[dpdk-dev] [PATH 0/5] use IOVAs check based on DMA mask

[dpdk-dev] [PATH 0/5] use IOVAs check based on DMA mask

[Alejandro Lucero]

I sent a patchset about this to be applied on 17.11 stable. The memory
code has had main changes since that version, so here it is the patchset
adjusted to current master repo.

This patchset adds, mainly, a check for ensuring IOVAs are within a
restricted range due to addressing limitations with some devices. There
are two known cases: NFP and IOMMU VT-d emulation.

With this check IOVAs out of range are detected and PMDs can abort
initialization. For the VT-d case, IOVA VA mode is allowed as long as
IOVAs are within the supported range, avoiding to forbid IOVA VA by
default.

For the addressing limitations known cases, there are just 40(NFP) or
39(VT-d) bits for handling IOVAs. When using IOVA PA, those limitations
imply 1TB(NFP) or 512M(VT-d) as upper limits, which is likely enough for
most systems. With machines using more memory, the added check will
ensure IOVAs within the range.

With IOVA VA, and because the way the Linux kernel serves mmap calls
in 64 bits systems, 39 or 40 bits are not enough. It is possible to
give an address hint with a lower starting address than the default one
used by the kernel, and then ensuring the mmap uses that hint or hint plus
some offset. With 64 bits systems, the process virtual address space is
large enoguh for doing the hugepages mmaping within the supported range
when those addressing limitations exist. This patchset also adds a change
for using such a hint making the use of IOVA VA a more than likely
possibility when there are those addressing limitations.

The check is not done by default but just when it is required. This
patchset adds the check for NFP initialization and for setting the IOVA
mode is an emulated VT-d is detected. Also, because the recent patchset
adding dynamic memory allocation, the check is also invoked for ensuring
the new memsegs are within the required range.

This patchset could be applied to stable 18.05.

[dpdk-dev] [PATCH 1/5] mem: add function for checking memsegs IOVAs addresses

[Alejandro Lucero]

A device can suffer addressing limitations. This functions checks
memsegs have iovas within the supported range based on dma mask.

PMD should use this during initialization if supported devices
suffer addressing limitations, returning an error if this function
returns memsegs out of range.

Another potential usage is for emulated IOMMU hardware with addressing
limitations.

It is necessary to save the most restricted dma mask for checking
memory allocated dynamically after initialization.

Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
---
 lib/librte_eal/common/eal_common_memory.c         | 56 +++++++++++++++++++++++
 lib/librte_eal/common/include/rte_eal_memconfig.h |  3 ++
 lib/librte_eal/common/include/rte_memory.h        |  3 ++
 lib/librte_eal/common/malloc_heap.c               | 12 +++++
 lib/librte_eal/linuxapp/eal/eal.c                 |  2 +
 lib/librte_eal/rte_eal_version.map                |  1 +
 6 files changed, 77 insertions(+)

diff --git a/lib/librte_eal/common/eal_common_memory.c b/lib/librte_eal/common/eal_common_memory.c
index fbfb1b0..1e8312b 100644
--- a/lib/librte_eal/common/eal_common_memory.c
+++ b/lib/librte_eal/common/eal_common_memory.c
@@ -383,6 +383,62 @@ struct virtiova {
 	rte_memseg_walk(dump_memseg, f);
 }
 
+static int
+check_iova(const struct rte_memseg_list *msl __rte_unused,
+		const struct rte_memseg *ms, void *arg)
+{
+	uint64_t *mask = arg;
+	rte_iova_t iova;
+
+	/* higher address within segment */
+	iova = (ms->iova + ms->len) - 1;
+	if (!(iova & *mask))
+		return 0;
+
+	RTE_LOG(INFO, EAL, "memseg iova %"PRIx64", len %zx, out of range\n",
+			   ms->iova, ms->len);
+
+	RTE_LOG(INFO, EAL, "\tusing dma mask %"PRIx64"\n", *mask);
+	/* Stop the walk and change mask */
+	*mask = 0;
+	return 1;
+}
+
+#if defined(RTE_ARCH_64)
+#define X86_VA_WIDTH 63
+#else
+#define MAX_DMA_MASK_BITS 31
+#endif
+
+/* check memseg iovas are within the required range based on dma mask */
+int __rte_experimental
+rte_eal_check_dma_mask(uint8_t maskbits)
+{
+	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
+	uint64_t mask;
+
+	/* sanity check */
+	if (maskbits > MAX_DMA_MASK_BITS) {
+		RTE_LOG(INFO, EAL, "wrong dma mask size %u (Max: %u)\n",
+				   maskbits, MAX_DMA_MASK_BITS);
+		return -1;
+	}
+
+	/* keep the more restricted maskbit */
+	if (!mcfg->dma_maskbits || maskbits < mcfg->dma_maskbits)
+		mcfg->dma_maskbits = maskbits;
+
+	/* create dma mask */
+	mask = ~((1ULL << maskbits) - 1);
+
+	rte_memseg_walk(check_iova, &mask);
+
+	if (!mask)
+		return -1;
+
+	return 0;
+}
+
 /* return the number of memory channels */
 unsigned rte_memory_get_nchannel(void)
 {
diff --git a/lib/librte_eal/common/include/rte_eal_memconfig.h b/lib/librte_eal/common/include/rte_eal_memconfig.h
index aff0688..aea44cb 100644
--- a/lib/librte_eal/common/include/rte_eal_memconfig.h
+++ b/lib/librte_eal/common/include/rte_eal_memconfig.h
@@ -77,6 +77,9 @@ struct rte_mem_config {
 	 * exact same address the primary process maps it.
 	 */
 	uint64_t mem_cfg_addr;
+
+	/* keeps the more restricted dma mask */
+	uint8_t dma_maskbits;
 } __attribute__((__packed__));
 
 
diff --git a/lib/librte_eal/common/include/rte_memory.h b/lib/librte_eal/common/include/rte_memory.h
index c4b7f4c..cd439e3 100644
--- a/lib/librte_eal/common/include/rte_memory.h
+++ b/lib/librte_eal/common/include/rte_memory.h
@@ -357,6 +357,9 @@ typedef int (*rte_memseg_list_walk_t)(const struct rte_memseg_list *msl,
  */
 unsigned rte_memory_get_nrank(void);
 
+/* check memsegs iovas are within a range based on dma mask */
+int rte_eal_check_dma_mask(uint8_t maskbits);
+
 /**
  * Drivers based on uio will not load unless physical
  * addresses are obtainable. It is only possible to get
diff --git a/lib/librte_eal/common/malloc_heap.c b/lib/librte_eal/common/malloc_heap.c
index 12aaf2d..255d717 100644
--- a/lib/librte_eal/common/malloc_heap.c
+++ b/lib/librte_eal/common/malloc_heap.c
@@ -259,11 +259,13 @@ struct malloc_elem *
 		int socket, unsigned int flags, size_t align, size_t bound,
 		bool contig, struct rte_memseg **ms, int n_segs)
 {
+	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 	struct rte_memseg_list *msl;
 	struct malloc_elem *elem = NULL;
 	size_t alloc_sz;
 	int allocd_pages;
 	void *ret, *map_addr;
+	uint64_t mask;
 
 	alloc_sz = (size_t)pg_sz * n_segs;
 
@@ -291,6 +293,16 @@ struct malloc_elem *
 		goto fail;
 	}
 
+	if (mcfg->dma_maskbits) {
+		mask = ~((1ULL << mcfg->dma_maskbits) - 1);
+		if (rte_eal_check_dma_mask(mask)) {
+			RTE_LOG(DEBUG, EAL,
+				"%s(): couldn't allocate memory due to DMA mask\n",
+				__func__);
+			goto fail;
+		}
+	}
+
 	/* add newly minted memsegs to malloc heap */
 	elem = malloc_heap_add_memory(heap, msl, map_addr, alloc_sz);
 
diff --git a/lib/librte_eal/linuxapp/eal/eal.c b/lib/librte_eal/linuxapp/eal/eal.c
index e59ac65..616723e 100644
--- a/lib/librte_eal/linuxapp/eal/eal.c
+++ b/lib/librte_eal/linuxapp/eal/eal.c
@@ -263,6 +263,8 @@ enum rte_iova_mode
 	 * processes could later map the config into this exact location */
 	rte_config.mem_config->mem_cfg_addr = (uintptr_t) rte_mem_cfg_addr;
 
+	rte_config.mem_config->dma_maskbits = 0;
+
 }
 
 /* attach to an existing shared memory config */
diff --git a/lib/librte_eal/rte_eal_version.map b/lib/librte_eal/rte_eal_version.map
index 344a43d..85e6212 100644
--- a/lib/librte_eal/rte_eal_version.map
+++ b/lib/librte_eal/rte_eal_version.map
@@ -284,6 +284,7 @@ EXPERIMENTAL {
 	rte_devargs_parsef;
 	rte_devargs_remove;
 	rte_devargs_type_count;
+	rte_eal_check_dma_mask;
 	rte_eal_cleanup;
 	rte_eal_hotplug_add;
 	rte_eal_hotplug_remove;
-- 
1.9.1

[dpdk-dev] [PATCH 2/5] mem: use address hint for mapping hugepages

[Alejandro Lucero]

Linux kernel uses a really high address as starting address for
serving mmaps calls. If there exist addressing limitations and
IOVA mode is VA, this starting address is likely too high for
those devices. However, it is possible to use a lower address in
the process virtual address space as with 64 bits there is a lot
of available space.

This patch adds an address hint as starting address for 64 bits
systems.

Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
---
 lib/librte_eal/common/eal_common_memory.c | 35 ++++++++++++++++++++++++++++++-
 1 file changed, 34 insertions(+), 1 deletion(-)

diff --git a/lib/librte_eal/common/eal_common_memory.c b/lib/librte_eal/common/eal_common_memory.c
index 1e8312b..914f1d8 100644
--- a/lib/librte_eal/common/eal_common_memory.c
+++ b/lib/librte_eal/common/eal_common_memory.c
@@ -37,6 +37,23 @@
 static void *next_baseaddr;
 static uint64_t system_page_sz;
 
+#ifdef RTE_ARCH_64
+/*
+ * Linux kernel uses a really high address as starting address for serving
+ * mmaps calls. If there exists addressing limitations and IOVA mode is VA,
+ * this starting address is likely too high for those devices. However, it
+ * is possible to use a lower address in the process virtual address space
+ * as with 64 bits there is a lot of available space.
+ *
+ * Current known limitations are 39 or 40 bits. Setting the starting address
+ * at 4GB implies there are 508GB or 1020GB for mapping the available
+ * hugepages. This is likely enough for most systems, although a device with
+ * addressing limitations should call rte_eal_check_dma_mask for ensuring all
+ * memory is within supported range.
+ */
+static uint64_t baseaddr = 0x100000000;
+#endif
+
 void *
 eal_get_virtual_area(void *requested_addr, size_t *size,
 		size_t page_sz, int flags, int mmap_flags)
@@ -60,6 +77,11 @@
 			rte_eal_process_type() == RTE_PROC_PRIMARY)
 		next_baseaddr = (void *) internal_config.base_virtaddr;
 
+#ifdef RTE_ARCH_64
+	if (next_baseaddr == NULL && internal_config.base_virtaddr == 0 &&
+			rte_eal_process_type() == RTE_PROC_PRIMARY)
+		next_baseaddr = (void *) baseaddr;
+#endif
 	if (requested_addr == NULL && next_baseaddr != NULL) {
 		requested_addr = next_baseaddr;
 		requested_addr = RTE_PTR_ALIGN(requested_addr, page_sz);
@@ -89,9 +111,20 @@
 
 		mapped_addr = mmap(requested_addr, (size_t)map_sz, PROT_READ,
 				mmap_flags, -1, 0);
+
 		if (mapped_addr == MAP_FAILED && allow_shrink)
 			*size -= page_sz;
-	} while (allow_shrink && mapped_addr == MAP_FAILED && *size > 0);
+
+		if (mapped_addr != MAP_FAILED && addr_is_hint &&
+		    mapped_addr != requested_addr) {
+			/* hint was not used. Try with another offset */
+			munmap(mapped_addr, map_sz);
+			mapped_addr = MAP_FAILED;
+			next_baseaddr = RTE_PTR_ADD(next_baseaddr, 0x100000000);
+			requested_addr = next_baseaddr;
+		}
+	} while ((allow_shrink || addr_is_hint) &&
+		 mapped_addr == MAP_FAILED && *size > 0);
 
 	/* align resulting address - if map failed, we will ignore the value
 	 * anyway, so no need to add additional checks.
-- 
1.9.1

[dpdk-dev] [PATCH 3/5] bus/pci: use IOVAs check when setting IOVA mode

[Alejandro Lucero]

Although VT-d emulation currently only supports 39 bits, it could
be iovas being within that supported range. This patch allows
IOVA mode in such a case.

Indeed, memory initialization code can be modified for using lower
virtual addresses than those used by the kernel for 64 bits processes
by default, and therefore memsegs iovas can use 39 bits or less for
most system. And this is likely 100% true for VMs.

Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
---
 drivers/bus/pci/linux/pci.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/drivers/bus/pci/linux/pci.c b/drivers/bus/pci/linux/pci.c
index 04648ac..215dc10 100644
--- a/drivers/bus/pci/linux/pci.c
+++ b/drivers/bus/pci/linux/pci.c
@@ -588,10 +588,11 @@
 	fclose(fp);
 
 	mgaw = ((vtd_cap_reg & VTD_CAP_MGAW_MASK) >> VTD_CAP_MGAW_SHIFT) + 1;
-	if (mgaw < X86_VA_WIDTH)
-		return false;
 
-	return true;
+	if (!rte_eal_check_dma_mask(mgaw))
+		return true;
+	else
+		return false;
 }
 #elif defined(RTE_ARCH_PPC_64)
 static bool
@@ -615,13 +616,17 @@
 {
 	struct rte_pci_device *dev = NULL;
 	struct rte_pci_driver *drv = NULL;
+	int iommu_dma_mask_check_done = 0;
 
 	FOREACH_DRIVER_ON_PCIBUS(drv) {
 		FOREACH_DEVICE_ON_PCIBUS(dev) {
 			if (!rte_pci_match(drv, dev))
 				continue;
-			if (!pci_one_device_iommu_support_va(dev))
-				return false;
+			if (!iommu_dma_mask_check_done) {
+				if (!pci_one_device_iommu_support_va(dev))
+					return false;
+				iommu_dma_mask_check_done  = 1;
+			}
 		}
 	}
 	return true;
-- 
1.9.1

[dpdk-dev] [PATCH 4/5] net/nfp: check hugepages IOVAs based on DMA mask

[Alejandro Lucero]

NFP devices can not handle DMA addresses requiring more than
40 bits. This patch uses rte_dev_check_dma_mask with 40 bits
and avoids device initialization if memory out of NFP range.

Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
---
 drivers/net/nfp/nfp_net.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/net/nfp/nfp_net.c b/drivers/net/nfp/nfp_net.c
index 6e5e305..aedca31 100644
--- a/drivers/net/nfp/nfp_net.c
+++ b/drivers/net/nfp/nfp_net.c
@@ -2688,6 +2688,14 @@ uint32_t nfp_net_txq_full(struct nfp_net_txq *txq)
 
 	pci_dev = RTE_ETH_DEV_TO_PCI(eth_dev);
 
+	/* NFP can not handle DMA addresses requiring more than 40 bits */
+	if (rte_eal_check_dma_mask(40) < 0) {
+		RTE_LOG(INFO, PMD, "device %s can not be used:",
+				   pci_dev->device.name);
+		RTE_LOG(INFO, PMD, "\trestricted dma mask to 40 bits!\n");
+		return -ENODEV;
+	};
+
 	if ((pci_dev->id.device_id == PCI_DEVICE_ID_NFP4000_PF_NIC) ||
 	    (pci_dev->id.device_id == PCI_DEVICE_ID_NFP6000_PF_NIC)) {
 		port = get_pf_port_number(eth_dev->data->name);
-- 
1.9.1

[dpdk-dev] [PATCH 5/5] net/nfp: support IOVA VA mode

[Alejandro Lucero]

NFP can handle IOVA as VA. It requires to check those IOVAs
being in the supported range what is done during initialization.

Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
---
 drivers/net/nfp/nfp_net.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/drivers/net/nfp/nfp_net.c b/drivers/net/nfp/nfp_net.c
index aedca31..1a7d58c 100644
--- a/drivers/net/nfp/nfp_net.c
+++ b/drivers/net/nfp/nfp_net.c
@@ -3273,14 +3273,16 @@ static int eth_nfp_pci_remove(struct rte_pci_device *pci_dev)
 
 static struct rte_pci_driver rte_nfp_net_pf_pmd = {
 	.id_table = pci_id_nfp_pf_net_map,
-	.drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC,
+	.drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC |
+		     RTE_PCI_DRV_IOVA_AS_VA,
 	.probe = nfp_pf_pci_probe,
 	.remove = eth_nfp_pci_remove,
 };
 
 static struct rte_pci_driver rte_nfp_net_vf_pmd = {
 	.id_table = pci_id_nfp_vf_net_map,
-	.drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC,
+	.drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC |
+		     RTE_PCI_DRV_IOVA_AS_VA,
 	.probe = eth_nfp_pci_probe,
 	.remove = eth_nfp_pci_remove,
 };
-- 
1.9.1

Re: [dpdk-dev] [PATH 0/5] use IOVAs check based on DMA mask

[Alejandro Lucero]

There is a problem with the patch due to a last-hour change regarding
MAX_DMA_MASK_BITS. It turns out initial X86_VA_WIDTH definition was to 47
bits, but it is not enough in some systems leading to IOVA VA not
available. Because MAX_DMA_MASK_BITS was just used for a sanity check about
the mask length, it is harmless to increment the value to 63 for 64-bits
systems, but I was really clumsy doing the change and it does not compile
now.

I've already set the patches as not applicable at patchwork and I will send
another version fixing that issue.



On Thu, Aug 30, 2018 at 5:21 PM, Alejandro Lucero <
alejandro.lucero@netronome.com> wrote:

> I sent a patchset about this to be applied on 17.11 stable. The memory
> code has had main changes since that version, so here it is the patchset
> adjusted to current master repo.
>
> This patchset adds, mainly, a check for ensuring IOVAs are within a
> restricted range due to addressing limitations with some devices. There
> are two known cases: NFP and IOMMU VT-d emulation.
>
> With this check IOVAs out of range are detected and PMDs can abort
> initialization. For the VT-d case, IOVA VA mode is allowed as long as
> IOVAs are within the supported range, avoiding to forbid IOVA VA by
> default.
>
> For the addressing limitations known cases, there are just 40(NFP) or
> 39(VT-d) bits for handling IOVAs. When using IOVA PA, those limitations
> imply 1TB(NFP) or 512M(VT-d) as upper limits, which is likely enough for
> most systems. With machines using more memory, the added check will
> ensure IOVAs within the range.
>
> With IOVA VA, and because the way the Linux kernel serves mmap calls
> in 64 bits systems, 39 or 40 bits are not enough. It is possible to
> give an address hint with a lower starting address than the default one
> used by the kernel, and then ensuring the mmap uses that hint or hint plus
> some offset. With 64 bits systems, the process virtual address space is
> large enoguh for doing the hugepages mmaping within the supported range
> when those addressing limitations exist. This patchset also adds a change
> for using such a hint making the use of IOVA VA a more than likely
> possibility when there are those addressing limitations.
>
> The check is not done by default but just when it is required. This
> patchset adds the check for NFP initialization and for setting the IOVA
> mode is an emulated VT-d is detected. Also, because the recent patchset
> adding dynamic memory allocation, the check is also invoked for ensuring
> the new memsegs are within the required range.
>
> This patchset could be applied to stable 18.05.
>
>