[dpdk-dev] [PATCH v2 1/3] vfio: revert changes for map contiguous areas in one go

[Nithin Dabilpuram]

In order to save DMA entries limited by kernel both for externel
memory and hugepage memory, an attempt was made to map physically
contiguous memory in one go. This cannot be done as VFIO IOMMU type1
does not support partially unmapping a previously mapped memory
region while Heap can request for multi page mapping and
partial unmapping.
Hence for going back to old method of mapping/unmapping at
memseg granularity, this commit reverts commit
d1c7c0cdf7ba ("vfio: map contiguous areas in one go")

Also add documentation on what module parameter needs to be used
to increase the per-container dma map limit for VFIO.

Fixes: d1c7c0cdf7ba ("vfio: map contiguous areas in one go")
Cc: anatoly.burakov at intel.com
Cc: stable at dpdk.org

Signed-off-by: Nithin Dabilpuram <ndabilpuram at marvell.com>
---
 doc/guides/linux_gsg/linux_drivers.rst | 10 ++++++
 lib/librte_eal/linux/eal_vfio.c        | 59 +++++-----------------------------
 2 files changed, 18 insertions(+), 51 deletions(-)

diff --git a/doc/guides/linux_gsg/linux_drivers.rst b/doc/guides/linux_gsg/linux_drivers.rst
index 080b449..bb43ab2 100644
--- a/doc/guides/linux_gsg/linux_drivers.rst
+++ b/doc/guides/linux_gsg/linux_drivers.rst
@@ -67,6 +67,16 @@ Note that in order to use VFIO, your kernel must support it.
 VFIO kernel modules have been included in the Linux kernel since version 3.6.0 and are usually present by default,
 however please consult your distributions documentation to make sure that is the case.
 
+For DMA mapping of either external memory or hugepages, VFIO interface is used.
+VFIO does not support partial unmap of once mapped memory. Hence DPDK's memory is
+mapped in hugepage granularity or system page granularity. Number of DMA
+mappings is limited by kernel with user locked memory limit of a process(rlimit)
+for system/hugepage memory. Another per-container overall limit applicable both
+for external memory and system memory was added in kernel 5.1 defined by
+VFIO module parameter ``dma_entry_limit`` with a default value of 64K.
+When application is out of DMA entries, these limits need to be adjusted to
+increase the allowed limit.
+
 The ``vfio-pci`` module since Linux version 5.7 supports the creation of virtual
 functions. After the PF is bound to vfio-pci module, the user can create the VFs
 by sysfs interface, and these VFs are bound to vfio-pci module automatically.
diff --git a/lib/librte_eal/linux/eal_vfio.c b/lib/librte_eal/linux/eal_vfio.c
index 380f2f4..dbefcba 100644
--- a/lib/librte_eal/linux/eal_vfio.c
+++ b/lib/librte_eal/linux/eal_vfio.c
@@ -516,11 +516,9 @@ static void
 vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
 		void *arg __rte_unused)
 {
-	rte_iova_t iova_start, iova_expected;
 	struct rte_memseg_list *msl;
 	struct rte_memseg *ms;
 	size_t cur_len = 0;
-	uint64_t va_start;
 
 	msl = rte_mem_virt2memseg_list(addr);
 
@@ -549,63 +547,22 @@ vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
 #endif
 	/* memsegs are contiguous in memory */
 	ms = rte_mem_virt2memseg(addr, msl);
-
-	/*
-	 * This memory is not guaranteed to be contiguous, but it still could
-	 * be, or it could have some small contiguous chunks. Since the number
-	 * of VFIO mappings is limited, and VFIO appears to not concatenate
-	 * adjacent mappings, we have to do this ourselves.
-	 *
-	 * So, find contiguous chunks, then map them.
-	 */
-	va_start = ms->addr_64;
-	iova_start = iova_expected = ms->iova;
 	while (cur_len < len) {
-		bool new_contig_area = ms->iova != iova_expected;
-		bool last_seg = (len - cur_len) == ms->len;
-		bool skip_last = false;
-
-		/* only do mappings when current contiguous area ends */
-		if (new_contig_area) {
-			if (type == RTE_MEM_EVENT_ALLOC)
-				vfio_dma_mem_map(default_vfio_cfg, va_start,
-						iova_start,
-						iova_expected - iova_start, 1);
-			else
-				vfio_dma_mem_map(default_vfio_cfg, va_start,
-						iova_start,
-						iova_expected - iova_start, 0);
-			va_start = ms->addr_64;
-			iova_start = ms->iova;
-		}
 		/* some memory segments may have invalid IOVA */
 		if (ms->iova == RTE_BAD_IOVA) {
 			RTE_LOG(DEBUG, EAL, "Memory segment at %p has bad IOVA, skipping\n",
 					ms->addr);
-			skip_last = true;
+			goto next;
 		}
-		iova_expected = ms->iova + ms->len;
+		if (type == RTE_MEM_EVENT_ALLOC)
+			vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
+					ms->iova, ms->len, 1);
+		else
+			vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
+					ms->iova, ms->len, 0);
+next:
 		cur_len += ms->len;
 		++ms;
-
-		/*
-		 * don't count previous segment, and don't attempt to
-		 * dereference a potentially invalid pointer.
-		 */
-		if (skip_last && !last_seg) {
-			iova_expected = iova_start = ms->iova;
-			va_start = ms->addr_64;
-		} else if (!skip_last && last_seg) {
-			/* this is the last segment and we're not skipping */
-			if (type == RTE_MEM_EVENT_ALLOC)
-				vfio_dma_mem_map(default_vfio_cfg, va_start,
-						iova_start,
-						iova_expected - iova_start, 1);
-			else
-				vfio_dma_mem_map(default_vfio_cfg, va_start,
-						iova_start,
-						iova_expected - iova_start, 0);
-		}
 	}
 #ifdef RTE_ARCH_PPC_64
 	cur_len = 0;
-- 
2.8.4