according to Vivek would result in some issue with Cheza.

Can you confirm that this is okay? (Or introduce the DT property for enabling the safe_toggle logic?)

Regards,
Bjorn

Quoted 42 lines
drivers/iommu/arm-smmu-qcom.c new file +25 -0
+	if (ret)
+		dev_warn(smmu->dev, "Failed to turn off SAFE logic\n");
+
+	return ret;
+}
+
+static const struct arm_smmu_impl qcom_smmu_impl = {
+	.reset = qcom_sdm845_smmu500_reset,
+};
+
+struct arm_smmu_device *qcom_smmu_impl_init(struct arm_smmu_device *smmu)
+{
+	struct qcom_smmu *qsmmu;
+
+	qsmmu = devm_kzalloc(smmu->dev, sizeof(*qsmmu), GFP_KERNEL);
+	if (!qsmmu)
+		return ERR_PTR(-ENOMEM);
+
+	qsmmu->smmu = *smmu;
+
+	qsmmu->smmu.impl = &qcom_smmu_impl;
+	devm_kfree(smmu->dev, smmu);
+
+	return &qsmmu->smmu;
+}
drivers/iommu/arm-smmu.h +3 -0
index ac9eac966cf5..4bc9e853c95d 100644
@@ -391,5 +391,8 @@ static inline void arm_smmu_writeq(struct arm_smmu_device *smmu, int page,
 	arm_smmu_writeq((s), ARM_SMMU_CB((s), (n)), (o), (v))
 
 struct arm_smmu_device *arm_smmu_impl_init(struct arm_smmu_device *smmu);
+struct arm_smmu_device *qcom_smmu_impl_init(struct arm_smmu_device *smmu);
+
+int arm_mmu500_reset(struct arm_smmu_device *smmu);
 
 #endif /* _ARM_SMMU_H */
Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation
#7 lore raw

Re: [PATCHv6 3/3] iommu: arm-smmu-impl: Add sdm845 implementation hook

Quoting Bjorn Andersson (2019-09-18 17:25:01)

Quoted 8 lines · Sai Prakash Ranjan
Quoted 7 lines

From: Vivek Gautam <[email protected]>

There are other boards such as cheza whose bootloaders don't enable this logic. Such boards don't implement callbacks to handle the specific SCM call so disabling this logic for such boards will be a no-op.

[...]

Quoted 37 lines
Quoted 22 lines
drivers/iommu/arm-smmu-qcom.c new file +16 -0
new file mode 100644
index 000000000000..24c071c1d8b0
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (c) 2019, The Linux Foundation. All rights reserved.
+ */
+
+#include <linux/qcom_scm.h>
+
+#include "arm-smmu.h"
+
+struct qcom_smmu {
+     struct arm_smmu_device smmu;
+};
+
+static int qcom_sdm845_smmu500_reset(struct arm_smmu_device *smmu)
+{
+     int ret;

drivers/iommu/arm-smmu-qcom.c line 26

Quoted 10 lines
drivers/iommu/arm-smmu-qcom.c new file +10 -0
+
+     arm_mmu500_reset(smmu);
+
+     /*
+      * To address performance degradation in non-real time clients,
+      * such as USB and UFS, turn off wait-for-safe on sdm845 based boards,
+      * such as MTP and db845, whose firmwares implement secure monitor
+      * call handlers to turn on/off the wait-for-safe logic.
+      */
+     ret = qcom_scm_qsmmu500_wait_safe_toggle(0);

In the transition to this new design we lost the ability to enable/disable the safe toggle per board, which according to Vivek would result in some issue with Cheza.

Quoted 3 lines

Can you confirm that this is okay? (Or introduce the DT property for enabling the safe_toggle logic?)

I can test this on Cheza. Not sure if anything will happen but it's worth a shot.

#8 lore raw

Re: [PATCHv6 3/3] iommu: arm-smmu-impl: Add sdm845 implementation hook

In reply to Bjorn Andersson
Quoted 7 lines · Bjorn Andersson

In the transition to this new design we lost the ability to enable/disable the safe toggle per board, which according to Vivek would result in some issue with Cheza.

Can you confirm that this is okay? (Or introduce the DT property for enabling the safe_toggle logic?)

Hmm, I don't remember Vivek telling about any issue on Cheza because of this logic. But I will test this on Cheza and let you know.

Thanks,
Sai

Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation
#9 lore raw

Re: [PATCHv6 3/3] iommu: arm-smmu-impl: Add sdm845 implementation hook

Quoted 5 lines · Sai Prakash Ranjan
Quoted 4 lines · Bjorn Andersson

In the transition to this new design we lost the ability to enable/disable the safe toggle per board, which according to Vivek would result in some issue with Cheza.

Quoted 8 lines
Quoted 3 lines

Can you confirm that this is okay? (Or introduce the DT property for enabling the safe_toggle logic?)

Hmm, I don't remember Vivek telling about any issue on Cheza because of this logic.
But I will test this on Cheza and let you know.

I tested this on Cheza and no perf degradation nor any other issue is seen atleast openly, although I see this below stack dump always with cant_sleep change added.

[    5.048860] BUG: assuming atomic context at /mnt/host/source/src/third_party/kernel/v5.3/drivers/firmware/qcom_scm-64.c:206
[    5.060303] in_atomic(): 0, irqs_disabled(): 0, pid: 1, name: swapper/0
[    5.067118] CPU: 4 PID: 1 Comm: swapper/0 Not tainted 5.3.0 #102
[    5.073299] Hardware name: Google Cheza (rev3+) (DT)
[    5.078416] Call trace:
[    5.080953]  dump_backtrace+0x0/0x16c
[    5.084727]  show_stack+0x20/0x2c
[    5.088156]  dump_stack+0x90/0xcc
[    5.091585]  __cant_sleep+0xb4/0xc4
[    5.095192]  __qcom_scm_qsmmu500_wait_safe_toggle+0x5c/0xa0
[    5.100929]  qcom_scm_qsmmu500_wait_safe_toggle+0x28/0x34
[    5.106491]  qcom_sdm845_smmu500_reset+0x24/0x50
[    5.111249]  arm_smmu_device_reset+0x1a4/0x25c
[    5.115827]  arm_smmu_device_probe+0x418/0x50c
[    5.120406]  platform_drv_probe+0x90/0xb0
[    5.124542]  really_probe+0x14c/0x3b8
[    5.128327]  driver_probe_device+0x70/0x140
[    5.132643]  device_driver_attach+0x4c/0x6c
[    5.136960]  __driver_attach+0xc8/0x150
[    5.140917]  bus_for_each_dev+0x84/0xcc
[    5.144873]  driver_attach+0x2c/0x38
[    5.148555]  bus_add_driver+0x108/0x1fc
[    5.152512]  driver_register+0x64/0xf8
[    5.156375]  __platform_driver_register+0x4c/0x58
[    5.161226]  arm_smmu_driver_init+0x1c/0x24
[    5.165545]  do_one_initcall+0x11c/0x2dc
[    5.169595]  do_initcall_level+0x14c/0x174
[    5.173822]  do_basic_setup+0x30/0x48
[    5.177595]  kernel_init_freeable+0xc4/0x144
[    5.181990]  kernel_init+0x14/0x100
[    5.185584]  ret_from_fork+0x10/0x18
#10 lore raw

Re: [PATCHv6 3/3] iommu: arm-smmu-impl: Add sdm845 implementation hook

Quoting Sai Prakash Ranjan (2019-09-19 11:54:27)

Quoted 15 lines · Sai Prakash Ranjan
Quoted 5 lines · Bjorn Andersson
Quoted 4 lines

In the transition to this new design we lost the ability to enable/disable the safe toggle per board, which according to Vivek would result in some issue with Cheza.

Quoted 8 lines
Quoted 3 lines

Can you confirm that this is okay? (Or introduce the DT property for enabling the safe_toggle logic?)

Hmm, I don't remember Vivek telling about any issue on Cheza because of this logic.
But I will test this on Cheza and let you know.

Quoted 4 lines

I tested this on Cheza and no perf degradation nor any other issue is seen
atleast openly, although I see this below stack dump always with cant_sleep change added.

The usage of cant_sleep() here is wrong then, and the comment should be removed from the scm API as well because it looks like it's perfectly OK to call the function from a context that can sleep.

Quoted 18 lines
[    5.048860] BUG: assuming atomic context at 
/mnt/host/source/src/third_party/kernel/v5.3/drivers/firmware/qcom_scm-64.c:206
[    5.060303] in_atomic(): 0, irqs_disabled(): 0, pid: 1, name: 
swapper/0
[    5.067118] CPU: 4 PID: 1 Comm: swapper/0 Not tainted 5.3.0 #102
[    5.073299] Hardware name: Google Cheza (rev3+) (DT)
[    5.078416] Call trace:
[    5.080953]  dump_backtrace+0x0/0x16c
[    5.084727]  show_stack+0x20/0x2c
[    5.088156]  dump_stack+0x90/0xcc
[    5.091585]  __cant_sleep+0xb4/0xc4
[    5.095192]  __qcom_scm_qsmmu500_wait_safe_toggle+0x5c/0xa0
[    5.100929]  qcom_scm_qsmmu500_wait_safe_toggle+0x28/0x34
[    5.106491]  qcom_sdm845_smmu500_reset+0x24/0x50
[    5.111249]  arm_smmu_device_reset+0x1a4/0x25c
[    5.115827]  arm_smmu_device_probe+0x418/0x50c
[    5.120406]  platform_drv_probe+0x90/0xb0
#11 lore raw

Re: [PATCHv6 3/3] iommu: arm-smmu-impl: Add sdm845 implementation hook

Quoted 16 lines · Stephen Boyd

Quoting Sai Prakash Ranjan (2019-09-19 11:54:27)

Quoted 15 lines · Sai Prakash Ranjan
Quoted 13 lines · Bjorn Andersson
Quoted 7 lines

In the transition to this new design we lost the ability to enable/disable the safe toggle per board, which according to Vivek would result in some issue with Cheza.

Can you confirm that this is okay? (Or introduce the DT property for enabling the safe_toggle logic?)

Hmm, I don't remember Vivek telling about any issue on Cheza because of this logic.
But I will test this on Cheza and let you know.

Quoted 9 lines
Quoted 4 lines

I tested this on Cheza and no perf degradation nor any other issue is seen
atleast openly, although I see this below stack dump always with cant_sleep change added.

The usage of cant_sleep() here is wrong then, and the comment should be removed from the scm API as well because it looks like it's perfectly OK to call the function from a context that can sleep.

Looking more into this downstream, it says this *can be called in atomic context* and not *should be called only in atomic context*. So will remove cant_sleep and modify the comment accordingly.

Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation
#12 lore raw

Re: [PATCHv6 3/3] iommu: arm-smmu-impl: Add sdm845 implementation hook

In reply to Sai Prakash Ranjan

Quoting Sai Prakash Ranjan (2019-09-17 02:45:04)

drivers/iommu/arm-smmu-impl.c line 14

Quoted 12 lines
drivers/iommu/arm-smmu-impl.c +0 -1
index 3f88cd078dd5..d62da270f430 100644
@@ -9,7 +9,6 @@
 
 #include "arm-smmu.h"
 
-
 static int arm_smmu_gr0_ns(int offset)
 {
        switch(offset) {

Why is this hunk still around?

Quoted 6 lines
drivers/iommu/arm-smmu-qcom.c new file +0 -0
new file mode 100644
index 000000000000..24c071c1d8b0
@@ -0,0 +1,51 @@

[...]

drivers/iommu/arm-smmu-qcom.c line 48

Quoted 12 lines
drivers/iommu/arm-smmu-qcom.c new file +12 -0
+struct arm_smmu_device *qcom_smmu_impl_init(struct arm_smmu_device *smmu)
+{
+       struct qcom_smmu *qsmmu;
+
+       qsmmu = devm_kzalloc(smmu->dev, sizeof(*qsmmu), GFP_KERNEL);
+       if (!qsmmu)
+               return ERR_PTR(-ENOMEM);
+
+       qsmmu->smmu = *smmu;
+
+       qsmmu->smmu.impl = &qcom_smmu_impl;
+       devm_kfree(smmu->dev, smmu);

This copy is interesting but OK I guess cavium does it.

Quoted 3 lines
drivers/iommu/arm-smmu-qcom.c new file +3 -0
+
+       return &qsmmu->smmu;
+}
#13 lore raw

Re: [PATCHv6 3/3] iommu: arm-smmu-impl: Add sdm845 implementation hook

Quoted 7 lines · Stephen Boyd

Quoting Sai Prakash Ranjan (2019-09-17 02:45:04)

Quoted 6 lines
drivers/iommu/arm-smmu-impl.c +0 -0
index 3f88cd078dd5..d62da270f430 100644
@@ -9,7 +9,6 @@
 
Quoted 8 lines

drivers/iommu/arm-smmu-impl.c line 14

Quoted 6 lines
drivers/iommu/arm-smmu-impl.c +0 -1
 #include "arm-smmu.h"
 
-
 static int arm_smmu_gr0_ns(int offset)
 {
        switch(offset) {

Why is this hunk still around?

I remember correcting this in previous version but somehow slipped in this version. Will correct it.

Quoted 8 lines
Quoted 7 lines
drivers/iommu/arm-smmu-qcom.c new file +0 -0
new file mode 100644
index 000000000000..24c071c1d8b0
@@ -0,0 +1,51 @@

[...]

Quoted 16 lines

drivers/iommu/arm-smmu-qcom.c line 48

Quoted 13 lines

+struct arm_smmu_device *qcom_smmu_impl_init(struct arm_smmu_device *smmu)

drivers/iommu/arm-smmu-qcom.c new file +11 -0
+{
+       struct qcom_smmu *qsmmu;
+
+       qsmmu = devm_kzalloc(smmu->dev, sizeof(*qsmmu), GFP_KERNEL);
+       if (!qsmmu)
+               return ERR_PTR(-ENOMEM);
+
+       qsmmu->smmu = *smmu;
+
+       qsmmu->smmu.impl = &qcom_smmu_impl;
+       devm_kfree(smmu->dev, smmu);

This copy is interesting but OK I guess cavium does it.

This is from nvidia impl since Robin pointed me at its implementation.

Quoted 3 lines

drivers/iommu/arm-smmu-qcom.c line 51

Quoted 3 lines
drivers/iommu/arm-smmu-qcom.c new file +3 -0
+
+       return &qsmmu->smmu;
+}

Thanks,
Sai

Signature · 6 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu