Skip to contentSkip to conversation

ath10k: Enable IOMMU support for WCN3990 target

PATCH In review Govind Singh 4 participants 7 messages Feb 15, 2018 – Mar 8, 2018 Qualcomm Atheros ATH10K wireless driver

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ath10k: Enable IOMMU support for WCN3990 target

When an IOMMU device is available on the platform bus, allocate an IOMMU domain and attach the wlan target to it. WCN3990 target can then attach an DMA I/O virtual address space to scan out of bound transactions.

---
 drivers/net/wireless/ath/ath10k/snoc.c | 100 ++++++++++++++++++++++++++++++++-
 drivers/net/wireless/ath/ath10k/snoc.h |   3 +
 2 files changed, 101 insertions(+), 2 deletions(-)
drivers/net/wireless/ath/ath10k/snoc.c +98 -2
index cd21b25..502263d 100644
@@ -26,6 +26,10 @@
 #include <linux/platform_device.h>
 #include <linux/regulator/consumer.h>
 #include <linux/clk.h>
+#include <asm/dma-iommu.h>
+#include <linux/iommu.h>
+#include <linux/dma-mapping.h>
+
 #define  WCN3990_CE_ATTR_FLAGS 0
 #define ATH10K_SNOC_RX_POST_RETRY_MS 50
 #define CE_POLL_PIPE 4
@@ -1292,6 +1296,88 @@ static int ath10k_hw_power_off(struct ath10k *ar)
 	return ret;
 }
 
+static int ath10k_smmu_attach(struct ath10k *ar)
+{
+	struct ath10k_snoc *ar_snoc = ath10k_snoc_priv(ar);
+	struct dma_iommu_mapping *mapping;
+	struct platform_device *pdev;
+	int ret = 0;
+
+	ath10k_dbg(ar, ATH10K_DBG_SNOC, "Initializing SMMU\n");
+
+	pdev = ar_snoc->dev;
+	mapping = arm_iommu_create_mapping(&platform_bus_type,
+					   ar_snoc->smmu_iova_start,
+					   ar_snoc->smmu_iova_len);
+	if (IS_ERR(mapping)) {
+		ath10k_err(ar, "create mapping failed, err = %d\n", ret);
+		ret = PTR_ERR(mapping);
+		goto map_fail;
+	}
+
+	ret = arm_iommu_attach_device(&pdev->dev, mapping);
+	if (ret < 0 && ret != -EEXIST) {
+		ath10k_err(ar, "iommu attach device failed, err = %d\n", ret);
+		goto attach_fail;
+	} else if (ret == -EEXIST) {
+		ret = 0;
+	}
+
+	ar_snoc->smmu_mapping = mapping;
+
+	return ret;
+
+attach_fail:
+	arm_iommu_release_mapping(mapping);
+map_fail:
+	return ret;
+}
+
+static void ath10k_smmu_deinit(struct ath10k *ar)
+{
+	struct ath10k_snoc *ar_snoc = ath10k_snoc_priv(ar);
+	struct platform_device *pdev;
+
+	pdev = ar_snoc->dev;
+
+	if (!ar_snoc->smmu_mapping)
+		return;
+
+	arm_iommu_detach_device(&pdev->dev);
+	arm_iommu_release_mapping(ar_snoc->smmu_mapping);
+
+	ar_snoc->smmu_mapping = NULL;
+}
+
+static int ath10k_smmu_init(struct ath10k *ar)
+{
+	struct ath10k_snoc *ar_snoc = ath10k_snoc_priv(ar);
+	struct platform_device *pdev;
+	struct resource *res;
+	int ret = 0;
+
+	pdev = ar_snoc->dev;
+	res = platform_get_resource_byname(pdev, IORESOURCE_MEM,
+					   "smmu_iova_base");
+	if (!res) {
+		ath10k_err(ar, "SMMU iova base not found\n");
+	} else {
+		ar_snoc->smmu_iova_start = res->start;
+		ar_snoc->smmu_iova_len = resource_size(res);
+		ath10k_dbg(ar, ATH10K_DBG_SNOC, "SMMU iova start: %pa, len: %zu\n",
+			   &ar_snoc->smmu_iova_start, ar_snoc->smmu_iova_len);
+
+		ret = ath10k_smmu_attach(ar);
+		if (ret < 0) {
+			ath10k_err(ar, "SMMU init failed, err = %d, start: %pad, len: %zx\n",
+				   ret, &ar_snoc->smmu_iova_start,
+				   ar_snoc->smmu_iova_len);
+		}
+	}
+
+	return ret;
+}
+
 static const struct of_device_id ath10k_snoc_dt_match[] = {
 	{ .compatible = "qcom,wcn3990-wifi",
 	 .data = &drv_priv,
@@ -1339,16 +1425,22 @@ static int ath10k_snoc_probe(struct platform_device *pdev)
 	ar_snoc->ce.bus_ops = &ath10k_snoc_bus_ops;
 	ar->ce_priv = &ar_snoc->ce;
 
+	ret = ath10k_smmu_init(ar);
+	if (ret) {
+		ath10k_warn(ar, "failed to int SMMU: %d\n", ret);
+		goto err_core_destroy;
+	}
+
 	ath10k_snoc_resource_init(ar);
 	if (ret) {
 		ath10k_warn(ar, "failed to initialize resource: %d\n", ret);
-		goto err_core_destroy;
+		goto err_smmu_deinit;
 	}
 
 	ath10k_snoc_setup_resource(ar);
 	if (ret) {
 		ath10k_warn(ar, "failed to setup resource: %d\n", ret);
-		goto err_core_destroy;
+		goto err_smmu_deinit;
 	}
 	ret = ath10k_snoc_request_irq(ar);
 	if (ret) {
@@ -1396,6 +1488,9 @@ static int ath10k_snoc_probe(struct platform_device *pdev)
 err_release_resource:
 	ath10k_snoc_release_resource(ar);
 
+err_smmu_deinit:
+	ath10k_smmu_deinit(ar);
+
 err_core_destroy:
 	ath10k_core_destroy(ar);
 
@@ -1409,6 +1504,7 @@ static int ath10k_snoc_remove(struct platform_device *pdev)
 	ath10k_dbg(ar, ATH10K_DBG_SNOC, "snoc remove\n");
 	ath10k_core_unregister(ar);
 	ath10k_hw_power_off(ar);
+	ath10k_smmu_deinit(ar);
 	ath10k_snoc_free_irq(ar);
 	ath10k_snoc_release_resource(ar);
 	ath10k_core_destroy(ar);
drivers/net/wireless/ath/ath10k/snoc.h +3 -0
index 05dc98f..a9bca2e 100644
@@ -82,6 +82,9 @@ struct ath10k_snoc {
 	struct timer_list rx_post_retry;
 	struct ath10k_wcn3990_vreg_info *vreg;
 	struct ath10k_wcn3990_clk_info *clk;
+	struct dma_iommu_mapping *smmu_mapping;
+	dma_addr_t smmu_iova_start;
+	size_t smmu_iova_len;
 };
 
 static inline struct ath10k_snoc *ath10k_snoc_priv(struct ath10k *ar)
Signature · 1 line
1.9.1
#1 lore raw
Quoted 16 lines · Govind Singh

When an IOMMU device is available on the platform bus, allocate an IOMMU domain and attach the wlan target to it. WCN3990 target can then attach an DMA I/O virtual address space to scan out of bound transactions.

---
 drivers/net/wireless/ath/ath10k/snoc.c | 100 ++++++++++++++++++++++++++++++++-
 drivers/net/wireless/ath/ath10k/snoc.h |   3 +
 2 files changed, 101 insertions(+), 2 deletions(-)
drivers/net/wireless/ath/ath10k/snoc.c +0 -0
index cd21b25..502263d 100644

drivers/net/wireless/ath/ath10k/snoc.c line 31

Quoted 6 lines
drivers/net/wireless/ath/ath10k/snoc.c +3 -0
@@ -26,6 +26,10 @@
 #include <linux/platform_device.h>
 #include <linux/regulator/consumer.h>
 #include <linux/clk.h>
+#include <asm/dma-iommu.h>
+#include <linux/iommu.h>
+#include <linux/dma-mapping.h>

Kbuild bot reported a problem with arm64 but strangely I cannot find the full report. Anyway, this was the warning:

drivers/net/wireless/ath/ath10k/snoc.c:29:10: fatal error:
asm/dma-iommu.h: No such file or directory

Any ideas? Adding also Arnd, the grand master of compilation problems :)

Full patch here:

https://patchwork.kernel.org/patch/10220719/

Signature · 1 line
Kalle Valo
#2 lore raw
Quoted 28 lines · Kalle Valo
Quoted 16 lines · Govind Singh

When an IOMMU device is available on the platform bus, allocate an IOMMU domain and attach the wlan target to it. WCN3990 target can then attach an DMA I/O virtual address space to scan out of bound transactions.

---
 drivers/net/wireless/ath/ath10k/snoc.c | 100 ++++++++++++++++++++++++++++++++-
 drivers/net/wireless/ath/ath10k/snoc.h |   3 +
 2 files changed, 101 insertions(+), 2 deletions(-)
drivers/net/wireless/ath/ath10k/snoc.c +0 -0
index cd21b25..502263d 100644

drivers/net/wireless/ath/ath10k/snoc.c line 31

Quoted 6 lines
drivers/net/wireless/ath/ath10k/snoc.c +3 -0
@@ -26,6 +26,10 @@
 #include <linux/platform_device.h>
 #include <linux/regulator/consumer.h>
 #include <linux/clk.h>
+#include <asm/dma-iommu.h>
+#include <linux/iommu.h>
+#include <linux/dma-mapping.h>

Kbuild bot reported a problem with arm64 but strangely I cannot find the full report. Anyway, this was the warning:

Quoted 9 lines

drivers/net/wireless/ath/ath10k/snoc.c:29:10: fatal error:
asm/dma-iommu.h: No such file or directory

Any ideas? Adding also Arnd, the grand master of compilation problems :)

Full patch here:

https://patchwork.kernel.org/patch/10220719/

The asm/dma-iommu.h header file exsists only on arm32, no other architecture.

I'm not sure about the purpose of the patch to start with:
it's normally up to the platform code to allocate IOMMU domains, device drivers should only need to manually interact with the IOMMU layer if they need more than one domain, but this ath10k patch appears to be using the default domain and should have no effect as long as the platform code works correctly.

Arnd

#3 lore raw
Quoted 1 line
Quoted 1 line

The asm/dma-iommu.h header file exsists only on arm32, no other architec=

ture.

Quoted 2 lines
Quoted 2 lines

I'm not sure about the purpose of the patch to start with:
it's normally up to the platform code to allocate IOMMU domains, device =

drivers should only need to manually interact with the IOMMU layer if they = need more than one domain, but this ath10k patch appears to be using the de= fault domain and should have no effect as long as the platform code works c= orrectly.
Thanks Arnd, I have fixed this and migrated to 64bit API's(iommu_attach_de= vice/iommu_detach_device/ iommu_get_domain_for_dev), will share the next re= vision.
I tried using the default domain by adding the stream ID and mask in dt and=
no manual interaction, but it is resulting in TZ error and unhandled conte=
xt fault.
Seems I need to provide explicit mapping range(aperture_start/ aperture_end= ) as this is only working combination for me..

BR,
Govind

-----Original Message-----
From: ath10k [mailto:[email protected]] On Behalf Of Arnd = Bergmann
Sent: Thursday, March 1, 2018 3:40 PM
To: Kalle Valo <[email protected]>
Cc: Govind Singh <[email protected]>; linux-wireless <linux-wireless@v= ger.kernel.org>; [email protected]
Subject: Re: [PATCH] ath10k: Enable IOMMU support for WCN3990 target

Quoted 10 lines · Kalle Valo
Quoted 8 lines · Govind Singh

When an IOMMU device is available on the platform bus, allocate an=20 IOMMU domain and attach the wlan target to it.
WCN3990 target can then attach an DMA I/O virtual address space to=20 scan out of bound transactions.

---
drivers/net/wireless/ath/ath10k/snoc.c | 100 ++++++++++++++++++++++++++=

++++++-

Quoted 21 lines
Quoted 9 lines
 drivers/net/wireless/ath/ath10k/snoc.h |   3 +
 2 files changed, 101 insertions(+), 2 deletions(-)
drivers/net/wireless/ath/ath10k/snoc.c +0 -0
index cd21b25..502263d 100644

drivers/net/wireless/ath/ath10k/snoc.c line 30

Quoted 5 lines
drivers/net/wireless/ath/ath10k/snoc.c +3 -0
@@ -26,6 +26,10 @@
 #include <linux/platform_device.h>
 #include <linux/regulator/consumer.h>  #include <linux/clk.h>
+#include <asm/dma-iommu.h>
+#include <linux/iommu.h>
+#include <linux/dma-mapping.h>

Kbuild bot reported a problem with arm64 but strangely I cannot find=20 the full report. Anyway, this was the warning:

drivers/net/wireless/ath/ath10k/snoc.c:29:10: fatal error:
asm/dma-iommu.h: No such file or directory

Quoted 7 lines

Any ideas? Adding also Arnd, the grand master of compilation problems=20 :)

Full patch here:

https://patchwork.kernel.org/patch/10220719/

The asm/dma-iommu.h header file exsists only on arm32, no other architectur= e.

I'm not sure about the purpose of the patch to start with:
it's normally up to the platform code to allocate IOMMU domains, device dri= vers should only need to manually interact with the IOMMU layer if they nee= d more than one domain, but this ath10k patch appears to be using the defau= lt domain and should have no effect as long as the platform code works corr= ectly.

Arnd

_______________________________________________
ath10k mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/ath10k

#4 lore raw

On Thu, Mar 1, 2018 at 2:18 PM, Govind Singh <[email protected]> wro= te:

Quoted 1 line
Quoted 1 line
Quoted 1 line

The asm/dma-iommu.h header file exsists only on arm32, no other archite=

cture.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

I'm not sure about the purpose of the patch to start with:
it's normally up to the platform code to allocate IOMMU domains, device=

drivers should only need to manually interact with the IOMMU layer if they= need more than one domain, but this ath10k patch appears to be using the d=
efault domain and should have no effect as long as the platform code works = correctly.

Quoted 1 line

Thanks Arnd, I have fixed this and migrated to 64bit API's(iommu_attach_=

device/iommu_detach_device/ iommu_get_domain_for_dev), will share the next = revision.

Quoted 1 line

I tried using the default domain by adding the stream ID and mask in dt a=

nd no manual interaction, but it is resulting in TZ error and unhandled con= text fault.

Quoted 1 line

Seems I need to provide explicit mapping range(aperture_start/ aperture_e=

nd) as this is only working combination for me..

I don't see why you need to do that at all, can you clarify?

The IOMMU should be set up implicitly for you here based on the iommus property in DT,
with no driver changes at all. This should work on all architectures/

Arnd

#5 lore raw
Quoted 17 lines · Arnd Bergmann
Quoted 7 lines · Govind Singh
Quoted 7 lines
Quoted 7 lines

The asm/dma-iommu.h header file exsists only on arm32, no other architecture. I'm not sure about the purpose of the patch to start with:
it's normally up to the platform code to allocate IOMMU domains, device drivers should only need to manually interact with the IOMMU layer if they need more than one domain, but this ath10k patch appears to be using the default domain and should have no effect as long as the platform code works correctly.

Quoted 8 lines

Thanks Arnd, I have fixed this and migrated to 64bit API's(iommu_attach_device/iommu_detach_device/ iommu_get_domain_for_dev), will share the next revision. I tried using the default domain by adding the stream ID and mask in dt and no manual interaction, but it is resulting in TZ error and unhandled context fault.
Seems I need to provide explicit mapping range(aperture_start/ aperture_end) as this is only working combination for me..

Quoted 5 lines

I don't see why you need to do that at all, can you clarify?

The IOMMU should be set up implicitly for you here based on the iommus property in DT, with no driver changes at all. This should work on all architectures/

Maybe Govind is using some out-of-tree tree which is buggy in this regard?

Signature · 1 line
Kalle Valo
#6 lore raw
Quoted 19 lines · Kalle Valo
Quoted 17 lines · Arnd Bergmann
Quoted 15 lines · Govind Singh
Quoted 7 lines
Quoted 7 lines

The asm/dma-iommu.h header file exsists only on arm32, no other architecture. I'm not sure about the purpose of the patch to start with:
it's normally up to the platform code to allocate IOMMU domains, device drivers should only need to manually interact with the IOMMU layer if they need more than one domain, but this ath10k patch appears to be using the default domain and should have no effect as long as the platform code works correctly.

Thanks Arnd, I have fixed this and migrated to 64bit API's(iommu_attach_device/iommu_detach_device/ iommu_get_domain_for_dev), will share the next revision. I tried using the default domain by adding the stream ID and mask in dt and no manual interaction, but it is resulting in TZ error and unhandled context fault.
Seems I need to provide explicit mapping range(aperture_start/ aperture_end) as this is only working combination for me..

Quoted 8 lines
Quoted 5 lines

I don't see why you need to do that at all, can you clarify?

The IOMMU should be set up implicitly for you here based on the iommus property in DT, with no driver changes at all. This should work on all architectures/

Maybe Govind is using some out-of-tree tree which is buggy in this regard?

Actually there is limitations in using the iova address range for wlan IP. It can allow certain iova range and i was attaching the iommu to specify the iova range. I am exploring if i can use "dma-ranges" in dt to avoid the manual interaction apart from stream ID and mask.

BR,
Govind