Skip to contentSkip to conversation

x86/hyperv: Use memremap()/memunmap() instead of ioremap_cache()/iounmap()

PATCH Applied [email protected] 2 participants 2 messages Jan 19 – Feb 4

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/hyperv: Use memremap()/memunmap() instead of ioremap_cache()/iounmap()

From: Michael Kelley <[email protected]>

When running with a paravisor and SEV-SNP, the GHCB page is provided by the paravisor instead of being allocated by Linux. The provided page is normal memory, but is outside of the physical address space seen by Linux. As such it cannot be accessed via the kernel's direct map, and must be explicitly mapped to a kernel virtual address.

Current code uses ioremap_cache() and iounmap() to map and unmap the page. These functions are for use on I/O address space that may not behave as normal memory, so they generate or expect addresses with the __iomem attribute. For normal memory, the preferred functions are memremap() and memunmap(), which operate similarly but without __iomem.

At the time of the original work on CoCo VMs on Hyper-V, memremap() did not support creating a decrypted mapping, so ioremap_cache() was used instead, since I/O address space is always mapped decrypted. memremap() has since been enhanced to allow decrypted mappings, so replace ioremap_cache() with memremap() when mapping the GHCB page. Similarly, replace iounmap() with memunmap(). As a side benefit, the replacement cleans up 'sparse' warnings about __iomem mismatches.

The replacement is done to use the correct functions as long-term goodness and to clean up the sparse warnings. No runtime bugs are fixed.

---
 arch/x86/hyperv/hv_init.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/x86/hyperv/hv_init.c +3 -3
index 14de43f4bc6c..ca992ca6394c 100644
@@ -103,9 +103,9 @@ static int hyperv_init_ghcb(void)
 	 */
 	rdmsrq(MSR_AMD64_SEV_ES_GHCB, ghcb_gpa);
 
-	/* Mask out vTOM bit. ioremap_cache() maps decrypted */
+	/* Mask out vTOM bit and map as decrypted */
 	ghcb_gpa &= ~ms_hyperv.shared_gpa_boundary;
-	ghcb_va = (void *)ioremap_cache(ghcb_gpa, HV_HYP_PAGE_SIZE);
+	ghcb_va = memremap(ghcb_gpa, HV_HYP_PAGE_SIZE, MEMREMAP_WB | MEMREMAP_DEC);
 	if (!ghcb_va)
 		return -ENOMEM;
 
@@ -277,7 +277,7 @@ static int hv_cpu_die(unsigned int cpu)
 	if (hv_ghcb_pg) {
 		ghcb_va = (void **)this_cpu_ptr(hv_ghcb_pg);
 		if (*ghcb_va)
-			iounmap(*ghcb_va);
+			memunmap(*ghcb_va);
 		*ghcb_va = NULL;
 	}
 
Signature · 1 line
2.25.1
maintainer Applied
#1 lore raw
Quoted 22 lines · On Mon, Jan 19, 2026 at 07:59:37AM -0800, [email protected] wrote:

From: Michael Kelley <[email protected]>

When running with a paravisor and SEV-SNP, the GHCB page is provided by the paravisor instead of being allocated by Linux. The provided page is normal memory, but is outside of the physical address space seen by Linux. As such it cannot be accessed via the kernel's direct map, and must be explicitly mapped to a kernel virtual address.

Current code uses ioremap_cache() and iounmap() to map and unmap the page. These functions are for use on I/O address space that may not behave as normal memory, so they generate or expect addresses with the __iomem attribute. For normal memory, the preferred functions are memremap() and memunmap(), which operate similarly but without __iomem.

At the time of the original work on CoCo VMs on Hyper-V, memremap() did not support creating a decrypted mapping, so ioremap_cache() was used instead, since I/O address space is always mapped decrypted. memremap() has since been enhanced to allow decrypted mappings, so replace ioremap_cache() with memremap() when mapping the GHCB page. Similarly, replace iounmap() with memunmap(). As a side benefit, the replacement cleans up 'sparse' warnings about __iomem mismatches.

Quoted 6 lines

The replacement is done to use the correct functions as long-term goodness and to clean up the sparse warnings. No runtime bugs are fixed.

Applied to hyperv-next. Thanks.