Skip to contentSkip to conversation

bus/pci: support iova=va on PowerNV systems

PATCH Superseded David Christensen 2 participants 5 messages Feb 26, 2020 – Apr 25, 2020

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bus/pci: support iova=va on PowerNV systems

Bare metal PowerNV systems include a DPDK supported IOMMU that allows IOVA=VA support. Test for the platform type and report virtual address support if running on a PowerNV system.

---
 drivers/bus/pci/linux/pci.c | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)
drivers/bus/pci/linux/pci.c +29 -0
index 740a2cdad..696dc177e 100644
@@ -549,6 +549,35 @@ pci_device_iommu_support_va(const struct rte_pci_device *dev)
 bool
 pci_device_iommu_support_va(__rte_unused const struct rte_pci_device *dev)
 {
+	/*
+	 * IOMMU is always present on a PowerNV host (IOMMUv2).
+	 * IOMMU is also present in a KVM/QEMU VM (IOMMUv1) but is not
+	 * currently supported by DPDK. Test for our current environment
+	 * and report VA support as appropriate.
+	 */
+
+	char *line = 0;
+	size_t len = 0;
+	char filename[PATH_MAX] = "/proc/cpuinfo";
+	FILE *fp = fopen(filename, "r");
+
+	if (fp == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): can't open %s: %s\n",
+			__func__, filename, strerror(errno));
+		return false;
+	}
+
+	/* Check for a PowerNV platform */
+	while (getline(&line, &len, fp) != -1) {
+		if (strstr(line, "platform") != NULL)
+			if (strstr(line, "PowerNV") != NULL) {
+				RTE_LOG(DEBUG, EAL, "Running on a PowerNV system\n");
+				fclose(fp);
+				return true;
+			}
+	}
+	fclose(fp);
+
 	return false;
 }
 #else
Signature · 1 line
2.18.1
#1 lore raw
Quoted 4 lines · David Christensen

Bare metal PowerNV systems include a DPDK supported IOMMU that allows IOVA=VA support. Test for the platform type and report virtual address support if running on a PowerNV system.

I can't test it, so I'll trust you on this :-).
Comments below.

Quoted 12 lines
---
 drivers/bus/pci/linux/pci.c | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)
drivers/bus/pci/linux/pci.c +0 -0
index 740a2cdad..696dc177e 100644
@@ -549,6 +549,35 @@ pci_device_iommu_support_va(const struct rte_pci_device *dev)
 bool

drivers/bus/pci/linux/pci.c pci_device_iommu_support_va(const struct rte_pci_device *dev) · line 559

Quoted 10 lines
drivers/bus/pci/linux/pci.c +8 -0
 pci_device_iommu_support_va(__rte_unused const struct rte_pci_device *dev)
 {
+       /*
+        * IOMMU is always present on a PowerNV host (IOMMUv2).
+        * IOMMU is also present in a KVM/QEMU VM (IOMMUv1) but is not
+        * currently supported by DPDK. Test for our current environment
+        * and report VA support as appropriate.
+        */
+
+       char *line = 0;

Nit: NULL

drivers/bus/pci/linux/pci.c pci_device_iommu_support_va(const struct rte_pci_device *dev) · line 571

Quoted 12 lines
drivers/bus/pci/linux/pci.c +12 -0
+       size_t len = 0;
+       char filename[PATH_MAX] = "/proc/cpuinfo";
+       FILE *fp = fopen(filename, "r");
+
+       if (fp == NULL) {
+               RTE_LOG(ERR, EAL, "%s(): can't open %s: %s\n",
+                       __func__, filename, strerror(errno));
+               return false;
+       }
+
+       /* Check for a PowerNV platform */
+       while (getline(&line, &len, fp) != -1) {

From here, you leak line.

man getline

ssize_t getline(char **lineptr, size_t *n, FILE *stream);

If *lineptr is NULL, then getline() will allocate a buffer for storing
the line, which should be freed by the user program. (In this case,
the value in *n is ignored.)

drivers/bus/pci/linux/pci.c pci_device_iommu_support_va(const struct rte_pci_device *dev) · line 572

Quoted 1 line
drivers/bus/pci/linux/pci.c +1 -0
+               if (strstr(line, "platform") != NULL)

Nit, to avoid multiple level of indent, how about invert the check:

if (strstr(line, "platform") == NULL)

    continue;

drivers/bus/pci/linux/pci.c pci_device_iommu_support_va(const struct rte_pci_device *dev) · line 574

Quoted 2 lines
drivers/bus/pci/linux/pci.c +2 -0
+                       if (strstr(line, "PowerNV") != NULL) {
+                               RTE_LOG(DEBUG, EAL, "Running on a PowerNV system\n");

Not really helpful, it does not indicate that this system iommu supports VA.

Quoted 12 lines
drivers/bus/pci/linux/pci.c +6 -1
+                               fclose(fp);
+                               return true;
+                       }
+       }
+       fclose(fp);
+
        return false;
 }
 #else
--

2.18.1

Signature · 1 line
David Marchand
#2 lore raw
Quoted 3 lines
Quoted 3 lines

Bare metal PowerNV systems include a DPDK supported IOMMU that allows IOVA=VA support. Test for the platform type and report virtual address support if running on a PowerNV system.

...

Quoted 5 lines

drivers/bus/pci/linux/pci.c pci_device_iommu_support_va(const struct rte_pci_device *dev) · line 559

Quoted 3 lines
drivers/bus/pci/linux/pci.c +2 -0
+
+       char *line = 0;

Nit: NULL

Fixed.

Quoted 5 lines

drivers/bus/pci/linux/pci.c pci_device_iommu_support_va(const struct rte_pci_device *dev) · line 571

Quoted 2 lines
drivers/bus/pci/linux/pci.c +2 -0
+       /* Check for a PowerNV platform */
+       while (getline(&line, &len, fp) != -1) {

From here, you leak line.

Fixed.

Quoted 5 lines

Nit, to avoid multiple level of indent, how about invert the check:

if (strstr(line, "platform") == NULL)

     continue;

Fixed.

Quoted 4 lines

drivers/bus/pci/linux/pci.c pci_device_iommu_support_va(const struct rte_pci_device *dev) · line 574

Quoted 2 lines
drivers/bus/pci/linux/pci.c +2 -0
+                       if (strstr(line, "PowerNV") != NULL) {
+                               RTE_LOG(DEBUG, EAL, "Running on a PowerNV system\n");

Not really helpful, it does not indicate that this system iommu supports VA.

This is the advice given by the kernel developer who maintains the IOMMU code for PPC. There's nothing in the /sys filesystem that describes the IOMMU like there is in x86_64 platforms. All POWER systems supported by DPDK have an IOMMU (you can't turn it off). This is true for both bare metal systems like PowerNV (Power Non-Virtualized) as well as virtualized systems like QEMU/KVM and PowerVM LPARs. In the case of virtualized systems the IOMMU has different capabilities and is not currently supported in DPDK. Thus, if I know the platform I'm running on then I know the IOMMU is present and enabled.

Dave

#3 lore raw

PATCH v2 1/1 bus/pci: support iova=va on PowerNV systems

In reply to David Christensen

All recent POWER systems, Power 8 and 9 specifically, support an IOMMU (it can't be disabled). The functionality of the IOMMU is different depending on whether it's running on a bare metal PowerNV system or in a virtual environment (PowerVM LPAR or KVM/QEMU). DPDK currently supports the IOMMU found on PowerNV platforms, sPAPRv2, so IOVA=VA mode can be enabled when the correct platform is detected.

The POWER IOMMU type can't be detected through mechansims such as parsing files in the /sys heirarchy like x86_64 systems so the /proc/cpuinfo file is parsed to determine whether Linux is running on bare metal (i.e. PowerNV) or in a virtual environment (KVM/QEMU).

---
v2:
* Fixed memory leak from use of getline
* Fixed nits for NULL memory address and indent levels
* Modified commit message description on logic for detecting IOMMU on a POWER system

---
 drivers/bus/pci/linux/pci.c | 35 ++++++++++++++++++++++++++++++++++-
 1 file changed, 34 insertions(+), 1 deletion(-)
drivers/bus/pci/linux/pci.c +34 -1
index 740a2cdad..a1e30b721 100644
@@ -549,7 +549,40 @@ pci_device_iommu_support_va(const struct rte_pci_device *dev)
 bool
 pci_device_iommu_support_va(__rte_unused const struct rte_pci_device *dev)
 {
-	return false;
+	/*
+	 * IOMMU is always present on a PowerNV host (IOMMUv2).
+	 * IOMMU is also present in a KVM/QEMU VM (IOMMUv1) but is not
+	 * currently supported by DPDK. Test for our current environment
+	 * and report VA support as appropriate.
+	 */
+
+	char *line = NULL;
+	size_t len = 0;
+	char filename[PATH_MAX] = "/proc/cpuinfo";
+	FILE *fp = fopen(filename, "r");
+	bool ret = false;
+
+	if (fp == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): can't open %s: %s\n",
+			__func__, filename, strerror(errno));
+		return ret;
+	}
+
+	/* Check for a PowerNV platform */
+	while (getline(&line, &len, fp) != -1) {
+		if (strstr(line, "platform") != NULL)
+			continue;
+
+		if (strstr(line, "PowerNV") != NULL) {
+			RTE_LOG(DEBUG, EAL, "Running on a PowerNV system\n");
+			ret = true;
+			break;
+		}
+	}
+
+	free(line);
+	fclose(fp);
+	return ret;
 }
 #else
 bool
Signature · 1 line
2.18.1
#4 lore raw
Quoted 9 lines · David Christensen

All recent POWER systems, Power 8 and 9 specifically, support an IOMMU (it can't be disabled). The functionality of the IOMMU is different depending on whether it's running on a bare metal PowerNV system or in a virtual environment (PowerVM LPAR or KVM/QEMU). DPDK currently supports the IOMMU found on PowerNV platforms, sPAPRv2, so IOVA=VA mode can be enabled when the correct platform is detected.

The POWER IOMMU type can't be detected through mechansims such as

mechanisms

Quoted 1 line

parsing files in the /sys heirarchy like x86_64 systems so the

hierarchy

Quoted 4 lines

/proc/cpuinfo file is parsed to determine whether Linux is running on bare metal (i.e. PowerNV) or in a virtual environment (KVM/QEMU).

Applied, thanks.

Signature · 1 line
David Marchand