[PATCH v2] vhost-scsi: use FOLL_PIN for DMA page pinning

[PATCH v2] vhost-scsi: use FOLL_PIN for DMA page pinning

[Karthikeyan KS]

vhost_scsi_map_to_sgl() takes references on guest buffer pages with
iov_iter_get_pages2() (FOLL_GET) and then performs DMA to/from them.
Per Documentation/core-api/pin_user_pages.rst (Case 1: Direct IO),
pages that are the target of DMA must be pinned with FOLL_PIN instead.
Only FOLL_PIN pins are visible to folio_maybe_dma_pinned(), which fork()
uses to copy rather than COW-share anonymous pages under DMA, and only
FOLL_PIN guarantees a pinned anonymous page is exclusive to the process,
so a later write fault cannot replace it while DMA is in flight.

Switch to iov_iter_extract_pages(), which pins user-backed iterators,
and release the pages with unpin_user_page(). Kernel-backed iterators
take neither a reference nor a pin, so track with
iov_iter_extract_will_pin() whether an unpin is needed.

Cc: Jason Wang <jasowangio@gmail.com>
Cc: Mike Christie <michael.christie@oracle.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: virtualization@lists.linux.dev
Signed-off-by: Karthikeyan KS <karthiproffesional@gmail.com>
---
v1 -> v2:
 - Rebased on vhost next. Sashiko reported that the revert loop in
   vhost_scsi_map_to_sgl() leaks a page when the misaligned-I/O check
   trips on the first page with a non-zero offset. That bug is older
   than this patch (the loop leaked a put_page() reference) and is fixed
   in vhost next by Jia Jia's 058fc59326af ("vhost-scsi: fix page
   reference leak in misaligned I/O fallback"). On top of it, this patch
   only calls unpin_user_page() instead of put_page() in that loop.
 - Reworded the commit message to describe FOLL_GET and FOLL_PIN more
   accurately.

Testing: in QEMU (fileio backend), a 1 MiB write at page offset 3840 is
257 pages, more than BIO_MAX_VECS, so the misaligned-I/O check trips on
the first page and dmesg shows the "vhost-scsi detected misaligned IO"
warning. On v1 (before 058fc59326af) the nr_foll_pin_acquired and
nr_foll_pin_released counters end at 257 and 256 for each such command;
with v2 on vhost next they balance. Plain (256 pages) and T10-PI (260
pages) writes balance too.

v1:       https://lore.kernel.org/20261007163716.707102-1-karthiproffesional@gmail.com
RESEND:   https://lore.kernel.org/20261020182640.2048506-1-karthiproffesional@gmail.com
review:   https://lore.kernel.org/sashiko-outbox-180555@kernel.org
leak fix: https://lore.kernel.org/20261023104035.6686-1-physicalmtea@gmail.com

 drivers/vhost/scsi.c | 50 +++++++++++++++++++++++++++++++-------------
 1 file changed, 35 insertions(+), 15 deletions(-)

diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
index e44044449cef..8beeff06a7a7 100644
--- a/drivers/vhost/scsi.c
+++ b/drivers/vhost/scsi.c
@@ -115,6 +115,8 @@ struct vhost_scsi_cmd {
 	u32 tvc_sgl_count;
 	u32 tvc_prot_sgl_count;
 	u32 copied_iov:1;
+	u32 tvc_need_unpin:1;
+	u32 tvc_prot_need_unpin:1;
 	const void *read_iov;
 	struct iov_iter *read_iter;
 	struct scatterlist *sgl;
@@ -463,8 +465,13 @@ static void vhost_scsi_release_cmd_res(struct se_cmd *se_cmd)
 
 			if (tv_cmd->copied_iov)
 				__free_page(page);
-			else
-				put_page(page);
+			else if (tv_cmd->tvc_need_unpin)
+				unpin_user_page(page);
+			/*
+			 * Else: iov_iter_extract_pages() took neither a ref
+			 * nor a pin on this page (ITER_KVEC/BVEC/etc), so
+			 * there is nothing to release here.
+			 */
 		}
 		kfree(tv_cmd->read_iter);
 		kfree(tv_cmd->read_iov);
@@ -473,8 +480,8 @@ static void vhost_scsi_release_cmd_res(struct se_cmd *se_cmd)
 	if (tv_cmd->tvc_prot_sgl_count) {
 		for_each_sgtable_sg(&tv_cmd->prot_table, sg, i) {
 			page = sg_page(sg);
-			if (page)
-				put_page(page);
+			if (page && tv_cmd->tvc_prot_need_unpin)
+				unpin_user_page(page);
 		}
 		sg_free_table_chained(&tv_cmd->prot_table, vs->inline_sg_cnt);
 	}
@@ -820,7 +827,8 @@ vhost_scsi_get_cmd(struct vhost_virtqueue *vq, u64 scsi_tag)
 
 static void vhost_scsi_revert_map_iov_to_sgl(struct iov_iter *iter,
 					     struct scatterlist *curr,
-					     struct scatterlist *end)
+					     struct scatterlist *end,
+					     bool need_unpin)
 {
 	size_t revert_bytes = 0;
 	struct page *page;
@@ -829,7 +837,8 @@ static void vhost_scsi_revert_map_iov_to_sgl(struct iov_iter *iter,
 		page = sg_page(curr);
 
 		if (page) {
-			put_page(page);
+			if (need_unpin)
+				unpin_user_page(page);
 			revert_bytes += curr->length;
 		}
 		/* Clear so we can re-use it for the copy path */
@@ -859,9 +868,9 @@ vhost_scsi_map_to_sgl(struct vhost_scsi_cmd *cmd,
 	size_t offset;
 	unsigned int n, npages = 0;
 
-	bytes = iov_iter_get_pages2(iter, pages, LONG_MAX,
-				VHOST_SCSI_PREALLOC_UPAGES, &offset);
-	/* No pages were pinned */
+	bytes = iov_iter_extract_pages(iter, &pages, LONG_MAX,
+				       VHOST_SCSI_PREALLOC_UPAGES, 0, &offset);
+	/* No pages were extracted */
 	if (bytes <= 0)
 		return bytes < 0 ? bytes : -EFAULT;
 
@@ -884,7 +893,7 @@ vhost_scsi_map_to_sgl(struct vhost_scsi_cmd *cmd,
 		    cmd->tvc_sgl_count > BIO_MAX_VECS) {
 			WARN_ONCE(true,
 				  "vhost-scsi detected misaligned IO. Performance may be degraded.");
-			goto revert_iter_get_pages;
+			goto revert_iter_extract_pages;
 		}
 
 		sg_set_page(sg, pages[npages++], n, offset);
@@ -896,8 +905,10 @@ vhost_scsi_map_to_sgl(struct vhost_scsi_cmd *cmd,
 	*sgl = sg;
 	return npages;
 
-revert_iter_get_pages:
-	vhost_scsi_revert_map_iov_to_sgl(iter, *sgl, sg);
+revert_iter_extract_pages:
+	vhost_scsi_revert_map_iov_to_sgl(iter, *sgl, sg,
+					 is_prot ? cmd->tvc_prot_need_unpin :
+						   cmd->tvc_need_unpin);
 
 	iov_iter_revert(iter, bytes);
 	/*
@@ -907,7 +918,9 @@ vhost_scsi_map_to_sgl(struct vhost_scsi_cmd *cmd,
 	while (bytes) {
 		n = min_t(unsigned int, PAGE_SIZE - offset, bytes);
 
-		put_page(pages[npages++]);
+		if (is_prot ? cmd->tvc_prot_need_unpin : cmd->tvc_need_unpin)
+			unpin_user_page(pages[npages]);
+		npages++;
 		bytes -= n;
 		offset = 0;
 	}
@@ -983,6 +996,7 @@ vhost_scsi_copy_iov_to_sgl(struct vhost_scsi_cmd *cmd, struct iov_iter *iter,
 	}
 
 	cmd->copied_iov = 1;
+	cmd->tvc_need_unpin = 0;
 	return 0;
 
 err:
@@ -1007,11 +1021,17 @@ vhost_scsi_map_iov_to_sgl(struct vhost_scsi_cmd *cmd, struct iov_iter *iter,
 	struct scatterlist *sg = sg_table->sgl;
 	int ret;
 
+	if (is_prot)
+		cmd->tvc_prot_need_unpin = iov_iter_extract_will_pin(iter);
+	else
+		cmd->tvc_need_unpin = iov_iter_extract_will_pin(iter);
+
 	while (iov_iter_count(iter)) {
 		ret = vhost_scsi_map_to_sgl(cmd, iter, sg_table, &sg, is_prot);
 		if (ret < 0) {
 			vhost_scsi_revert_map_iov_to_sgl(iter, sg_table->sgl,
-							 sg);
+					sg, is_prot ? cmd->tvc_prot_need_unpin :
+						      cmd->tvc_need_unpin);
 			return ret;
 		}
 	}
@@ -1449,7 +1469,7 @@ vhost_scsi_handle_vq(struct vhost_scsi *vs, struct vhost_virtqueue *vq)
 		/*
 		 * If T10_PI header + payload is present, setup prot_iter values
 		 * and recalculate data_iter for vhost_scsi_mapal() mapping to
-		 * host scatterlists via get_user_pages_fast().
+		 * host scatterlists via iov_iter_extract_pages().
 		 */
 		if (t10_pi) {
 			if (v_req_pi.pi_bytesout) {

base-commit: 1a577f9402c1780cc1e663f48d5dd54c02051b09
-- 
2.43.0