MLNX_OFED 4.9-4.1.7.0 on Debian 11

kernel-5.10.patch

diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/include/linux/sched/mm.h include/linux/sched/mm.h
--- a/include/linux/sched/mm.h	2021-12-08 23:08:07.000000000 +0900
+++ b/include/linux/sched/mm.h	2022-08-05 17:28:35.539301521 +0900
@@ -35,18 +35,18 @@
 #if !defined (HAVE_MMGET_STILL_VALID) && !defined(HAVE_MMGET_STILL_VALID_IN_SCHED_H) && !defined(HAVE_MMGET_STILL_VALID_IN_MM_H)
 /*
  * This has to be called after a get_task_mm()/mmget_not_zero()
- * followed by taking the mmap_sem for writing before modifying the
+ * followed by taking the mmap_lock for writing before modifying the
  * vmas or anything the coredump pretends not to change from under it.
  *
  * NOTE: find_extend_vma() called from GUP context is the only place
- * that can modify the "mm" (notably the vm_start/end) under mmap_sem
+ * that can modify the "mm" (notably the vm_start/end) under mmap_lock
  * for reading and outside the context of the process, so it is also
- * the only case that holds the mmap_sem for reading that must call
- * this function. Generally if the mmap_sem is hold for reading
+ * the only case that holds the mmap_lock for reading that must call
+ * this function. Generally if the mmap_lock is hold for reading
  * there's no need of this check after get_task_mm()/mmget_not_zero().
  *
  * This function can be obsoleted and the check can be removed, after
- * the coredump code will hold the mmap_sem for writing before
+ * the coredump code will hold the mmap_lock for writing before
  * invoking the ->core_dump methods.
  */
 static inline bool mmget_still_valid(struct mm_struct *mm)
diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/include/rdma/ib_verbs.h include/rdma/ib_verbs.h
--- a/include/rdma/ib_verbs.h	2022-08-05 17:32:01.725190251 +0900
+++ b/include/rdma/ib_verbs.h	2022-08-05 18:09:39.200671886 +0900
@@ -78,6 +78,10 @@
 #include <uapi/rdma/ib_user_ioctl_verbs.h>
 #define IB_FW_VERSION_NAME_MAX	ETHTOOL_FWVERS_LEN
 
+#ifndef uninitialized_var
+#define uninitialized_var(x) x = x
+#endif
+
 struct ib_umem_odp;
 struct ib_uqp_object;
 struct ib_usrq_object;
diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/drivers/infiniband/core/cmem.c drivers/infiniband/core/cmem.c
--- a/drivers/infiniband/core/cmem.c	2022-08-05 17:32:01.685190660 +0900
+++ b/drivers/infiniband/core/cmem.c	2022-08-05 17:28:35.543301480 +0900
@@ -22,7 +22,7 @@
 	}
 	/* no locking is needed:
 	  * ib_cmem_release is called from vm_close which is always called
-	  * with mm->mmap_sem held for writing.
+	  * with mm->mmap_lock held for writing.
 	  * The only exception is when the process shutting down but in that case
 	  * counter not relevant any more.
 	  */
@@ -58,7 +58,7 @@
 
 	ib_cmem = (struct ib_cmem *)(area->vm_private_data);
 
-	/* vm_open and vm_close are always called with mm->mmap_sem held for
+	/* vm_open and vm_close are always called with mm->mmap_lock held for
 	  * writing. The only exception is when the process is shutting down, at
 	  * which point vm_close is called with no locks held, but since it is
 	  * after the VMAs have been detached, it is impossible that vm_open will
@@ -190,7 +190,7 @@
 	  */
 	ntotal_pages = PAGE_ALIGN(total_size) >> PAGE_SHIFT;
 	/* ib_cmem_alloc_contiguous_pages is called as part of mmap
-	  * with mm->mmap_sem held for writing.
+	  * with mm->mmap_lock held for writing.
 	  * No need to lock
 	  */
 #ifdef HAVE_PINNED_VM
diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/drivers/infiniband/core/umem.c drivers/infiniband/core/umem.c
--- a/drivers/infiniband/core/umem.c	2022-08-05 17:32:01.709190415 +0900
+++ b/drivers/infiniband/core/umem.c	2022-08-05 17:28:35.543301480 +0900
@@ -460,7 +460,7 @@
 	new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
 	if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
 #else
-	down_write(&mm->mmap_sem);
+	down_write(&mm->mmap_lock);
 #ifdef HAVE_PINNED_VM
 	if (check_add_overflow(mm->pinned_vm, npages, &new_pinned) ||
 	    (new_pinned > lock_limit && !capable(CAP_IPC_LOCK))) {
@@ -473,7 +473,7 @@
 #ifdef HAVE_ATOMIC_PINNED_VM
 		atomic64_sub(npages, &mm->pinned_vm);
 #else
-		up_write(&mm->mmap_sem);
+		up_write(&mm->mmap_lock);
 #ifdef HAVE_PINNED_VM
 		pr_debug("%s: requested to lock(%lu) while limit is(%lu)\n",
 		       __func__, new_pinned, lock_limit);
@@ -489,7 +489,7 @@
 #ifdef HAVE_PINNED_VM
 	mm->pinned_vm = new_pinned;
 #endif /* HAVE_PINNED_VM */
-	up_write(&mm->mmap_sem);
+	up_write(&mm->mmap_lock);
 #endif /* HAVE_ATOMIC_PINNED_VM */
 	cur_base = addr & PAGE_MASK;
 
@@ -508,7 +508,7 @@
 	sg = umem->sg_head.sgl;
 
 	while (npages) {
-		down_read(&mm->mmap_sem);
+		down_read(&mm->mmap_lock);
 #ifdef HAVE_FOLL_LONGTERM
 		ret = get_user_pages(cur_base,
 				     min_t(unsigned long, npages,
@@ -552,7 +552,7 @@
 				     PAGE_SIZE / sizeof(struct page *)));
 #endif
 
-			up_read(&mm->mmap_sem);
+			up_read(&mm->mmap_lock);
 			goto umem_release;
 		}
 
@@ -563,7 +563,7 @@
 			dma_get_max_seg_size(context->device->dma_device),
 			&umem->sg_nents);
 #if !defined(HAVE_FOLL_LONGTERM) && !defined(HAVE_GET_USER_PAGES_LONGTERM)
-		/* Continue to hold the mmap_sem as vma_list access
+		/* Continue to hold the mmap_lock as vma_list access
 		 *               * needs to be protected.
 		 *                                */
 		for (i = 0; i < ret && umem->hugetlb; i++) {
@@ -571,7 +571,7 @@
 				umem->hugetlb = 0;
 		}
 #endif
-		up_read(&mm->mmap_sem);
+		up_read(&mm->mmap_lock);
 	}
 
 	sg_mark_end(sg);
@@ -610,13 +610,13 @@
 #ifdef HAVE_ATOMIC_PINNED_VM
 	atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
 #else
-	down_write(&mm->mmap_sem);
+	down_write(&mm->mmap_lock);
 #ifdef HAVE_PINNED_VM
 	mm->pinned_vm -= ib_umem_num_pages(umem);
 #else
 	mm->locked_vm -= ib_umem_num_pages(umem);
 #endif /* HAVE_PINNED_VM */
-	up_write(&mm->mmap_sem);
+	up_write(&mm->mmap_lock);
 #endif /* HAVE_ATOMIC_PINNED_VM */
 out:
 #if !defined(HAVE_FOLL_LONGTERM) && !defined(HAVE_GET_USER_PAGES_LONGTERM)
@@ -681,13 +681,13 @@
 #ifdef HAVE_ATOMIC_PINNED_VM
 	atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
 #else
-	down_write(&umem->owning_mm->mmap_sem);
+	down_write(&umem->owning_mm->mmap_lock);
 #ifdef HAVE_PINNED_VM
 	umem->owning_mm->pinned_vm -= ib_umem_num_pages(umem);
 #else
 	umem->owning_mm->locked_vm -= ib_umem_num_pages(umem);
 #endif /* HAVE_PINNED_VM */
-	up_write(&umem->owning_mm->mmap_sem);
+	up_write(&umem->owning_mm->mmap_lock);
 #endif /*HAVE_ATOMIC_PINNED_VM*/
 
 	__ib_umem_release_tail(umem);
diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/drivers/infiniband/core/umem_odp.c drivers/infiniband/core/umem_odp.c
--- a/drivers/infiniband/core/umem_odp.c	2022-08-05 17:32:01.713190374 +0900
+++ b/drivers/infiniband/core/umem_odp.c	2022-08-05 18:01:24.459140610 +0900
@@ -571,15 +571,15 @@
 		struct vm_area_struct *vma;
 		struct hstate *h;
 
-		down_read(&mm->mmap_sem);
+		down_read(&mm->mmap_lock);
 		vma = find_vma(mm, ib_umem_start(umem));
 		if (!vma || !is_vm_hugetlb_page(vma)) {
-			up_read(&mm->mmap_sem);
+			up_read(&mm->mmap_lock);
 			return -EINVAL;
 		}
 		h = hstate_vma(vma);
 		umem->page_shift = huge_page_shift(h);
-		up_read(&mm->mmap_sem);
+		up_read(&mm->mmap_lock);
 #if !defined(HAVE_FOLL_LONGTERM) && !defined(HAVE_GET_USER_PAGES_LONGTERM)
 		umem->hugetlb = 1;
 	} else {
@@ -816,7 +816,7 @@
 				(bcnt + BIT(page_shift) - 1) >> page_shift,
 				PAGE_SIZE / sizeof(struct page *));
 
-		down_read(&owning_mm->mmap_sem);
+		down_read(&owning_mm->mmap_lock);
 		/*
 		 * Note: this might result in redundent page getting. We can
 		 * avoid this by checking dma_list to be 0 before calling
@@ -824,6 +824,12 @@
 		 * complex (and doesn't gain us much performance in most use
 		 * cases).
 		 */
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(5, 8, 0)
+		npages = get_user_pages_remote(owning_mm,
+				user_virt, gup_num_pages,
+				access_mask & ODP_WRITE_ALLOWED_BIT,
+				local_page_list, NULL, NULL);
+#else
 #if defined(HAVE_GET_USER_PAGES_REMOTE_8_PARAMS) || defined(HAVE_GET_USER_PAGES_REMOTE_7_PARAMS) || defined(HAVE_GET_USER_PAGES_REMOTE_8_PARAMS_W_LOCKED)
 		npages = get_user_pages_remote(owning_process, owning_mm,
 				user_virt, gup_num_pages,
@@ -832,11 +838,13 @@
 				flags, local_page_list, NULL, NULL);
 #else
 				flags, local_page_list, NULL);
-#endif
+#endif /* HAVE_GET_USER_PAGES_REMOTE_8_PARAMS_W_LOCKED */
 #else
 				access_mask & ODP_WRITE_ALLOWED_BIT, 0,
 				local_page_list, NULL);
-#endif
+#endif /* HAVE_GET_USER_PAGES_GUP_FLAGS */
+#else
+#ifdef HAVE_GET_USER_PAGES_5_PARAMS
 #else
 		npages = get_user_pages(owning_process, owning_mm,
 				user_virt, gup_num_pages,
@@ -845,9 +853,11 @@
 #else
 				access_mask & ODP_WRITE_ALLOWED_BIT,
 				0, local_page_list, NULL);
-#endif
-#endif
-		up_read(&owning_mm->mmap_sem);
+#endif /* HAVE_GET_USER_PAGES_7_PARAMS */
+#endif /* HAVE_GET_USER_PAGES_5_PARAMS */
+#endif /* defined(...) */
+#endif /* LINUX_VERSION_CODE >= KERNEL_VERSION(5, 8, 0) */
+		up_read(&owning_mm->mmap_lock);
 
 		if (npages < 0) {
 			if (npages != -EAGAIN)
diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/drivers/infiniband/core/uverbs_cmd.c drivers/infiniband/core/uverbs_cmd.c
--- a/drivers/infiniband/core/uverbs_cmd.c	2022-08-05 17:32:01.717190333 +0900
+++ b/drivers/infiniband/core/uverbs_cmd.c	2022-08-05 18:09:48.096560360 +0900
@@ -49,7 +49,6 @@
 #include "core_priv.h"
 #include "uverbs_exp.h"
 
-
 /*
  * Copy a response to userspace. If the provided 'resp' is larger than the
  * user buffer it is silently truncated. If the user provided a larger buffer
diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/drivers/infiniband/core/uverbs_main.c drivers/infiniband/core/uverbs_main.c
--- a/drivers/infiniband/core/uverbs_main.c	2022-08-05 17:32:01.717190333 +0900
+++ b/drivers/infiniband/core/uverbs_main.c	2022-08-05 17:28:35.543301480 +0900
@@ -1018,12 +1018,12 @@
 			return;
 
 		/*
-		 * The umap_lock is nested under mmap_sem since it used within
+		 * The umap_lock is nested under mmap_lock since it used within
 		 * the vma_ops callbacks, so we have to clean the list one mm
 		 * at a time to get the lock ordering right. Typically there
 		 * will only be one mm, so no big deal.
 		 */
-		down_read(&mm->mmap_sem);
+		down_read(&mm->mmap_lock);
 		if (!mmget_still_valid(mm))
 			goto skip_mm;
 		mutex_lock(&ufile->umap_lock);
@@ -1050,7 +1050,7 @@
 		}
 		mutex_unlock(&ufile->umap_lock);
 	skip_mm:
-		up_read(&mm->mmap_sem);
+		up_read(&mm->mmap_lock);
 		mmput(mm);
 	}
 }
diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/drivers/infiniband/hw/mlx4/mr.c drivers/infiniband/hw/mlx4/mr.c
--- a/drivers/infiniband/hw/mlx4/mr.c	2021-12-08 23:08:07.000000000 +0900
+++ b/drivers/infiniband/hw/mlx4/mr.c	2022-08-05 17:28:35.543301480 +0900
@@ -380,7 +380,7 @@
 	if (!ib_access_writable(access_flags)) {
 		struct vm_area_struct *vma;
 
-		down_read(&current->mm->mmap_sem);
+		down_read(&current->mm->mmap_lock);
 		/*
 		 * FIXME: Ideally this would iterate over all the vmas that
 		 * cover the memory, but for now it requires a single vma to
@@ -395,7 +395,7 @@
 			access_flags |= IB_ACCESS_LOCAL_WRITE;
 		}
 
-		up_read(&current->mm->mmap_sem);
+		up_read(&current->mm->mmap_lock);
 	}
 
 	return ib_umem_get(udata, start, length, access_flags, 0, peer_mem_flags);
diff -Naur /home/kyujin/mlnx-ofed-patch/mlnx-ofed-kernel-4.9-orig/drivers/infiniband/hw/mlx5/main_exp.c drivers/infiniband/hw/mlx5/main_exp.c
--- a/drivers/infiniband/hw/mlx5/main_exp.c	2022-08-05 17:32:01.797189514 +0900
+++ b/drivers/infiniband/hw/mlx5/main_exp.c	2022-08-05 17:28:35.547301439 +0900
@@ -2546,7 +2546,7 @@
 	dm->size = act_size;
 
 	if (context) {
-		down_read(&current->mm->mmap_sem);
+		down_read(&current->mm->mmap_lock);
 		vma = find_vma(current->mm, uaddr & PAGE_MASK);
 		if (!vma || (vma->vm_end - vma->vm_start < map_size)) {
 			ret = -EINVAL;
@@ -2568,7 +2568,7 @@
 			goto err_vma;
 		}
 
-		up_read(&current->mm->mmap_sem);
+		up_read(&current->mm->mmap_lock);
 	} else {
 		dm->dm_base_addr = ioremap(memic_addr, length);
 		if (!dm->dm_base_addr) {
@@ -2583,7 +2583,7 @@
 	return &dm->ibdm;
 
 err_vma:
-	up_read(&current->mm->mmap_sem);
+	up_read(&current->mm->mmap_lock);
 
 err_map:
 	mlx5_cmd_dealloc_memic(dm_db, memic_addr, act_size);

patch-ofed-kernel-dkms.sh

#! /bin/bash
if [ -z $1 ]; then
    echo "Usage: ./patch-ofed-kernel-dkms.sh <Path to original mlnx-ofed-kernel-dkms_4.9-OFED.4.9.4.1.7.1_all.deb file>"
    exit 1
fi
dpkg-deb -R $1 tmp  # unpack mlnx-ofed-kernel-dkms
cd tmp
./configure  # load backport patches to source
# Credit: https://forums.developer.nvidia.com/t/is-debian-11-bulleyes-not-supported-yet-on-the-latest-mlnx-ofed-4-9-lts-driver/219606
# Apply patch written by @firemeteor on NVIDIA Forum
curl -L https://pastebin.com/raw/e09pFd0h | tail -n '+3' | patch -p1
cd compat; ./autogen.sh; cd ..  # run autogen.sh
# Apply patches to run on Linux 5.10
curl -L https://gist.github.com/kyujin-cho/6c41f0c4be920e9b3c5183857fd4f4b3/raw/kernel-5.10.patch | patch -p1
cd ..
# Pack modified sources to new deb
dpkg-deb -b tmp mlnx-ofed-kernel-dkms_4.9-OFED.4.9.4.1.7.1_all.deb
# Replace unpatched deb with patched one
mv mlnx-ofed-kernel-dkms_4.9-OFED.4.9.4.1.7.1_all.deb $1

Hey, thanks for this

I think it's slightly off though- I needed this:

#! /bin/bash
if [ -z $1 ]; then
    echo "Usage: ./patch-ofed-kernel-dkms.sh <Path to original mlnx-ofed-kernel-dkms_4.9-OFED.4.9.4.1.7.1_all.deb file>"
    exit 1
fi
dpkg-deb -R $1 tmp  # unpack mlnx-ofed-kernel-dkms
- cd tmp
+ cd tmp/usr/src/mlnx-ofed-kernel-4.9
...

Thanks again for publishing this, I wasn't having fun trolling through the Mellanox (err, Nvidia, I guess) forums

Hello,

Thank you for this, just a little modification too :

...
curl -L https://pastebin.com/raw/e09pFd0h | tail -n '+3' | patch -p1
cd compat; ./autogen.sh; cd ..  # run autogen.sh
# Apply patches to run on Linux 5.10
curl -L https://gist.github.com/kyujin-cho/6c41f0c4be920e9b3c5183857fd4f4b3/raw/kernel-5.10.patch | patch -p1
cd ../../../..
# Pack modified sources to new deb
...

Thx!