dpdk/21.11: fix illegal instruction on non-AVX CPU

Backport a patch from main branch to fix illegal instruction
issue on CPU which has no AVX instructions.

Reference:
https://bugs.dpdk.org/show_bug.cgi?id=1038

Signed-off-by: Yongxin Liu <yongxin.liu@windriver.com>
Signed-off-by: Anuj Mittal <anuj.mittal@intel.com>

2 files changed, 172 insertions, 0 deletions

diff --git a/recipes-extended/dpdk/dpdk/0002-dma-idxd-fix-AVX2-in-non-datapath-functions.patch b/recipes-extended/dpdk/dpdk/0002-dma-idxd-fix-AVX2-in-non-datapath-functions.patch
new file mode 100644
index 0000000..4cd5164
--- /dev/null
+++ b/recipes-extended/dpdk/dpdk/0002-dma-idxd-fix-AVX2-in-non-datapath-functions.patch
@@ -0,0 +1,171 @@
+From aa802b10237c2f7d3b0d0498de9b2fb438f9b9a2 Mon Sep 17 00:00:00 2001
+From: Bruce Richardson <bruce.richardson@intel.com>
+Date: Fri, 17 Jun 2022 11:59:20 +0100
+Subject: [PATCH] dma/idxd: fix AVX2 in non-datapath functions
+
+While all systems which will use the idxd driver for hardware will
+support AVX2, if the driver is present the initialization functions e.g.
+to register logs, will be called on all systems - irrespective of HW
+support. This can cause issues if the system running DPDK does not have
+AVX2, and the compiler has included AVX instructions in the
+initialization code.
+
+To fix this, remove AVX2 instruction set from the whole build of the
+driver. Instead, we add "target(avx2)" attribute to all datapath
+functions, so those - and only those functions - will having AVX2
+instructions in them.
+
+Bugzilla ID: 1038
+Fixes: 3d36a0a1c7de ("dma/idxd: add data path job submission")
+Cc: stable@dpdk.org
+
+Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
+Acked-by: Conor Walsh <conor.walsh@intel.com>
+
+Upstream-Status: Backport [http://git.dpdk.org/dpdk/commit/?id=aa802b10237c2f7d3b0d0498de9b2fb438f9b9a2]
+
+Signed-off-by: Yongxin Liu <yongxin.liu@windriver.com>
+---
+ drivers/dma/idxd/idxd_common.c | 23 +++++++++++++++++++++++
+ drivers/dma/idxd/meson.build   |  1 -
+ 2 files changed, 23 insertions(+), 1 deletion(-)
+
+diff --git a/drivers/dma/idxd/idxd_common.c b/drivers/dma/idxd/idxd_common.c
+index ea6413cc7a..c77200a457 100644
+--- a/drivers/dma/idxd/idxd_common.c
++++ b/drivers/dma/idxd/idxd_common.c
+@@ -13,12 +13,23 @@
+ 
+ #define IDXD_PMD_NAME_STR "dmadev_idxd"
+ 
++/* systems with DSA all support AVX2 so allow our data-path functions to
++ * always use at least that instruction set
++ */
++#ifndef __AVX2__
++#define __use_avx2 __attribute__((target("avx2")))
++#else
++#define __use_avx2
++#endif
++
++__use_avx2
+ static __rte_always_inline rte_iova_t
+ __desc_idx_to_iova(struct idxd_dmadev *idxd, uint16_t n)
+ {
+        return idxd->desc_iova + (n * sizeof(struct idxd_hw_desc));
+ }
+ 
++__use_avx2
+ static __rte_always_inline void
+ __idxd_movdir64b(volatile void *dst, const struct idxd_hw_desc *src)
+ {
+@@ -28,6 +39,7 @@ __idxd_movdir64b(volatile void *dst, const struct idxd_hw_desc *src)
+                        : "memory");
+ }
+ 
++__use_avx2
+ static __rte_always_inline void
+ __submit(struct idxd_dmadev *idxd)
+ {
+@@ -74,6 +86,7 @@ __submit(struct idxd_dmadev *idxd)
+                        _mm256_setzero_si256());
+ }
+ 
++__use_avx2
+ static __rte_always_inline int
+ __idxd_write_desc(struct idxd_dmadev *idxd,
+                const uint32_t op_flags,
+@@ -112,6 +125,7 @@ __idxd_write_desc(struct idxd_dmadev *idxd,
+        return job_id;
+ }
+ 
++__use_avx2
+ int
+ idxd_enqueue_copy(void *dev_private, uint16_t qid __rte_unused, rte_iova_t src,
+                rte_iova_t dst, unsigned int length, uint64_t flags)
+@@ -126,6 +140,7 @@ idxd_enqueue_copy(void *dev_private, uint16_t qid __rte_unused, rte_iova_t src,
+                        flags);
+ }
+ 
++__use_avx2
+ int
+ idxd_enqueue_fill(void *dev_private, uint16_t qid __rte_unused, uint64_t pattern,
+                rte_iova_t dst, unsigned int length, uint64_t flags)
+@@ -136,6 +151,7 @@ idxd_enqueue_fill(void *dev_private, uint16_t qid __rte_unused, uint64_t pattern
+                        flags);
+ }
+ 
++__use_avx2
+ int
+ idxd_submit(void *dev_private, uint16_t qid __rte_unused)
+ {
+@@ -143,6 +159,7 @@ idxd_submit(void *dev_private, uint16_t qid __rte_unused)
+        return 0;
+ }
+ 
++__use_avx2
+ static enum rte_dma_status_code
+ get_comp_status(struct idxd_completion *c)
+ {
+@@ -163,6 +180,7 @@ get_comp_status(struct idxd_completion *c)
+        }
+ }
+ 
++__use_avx2
+ int
+ idxd_vchan_status(const struct rte_dma_dev *dev, uint16_t vchan __rte_unused,
+                enum rte_dma_vchan_status *status)
+@@ -180,6 +198,7 @@ idxd_vchan_status(const struct rte_dma_dev *dev, uint16_t vchan __rte_unused,
+        return 0;
+ }
+ 
++__use_avx2
+ static __rte_always_inline int
+ batch_ok(struct idxd_dmadev *idxd, uint16_t max_ops, enum rte_dma_status_code *status)
+ {
+@@ -224,6 +243,7 @@ batch_ok(struct idxd_dmadev *idxd, uint16_t max_ops, enum rte_dma_status_code *s
+        return -1; /* error case */
+ }
+ 
++__use_avx2
+ static inline uint16_t
+ batch_completed(struct idxd_dmadev *idxd, uint16_t max_ops, bool *has_error)
+ {
+@@ -275,6 +295,7 @@ batch_completed(struct idxd_dmadev *idxd, uint16_t max_ops, bool *has_error)
+        return ret;
+ }
+ 
++__use_avx2
+ static uint16_t
+ batch_completed_status(struct idxd_dmadev *idxd, uint16_t max_ops, enum rte_dma_status_code *status)
+ {
+@@ -366,6 +387,7 @@ batch_completed_status(struct idxd_dmadev *idxd, uint16_t max_ops, enum rte_dma_
+        return ret;
+ }
+ 
++__use_avx2
+ uint16_t
+ idxd_completed(void *dev_private, uint16_t qid __rte_unused, uint16_t max_ops,
+                uint16_t *last_idx, bool *has_error)
+@@ -383,6 +405,7 @@ idxd_completed(void *dev_private, uint16_t qid __rte_unused, uint16_t max_ops,
+        return ret;
+ }
+ 
++__use_avx2
+ uint16_t
+ idxd_completed_status(void *dev_private, uint16_t qid __rte_unused, uint16_t max_ops,
+                uint16_t *last_idx, enum rte_dma_status_code *status)
+diff --git a/drivers/dma/idxd/meson.build b/drivers/dma/idxd/meson.build
+index f1396be945..dcc0a297d7 100644
+--- a/drivers/dma/idxd/meson.build
++++ b/drivers/dma/idxd/meson.build
+@@ -5,7 +5,6 @@ build = dpdk_conf.has('RTE_ARCH_X86')
+ reason = 'only supported on x86'
+ 
+ deps += ['bus_pci']
+-cflags += '-mavx2' # all platforms with idxd HW support AVX
+ sources = files(
+         'idxd_common.c',
+         'idxd_pci.c',
+-- 
+2.32.0
+
diff --git a/recipes-extended/dpdk/dpdk_21.11.1.bb b/recipes-extended/dpdk/dpdk_21.11.1.bb
index f0d7f74..a54fc85 100644
--- a/recipes-extended/dpdk/dpdk_21.11.1.bb
+++ b/recipes-extended/dpdk/dpdk_21.11.1.bb
@@ -2,6 +2,7 @@ include dpdk.inc
 
 SRC_URI += " \
             file://0001-meson.build-march-and-mcpu-already-passed-by-Yocto-21.11.patch \
+            file://0002-dma-idxd-fix-AVX2-in-non-datapath-functions.patch \
 "
 
 STABLE = "-stable"