[PATCH v6 0/3] vfio/pci: Introduce vfio_pci driver for ISM devices

[PATCH v6 0/3] vfio/pci: Introduce vfio_pci driver for ISM devices

[Julian Ruess]

Hi all,

This series adds a vfio_pci variant driver for the s390-specific
Internal Shared Memory (ISM) devices used for inter-VM communication
including SMC-D.

This is a prerequisite for an in-development open-source user space
driver stack that will allow to use ISM devices to provide remote
console and block device functionality. This stack will be part of
s390-tools.

This driver would also allow QEMU to mediate access to an ISM device,
enabling a form of PCI pass-through even for guests whose hardware
cannot directly execute PCI accesses, such as nested guests.

On s390, kernel primitives such as ioread() and iowrite() are switched
over from function handle based PCI load/stores instructions to PCI
memory-I/O (MIO) loads/stores when these are available and not
explicitly disabled. Since these instructions cannot be used with ISM
devices, ensure that classic function handle-based PCI instructions are
used instead.

The driver is still required even when MIO instructions are disabled, as
the ISM device relies on the PCI store‑block (PCISTB) instruction to
perform write operations.

Thank you,
Julian

Signed-off-by: Julian Ruess <julianr@linux.ibm.com>
---
Changes in v6:
- Introduce ism_vfio_pci_init_dev() and ism_vfio_pci_release_dev().
- Add early return for __zpci_load().
- Use kmem_cache_alloc() instead of kmem_cache_zalloc().
- Rename ivdev to ivpcd for consistency.
- Rename ism_pci_open_device() to ism_vfio_pci_open_device().
- Link to v5: https://lore.kernel.org/r/20260317-vfio_pci_ism-v5-0-b73248b4e576@linux.ibm.com

Changes in v5:
- Support reads < 8 bytes.
- Link to v4: https://lore.kernel.org/r/20260313-vfio_pci_ism-v4-0-4765ae056f71@linux.ibm.com

Changes in v4:
- Fix bug with < 8 byte reads. For code simplicity, only support 8 byte reads.
- Fix leak of ivpcd.
- Fix cache replacement by implementing a per-device kmem_cache.
- Link to v3: https://lore.kernel.org/r/20260305-vfio_pci_ism-v3-0-1217076c81d9@linux.ibm.com

Changes in v3:
- Add comments to ism_vfio_pci_do_io_r() and ism_vfio_pci_do_io_w().
- Format Kconfig.
- Add 4k boundary check to ism_vfio_pci_do_io_w().
- Use kmem_cache instead of kzalloc in ism_vfio_pci_do_io_w().
- Add error handler to struct ism_vfio_pci_driver.
- Link to v2: https://lore.kernel.org/r/20260224-vfio_pci_ism-v2-0-f010945373fa@linux.ibm.com

Changes in v2:
- Remove common code patch that sets VFIO_PCI_OFFSET_SHIFT to 48.
- Implement ism_vfio_pci_ioctl_get_region_info() to have own region
  offsets.
- For config space accesses, rename vfio_config_do_rw() to
  vfio_pci_config_rw_single() and export it.
- Use zdev->maxstbl instead of ZPCI_BOUNDARY_SIZE.
- Add comment that zPCI must not use MIO instructions for config space
  access.
- Rework patch descriptions.
- Update license info.
- Link to v1: https://lore.kernel.org/r/20260212-vfio_pci_ism-v1-0-333262ade074@linux.ibm.com

---
Julian Ruess (3):
      vfio/pci: Rename vfio_config_do_rw() to vfio_pci_config_rw_single() and export it
      vfio/ism: Implement vfio_pci driver for ISM devices
      MAINTAINERS: add VFIO ISM PCI DRIVER section

 MAINTAINERS                        |   6 +
 drivers/vfio/pci/Kconfig           |   2 +
 drivers/vfio/pci/Makefile          |   2 +
 drivers/vfio/pci/ism/Kconfig       |  10 +
 drivers/vfio/pci/ism/Makefile      |   3 +
 drivers/vfio/pci/ism/main.c        | 400 +++++++++++++++++++++++++++++++++++++
 drivers/vfio/pci/vfio_pci_config.c |   8 +-
 drivers/vfio/pci/vfio_pci_priv.h   |   4 +
 8 files changed, 432 insertions(+), 3 deletions(-)
---
base-commit: 8a30aeb0d1b4e4aaf7f7bae72f20f2ae75385ccb
change-id: 20250227-vfio_pci_ism-0ccc2e472247

Best regards,
-- 
Julian Ruess <julianr@linux.ibm.com>

[PATCH v6 1/3] vfio/pci: Rename vfio_config_do_rw() to vfio_pci_config_rw_single() and export it

[Julian Ruess]

A follow-up patch adds a new variant driver for s390 ISM devices. Since
this device uses a 256 TiB BAR 0 that is never mapped, the variant
driver needs its own ISM_VFIO_PCI_OFFSET_MASK. To minimally mirror the
functionality of vfio_pci_config_rw() with such a custom mask, export
vfio_config_do_rw(). To better distinguish the now exported function
from vfio_pci_config_rw(), rename it to vfio_pci_config_rw_single()
emphasizing that it does a single config space read or write.

Reviewed-by: Niklas Schnelle <schnelle@linux.ibm.com>
Signed-off-by: Julian Ruess <julianr@linux.ibm.com>
---
 drivers/vfio/pci/vfio_pci_config.c | 8 +++++---
 drivers/vfio/pci/vfio_pci_priv.h   | 4 ++++
 2 files changed, 9 insertions(+), 3 deletions(-)

diff --git a/drivers/vfio/pci/vfio_pci_config.c b/drivers/vfio/pci/vfio_pci_config.c
index b4e39253f98da61a5e2b6dd0089b2f6aef4b85a0..fbb47b4ddb43d42b758b16778e6e701379d7e7db 100644
--- a/drivers/vfio/pci/vfio_pci_config.c
+++ b/drivers/vfio/pci/vfio_pci_config.c
@@ -1880,8 +1880,9 @@ static size_t vfio_pci_cap_remaining_dword(struct vfio_pci_core_device *vdev,
 	return i;
 }
 
-static ssize_t vfio_config_do_rw(struct vfio_pci_core_device *vdev, char __user *buf,
-				 size_t count, loff_t *ppos, bool iswrite)
+ssize_t vfio_pci_config_rw_single(struct vfio_pci_core_device *vdev,
+			      char __user *buf, size_t count, loff_t *ppos,
+			      bool iswrite)
 {
 	struct pci_dev *pdev = vdev->pdev;
 	struct perm_bits *perm;
@@ -1970,6 +1971,7 @@ static ssize_t vfio_config_do_rw(struct vfio_pci_core_device *vdev, char __user
 
 	return ret;
 }
+EXPORT_SYMBOL_GPL(vfio_pci_config_rw_single);
 
 ssize_t vfio_pci_config_rw(struct vfio_pci_core_device *vdev, char __user *buf,
 			   size_t count, loff_t *ppos, bool iswrite)
@@ -1981,7 +1983,7 @@ ssize_t vfio_pci_config_rw(struct vfio_pci_core_device *vdev, char __user *buf,
 	pos &= VFIO_PCI_OFFSET_MASK;
 
 	while (count) {
-		ret = vfio_config_do_rw(vdev, buf, count, &pos, iswrite);
+		ret = vfio_pci_config_rw_single(vdev, buf, count, &pos, iswrite);
 		if (ret < 0)
 			return ret;
 
diff --git a/drivers/vfio/pci/vfio_pci_priv.h b/drivers/vfio/pci/vfio_pci_priv.h
index 27ac280f00b975989f6cbc02c11aaca01f9badf3..28a3edf65aeecfa06cd1856637cd33eec1fa3006 100644
--- a/drivers/vfio/pci/vfio_pci_priv.h
+++ b/drivers/vfio/pci/vfio_pci_priv.h
@@ -37,6 +37,10 @@ int vfio_pci_set_irqs_ioctl(struct vfio_pci_core_device *vdev, uint32_t flags,
 ssize_t vfio_pci_config_rw(struct vfio_pci_core_device *vdev, char __user *buf,
 			   size_t count, loff_t *ppos, bool iswrite);
 
+ssize_t vfio_pci_config_rw_single(struct vfio_pci_core_device *vdev,
+			      char __user *buf, size_t count, loff_t *ppos,
+			      bool iswrite);
+
 ssize_t vfio_pci_bar_rw(struct vfio_pci_core_device *vdev, char __user *buf,
 			size_t count, loff_t *ppos, bool iswrite);
 

-- 
2.51.0

[PATCH v6 2/3] vfio/ism: Implement vfio_pci driver for ISM devices

[Julian Ruess]

Add a vfio_pci variant driver for the s390-specific Internal Shared
Memory (ISM) devices used for inter-VM communication.

This enables the development of vfio-pci-based user space drivers for
ISM devices.

On s390, kernel primitives such as ioread() and iowrite() are switched
over from function handle based PCI load/stores instructions to PCI
memory-I/O (MIO) loads/stores when these are available and not
explicitly disabled. Since these instructions cannot be used with ISM
devices, ensure that classic function handle-based PCI instructions are
used instead.

The driver is still required even when MIO instructions are disabled, as
the ISM device relies on the PCI store block (PCISTB) instruction to
perform write operations.

Stores are not fragmented, therefore one ioctl corresponds to exactly
one PCISTB instruction. User space must ensure to not write more than
4096 bytes at once to an ISM BAR which is the maximum payload of the
PCISTB instruction.

Signed-off-by: Julian Ruess <julianr@linux.ibm.com>
---
 drivers/vfio/pci/Kconfig      |   2 +
 drivers/vfio/pci/Makefile     |   2 +
 drivers/vfio/pci/ism/Kconfig  |  10 ++
 drivers/vfio/pci/ism/Makefile |   3 +
 drivers/vfio/pci/ism/main.c   | 400 ++++++++++++++++++++++++++++++++++++++++++
 5 files changed, 417 insertions(+)

diff --git a/drivers/vfio/pci/Kconfig b/drivers/vfio/pci/Kconfig
index 1e82b44bda1a0a544e1add7f4b06edecf35aaf81..296bf01e185ecacc388ebc69e92706c99e47c814 100644
--- a/drivers/vfio/pci/Kconfig
+++ b/drivers/vfio/pci/Kconfig
@@ -60,6 +60,8 @@ config VFIO_PCI_DMABUF
 
 source "drivers/vfio/pci/mlx5/Kconfig"
 
+source "drivers/vfio/pci/ism/Kconfig"
+
 source "drivers/vfio/pci/hisilicon/Kconfig"
 
 source "drivers/vfio/pci/pds/Kconfig"
diff --git a/drivers/vfio/pci/Makefile b/drivers/vfio/pci/Makefile
index e0a0757dd1d2b0bc69b7e4d79441d5cacf4e1cd8..6138f1bf241df04e7419f196b404abdf9b194050 100644
--- a/drivers/vfio/pci/Makefile
+++ b/drivers/vfio/pci/Makefile
@@ -11,6 +11,8 @@ obj-$(CONFIG_VFIO_PCI) += vfio-pci.o
 
 obj-$(CONFIG_MLX5_VFIO_PCI)           += mlx5/
 
+obj-$(CONFIG_ISM_VFIO_PCI)           += ism/
+
 obj-$(CONFIG_HISI_ACC_VFIO_PCI) += hisilicon/
 
 obj-$(CONFIG_PDS_VFIO_PCI) += pds/
diff --git a/drivers/vfio/pci/ism/Kconfig b/drivers/vfio/pci/ism/Kconfig
new file mode 100644
index 0000000000000000000000000000000000000000..02f47d25fed2d34c732b67b3a3655b64a7625467
--- /dev/null
+++ b/drivers/vfio/pci/ism/Kconfig
@@ -0,0 +1,10 @@
+# SPDX-License-Identifier: GPL-2.0
+config ISM_VFIO_PCI
+	tristate "VFIO support for ISM devices"
+	depends on S390
+	select VFIO_PCI_CORE
+	help
+	  This provides user space support for IBM Internal Shared Memory (ISM)
+	  Adapter devices using the VFIO framework.
+
+	  If you don't know what to do here, say N.
diff --git a/drivers/vfio/pci/ism/Makefile b/drivers/vfio/pci/ism/Makefile
new file mode 100644
index 0000000000000000000000000000000000000000..32cc3c66dd11395da85a2b6f05b3d97036ed8a35
--- /dev/null
+++ b/drivers/vfio/pci/ism/Makefile
@@ -0,0 +1,3 @@
+# SPDX-License-Identifier: GPL-2.0
+obj-$(CONFIG_ISM_VFIO_PCI) += ism-vfio-pci.o
+ism-vfio-pci-y := main.o
diff --git a/drivers/vfio/pci/ism/main.c b/drivers/vfio/pci/ism/main.c
new file mode 100644
index 0000000000000000000000000000000000000000..45e611f54795780cf09e540c859db889afa67323
--- /dev/null
+++ b/drivers/vfio/pci/ism/main.c
@@ -0,0 +1,400 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * vfio-ISM driver for s390
+ *
+ * Copyright IBM Corp.
+ */
+
+#include "../vfio_pci_priv.h"
+#include "linux/slab.h"
+
+#define ISM_VFIO_PCI_OFFSET_SHIFT   48
+#define ISM_VFIO_PCI_OFFSET_TO_INDEX(off) (off >> ISM_VFIO_PCI_OFFSET_SHIFT)
+#define ISM_VFIO_PCI_INDEX_TO_OFFSET(index) ((u64)(index) << ISM_VFIO_PCI_OFFSET_SHIFT)
+#define ISM_VFIO_PCI_OFFSET_MASK (((u64)(1) << ISM_VFIO_PCI_OFFSET_SHIFT) - 1)
+
+/*
+ * Use __zpci_load() to bypass automatic use of
+ * PCI MIO instructions which are not supported on ISM devices
+ */
+#define ISM_READ(size)                                                        \
+	static int ism_read##size(struct zpci_dev *zdev, int bar,             \
+				  ssize_t *filled, char __user *buf,          \
+				  loff_t off)                                 \
+	{                                                                     \
+		u64 req, tmp;                                                 \
+		u##size val;                                                  \
+		int ret;                                                      \
+									      \
+		req = ZPCI_CREATE_REQ(READ_ONCE(zdev->fh), bar, sizeof(val)); \
+		ret = __zpci_load(&tmp, req, off);                            \
+		if (ret)                                                      \
+			return ret;                                           \
+		val = (u##size)tmp;                                           \
+		if (copy_to_user(buf, &val, sizeof(val)))                     \
+			return -EFAULT;                                       \
+		*filled = sizeof(val);                                        \
+		return 0;						      \
+	}
+
+ISM_READ(64);
+ISM_READ(32);
+ISM_READ(16);
+ISM_READ(8);
+
+struct ism_vfio_pci_core_device {
+	struct vfio_pci_core_device core_device;
+	struct kmem_cache *store_block_cache;
+};
+
+static int ism_vfio_pci_open_device(struct vfio_device *core_vdev)
+{
+	struct ism_vfio_pci_core_device *ivpcd;
+	struct vfio_pci_core_device *vdev;
+	int ret;
+
+	ivpcd = container_of(core_vdev, struct ism_vfio_pci_core_device,
+			     core_device.vdev);
+	vdev = &ivpcd->core_device;
+
+	ret = vfio_pci_core_enable(vdev);
+	if (ret)
+		return ret;
+
+	vfio_pci_core_finish_enable(vdev);
+	return 0;
+}
+
+/*
+ * ism_vfio_pci_do_io_r()
+ *
+ * On s390, kernel primitives such as ioread() and iowrite() are switched over
+ * from function handle based PCI load/stores instructions to PCI memory-I/O (MIO)
+ * loads/stores when these are available and not explicitly disabled. Since these
+ * instructions cannot be used with ISM devices, ensure that classic function
+ * handle-based PCI instructions are used instead.
+ */
+static ssize_t ism_vfio_pci_do_io_r(struct vfio_pci_core_device *vdev,
+				    char __user *buf, loff_t off, size_t count,
+				    int bar)
+{
+	struct zpci_dev *zdev = to_zpci(vdev->pdev);
+	ssize_t done = 0;
+	int ret;
+
+	while (count) {
+		size_t filled;
+
+		if (count >= 8 && IS_ALIGNED(off, 8)) {
+			ret = ism_read64(zdev, bar, &filled, buf, off);
+			if (ret)
+				return ret;
+		} else if (count >= 4 && IS_ALIGNED(off, 4)) {
+			ret = ism_read32(zdev, bar, &filled, buf, off);
+			if (ret)
+				return ret;
+		} else if (count >= 2 && IS_ALIGNED(off, 2)) {
+			ret = ism_read16(zdev, bar, &filled, buf, off);
+			if (ret)
+				return ret;
+		} else {
+			ret = ism_read8(zdev, bar, &filled, buf, off);
+			if (ret)
+				return ret;
+		}
+
+		count -= filled;
+		done += filled;
+		off += filled;
+		buf += filled;
+	}
+
+	return done;
+}
+
+/*
+ * ism_vfio_pci_do_io_w()
+ *
+ * Ensure that the PCI store block (PCISTB) instruction is used as required by the
+ * ISM device. The ISM device also uses a 256 TiB BAR 0 for write operations,
+ * which requires a 48bit region address space (ISM_VFIO_PCI_OFFSET_SHIFT).
+ */
+static ssize_t ism_vfio_pci_do_io_w(struct vfio_pci_core_device *vdev,
+				    char __user *buf, loff_t off, size_t count,
+				    int bar)
+{
+	struct zpci_dev *zdev = to_zpci(vdev->pdev);
+	struct ism_vfio_pci_core_device *ivpcd;
+	ssize_t ret;
+	void *data;
+	u64 req;
+
+	if (count > zdev->maxstbl)
+		return -EINVAL;
+	if (((off % PAGE_SIZE) + count) > PAGE_SIZE)
+		return -EINVAL;
+
+	ivpcd = container_of(vdev, struct ism_vfio_pci_core_device,
+			     core_device);
+	data = kmem_cache_alloc(ivpcd->store_block_cache, GFP_KERNEL);
+	if (!data)
+		return -ENOMEM;
+
+	if (copy_from_user(data, buf, count)) {
+		ret = -EFAULT;
+		goto out_free;
+	}
+
+	req = ZPCI_CREATE_REQ(READ_ONCE(zdev->fh), bar, count);
+	ret = __zpci_store_block(data, req, off);
+	if (ret)
+		goto out_free;
+
+	ret = count;
+
+out_free:
+	kmem_cache_free(ivpcd->store_block_cache, data);
+	return ret;
+}
+
+static ssize_t ism_vfio_pci_bar_rw(struct vfio_pci_core_device *vdev,
+				   char __user *buf, size_t count, loff_t *ppos,
+				   bool iswrite)
+{
+	int bar = ISM_VFIO_PCI_OFFSET_TO_INDEX(*ppos);
+	loff_t pos = *ppos & ISM_VFIO_PCI_OFFSET_MASK;
+	resource_size_t end;
+	ssize_t done = 0;
+
+	if (pci_resource_start(vdev->pdev, bar))
+		end = pci_resource_len(vdev->pdev, bar);
+	else
+		return -EINVAL;
+
+	if (pos >= end)
+		return -EINVAL;
+
+	count = min(count, (size_t)(end - pos));
+
+	if (iswrite)
+		done = ism_vfio_pci_do_io_w(vdev, buf, pos, count, bar);
+	else
+		done = ism_vfio_pci_do_io_r(vdev, buf, pos, count, bar);
+
+	if (done >= 0)
+		*ppos += done;
+
+	return done;
+}
+
+static ssize_t ism_vfio_pci_config_rw(struct vfio_pci_core_device *vdev,
+				      char __user *buf, size_t count,
+				      loff_t *ppos, bool iswrite)
+{
+	loff_t pos = *ppos;
+	size_t done = 0;
+	int ret = 0;
+
+	pos &= ISM_VFIO_PCI_OFFSET_MASK;
+
+	while (count) {
+		/*
+		 * zPCI must not use MIO instructions for config space access,
+		 * so we can use common code path here.
+		 */
+		ret = vfio_pci_config_rw_single(vdev, buf, count, &pos, iswrite);
+		if (ret < 0)
+			return ret;
+
+		count -= ret;
+		done += ret;
+		buf += ret;
+		pos += ret;
+	}
+
+	*ppos += done;
+
+	return done;
+}
+
+static ssize_t ism_vfio_pci_rw(struct vfio_device *core_vdev, char __user *buf,
+			       size_t count, loff_t *ppos, bool iswrite)
+{
+	unsigned int index = ISM_VFIO_PCI_OFFSET_TO_INDEX(*ppos);
+	struct vfio_pci_core_device *vdev;
+	int ret;
+
+	vdev = container_of(core_vdev, struct vfio_pci_core_device, vdev);
+
+	if (!count)
+		return 0;
+
+	switch (index) {
+	case VFIO_PCI_CONFIG_REGION_INDEX:
+		ret = ism_vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
+		break;
+
+	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
+		ret = ism_vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
+		break;
+
+	default:
+		return -EINVAL;
+	}
+
+	return ret;
+}
+
+static ssize_t ism_vfio_pci_read(struct vfio_device *core_vdev,
+				 char __user *buf, size_t count, loff_t *ppos)
+{
+	return ism_vfio_pci_rw(core_vdev, buf, count, ppos, false);
+}
+
+static ssize_t ism_vfio_pci_write(struct vfio_device *core_vdev,
+				  const char __user *buf, size_t count,
+				  loff_t *ppos)
+{
+	return ism_vfio_pci_rw(core_vdev, (char __user *)buf, count, ppos,
+			       true);
+}
+
+static int ism_vfio_pci_ioctl_get_region_info(struct vfio_device *core_vdev,
+					      struct vfio_region_info *info,
+					      struct vfio_info_cap *caps)
+{
+	struct vfio_pci_core_device *vdev =
+		container_of(core_vdev, struct vfio_pci_core_device, vdev);
+	struct pci_dev *pdev = vdev->pdev;
+
+	switch (info->index) {
+	case VFIO_PCI_CONFIG_REGION_INDEX:
+		info->offset = ISM_VFIO_PCI_INDEX_TO_OFFSET(info->index);
+		info->size = pdev->cfg_size;
+		info->flags = VFIO_REGION_INFO_FLAG_READ |
+			      VFIO_REGION_INFO_FLAG_WRITE;
+		break;
+	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
+		info->offset = ISM_VFIO_PCI_INDEX_TO_OFFSET(info->index);
+		info->size = pci_resource_len(pdev, info->index);
+		if (!info->size) {
+			info->flags = 0;
+			break;
+		}
+		info->flags = VFIO_REGION_INFO_FLAG_READ |
+			      VFIO_REGION_INFO_FLAG_WRITE;
+		break;
+	default:
+		info->offset = 0;
+		info->size = 0;
+		info->flags = 0;
+	}
+	return 0;
+}
+
+static int ism_vfio_pci_init_dev(struct vfio_device *core_vdev)
+{
+	struct zpci_dev *zdev = to_zpci(to_pci_dev(core_vdev->dev));
+	struct ism_vfio_pci_core_device *ivpcd;
+	char cache_name[20];
+	int ret;
+
+	ivpcd = container_of(core_vdev, struct ism_vfio_pci_core_device,
+			     core_device.vdev);
+
+	snprintf(cache_name, sizeof(cache_name), "ism_sb_fid_%08x", zdev->fid);
+	ivpcd->store_block_cache =
+		kmem_cache_create(cache_name, zdev->maxstbl, 0, 0, NULL);
+	if (!ivpcd->store_block_cache)
+		return -ENOMEM;
+
+	ret = vfio_pci_core_init_dev(core_vdev);
+	if (ret)
+		kmem_cache_destroy(ivpcd->store_block_cache);
+
+	return ret;
+}
+
+static void ism_vfio_pci_release_dev(struct vfio_device *core_vdev)
+{
+	struct ism_vfio_pci_core_device *ivpcd = container_of(
+		core_vdev, struct ism_vfio_pci_core_device, core_device.vdev);
+
+	kmem_cache_destroy(ivpcd->store_block_cache);
+	vfio_pci_core_release_dev(core_vdev);
+}
+
+static const struct vfio_device_ops ism_pci_ops = {
+	.name = "ism-vfio-pci",
+	.init = ism_vfio_pci_init_dev,
+	.release = ism_vfio_pci_release_dev,
+	.open_device = ism_vfio_pci_open_device,
+	.close_device = vfio_pci_core_close_device,
+	.ioctl = vfio_pci_core_ioctl,
+	.get_region_info_caps = ism_vfio_pci_ioctl_get_region_info,
+	.device_feature = vfio_pci_core_ioctl_feature,
+	.read = ism_vfio_pci_read,
+	.write = ism_vfio_pci_write,
+	.request = vfio_pci_core_request,
+	.match = vfio_pci_core_match,
+	.match_token_uuid = vfio_pci_core_match_token_uuid,
+	.bind_iommufd = vfio_iommufd_physical_bind,
+	.unbind_iommufd = vfio_iommufd_physical_unbind,
+	.attach_ioas = vfio_iommufd_physical_attach_ioas,
+	.detach_ioas = vfio_iommufd_physical_detach_ioas,
+};
+
+static int ism_vfio_pci_probe(struct pci_dev *pdev,
+			      const struct pci_device_id *id)
+{
+	struct ism_vfio_pci_core_device *ivpcd;
+	int ret;
+
+	ivpcd = vfio_alloc_device(ism_vfio_pci_core_device, core_device.vdev,
+				  &pdev->dev, &ism_pci_ops);
+	if (IS_ERR(ivpcd))
+		return PTR_ERR(ivpcd);
+
+	dev_set_drvdata(&pdev->dev, &ivpcd->core_device);
+
+	ret = vfio_pci_core_register_device(&ivpcd->core_device);
+	if (ret)
+		vfio_put_device(&ivpcd->core_device.vdev);
+
+	return ret;
+}
+
+static void ism_vfio_pci_remove(struct pci_dev *pdev)
+{
+	struct vfio_pci_core_device *core_device;
+	struct ism_vfio_pci_core_device *ivpcd;
+
+	core_device = dev_get_drvdata(&pdev->dev);
+	ivpcd = container_of(core_device, struct ism_vfio_pci_core_device,
+			     core_device);
+
+	vfio_pci_core_unregister_device(&ivpcd->core_device);
+	vfio_put_device(&ivpcd->core_device.vdev);
+}
+
+static const struct pci_device_id ism_device_table[] = {
+	{ PCI_DRIVER_OVERRIDE_DEVICE_VFIO(PCI_VENDOR_ID_IBM,
+					  PCI_DEVICE_ID_IBM_ISM) },
+	{}
+};
+MODULE_DEVICE_TABLE(pci, ism_device_table);
+
+static struct pci_driver ism_vfio_pci_driver = {
+	.name = KBUILD_MODNAME,
+	.id_table = ism_device_table,
+	.probe = ism_vfio_pci_probe,
+	.remove = ism_vfio_pci_remove,
+	.err_handler = &vfio_pci_core_err_handlers,
+	.driver_managed_dma = true,
+};
+
+module_pci_driver(ism_vfio_pci_driver);
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("vfio-pci variant driver for the IBM Internal Shared Memory (ISM) device");
+MODULE_AUTHOR("IBM Corporation");

-- 
2.51.0

[PATCH v6 3/3] MAINTAINERS: add VFIO ISM PCI DRIVER section

[Julian Ruess]

ism_vfio_pci is a new kernel component that allows
to use the ISM device from userspace. Add myself
as a maintainer.

Reviewed-by: Niklas Schnelle <schnelle@linux.ibm.com>
Signed-off-by: Julian Ruess <julianr@linux.ibm.com>
---
 MAINTAINERS | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/MAINTAINERS b/MAINTAINERS
index d7241695df968f1f62f463f1cc01b2d3201e91ee..9d07dc0ae2547bb97f6ac05163f423c0b5b86ea7 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -27702,6 +27702,12 @@ L:	kvm@vger.kernel.org
 S:	Maintained
 F:	drivers/vfio/pci/hisilicon/
 
+VFIO ISM PCI DRIVER
+M:	Julian Ruess <julianr@linux.ibm.com>
+L:	kvm@vger.kernel.org
+S:	Maintained
+F:	drivers/vfio/pci/ism/
+
 VFIO MEDIATED DEVICE DRIVERS
 M:	Kirti Wankhede <kwankhede@nvidia.com>
 L:	kvm@vger.kernel.org

-- 
2.51.0

Re: [PATCH v6 2/3] vfio/ism: Implement vfio_pci driver for ISM devices

[Niklas Schnelle]

On Thu, 2026-03-19 at 16:42 +0100, Julian Ruess wrote:
> Add a vfio_pci variant driver for the s390-specific Internal Shared
> Memory (ISM) devices used for inter-VM communication.
> 
> This enables the development of vfio-pci-based user space drivers for
> ISM devices.
> 
> On s390, kernel primitives such as ioread() and iowrite() are switched
> over from function handle based PCI load/stores instructions to PCI
> memory-I/O (MIO) loads/stores when these are available and not
> explicitly disabled. Since these instructions cannot be used with ISM
> devices, ensure that classic function handle-based PCI instructions are
> used instead.
> 
> The driver is still required even when MIO instructions are disabled, as
> the ISM device relies on the PCI store block (PCISTB) instruction to
> perform write operations.
> 
> Stores are not fragmented, therefore one ioctl corresponds to exactly
> one PCISTB instruction. User space must ensure to not write more than
> 4096 bytes at once to an ISM BAR which is the maximum payload of the
> PCISTB instruction.
> 
> Signed-off-by: Julian Ruess <julianr@linux.ibm.com>
> ---
>  drivers/vfio/pci/Kconfig      |   2 +
>  drivers/vfio/pci/Makefile     |   2 +
>  drivers/vfio/pci/ism/Kconfig  |  10 ++
>  drivers/vfio/pci/ism/Makefile |   3 +
>  drivers/vfio/pci/ism/main.c   | 400 ++++++++++++++++++++++++++++++++++++++++++
>  5 files changed, 417 insertions(+)
> 

As a general note, ISM devices do not support any low power/sleep modes
which is why there is no PM handling done.

> +
> +/*
> + * ism_vfio_pci_do_io_w()
> + *
> + * Ensure that the PCI store block (PCISTB) instruction is used as required by the
> + * ISM device. The ISM device also uses a 256 TiB BAR 0 for write operations,
> + * which requires a 48bit region address space (ISM_VFIO_PCI_OFFSET_SHIFT).
> + */
> +static ssize_t ism_vfio_pci_do_io_w(struct vfio_pci_core_device *vdev,
> +				    char __user *buf, loff_t off, size_t count,
> +				    int bar)
> +{
> +	struct zpci_dev *zdev = to_zpci(vdev->pdev);
> +	struct ism_vfio_pci_core_device *ivpcd;
> +	ssize_t ret;
> +	void *data;
> +	u64 req;
> +
> +	if (count > zdev->maxstbl)
> +		return -EINVAL;
> +	if (((off % PAGE_SIZE) + count) > PAGE_SIZE)
> +		return -EINVAL;
> +
> +	ivpcd = container_of(vdev, struct ism_vfio_pci_core_device,
> +			     core_device);
> +	data = kmem_cache_alloc(ivpcd->store_block_cache, GFP_KERNEL);
> +	if (!data)
> +		return -ENOMEM;
> +
> +	if (copy_from_user(data, buf, count)) {
> +		ret = -EFAULT;
> +		goto out_free;
> +	}
> +
> +	req = ZPCI_CREATE_REQ(READ_ONCE(zdev->fh), bar, count);
> +	ret = __zpci_store_block(data, req, off);

Note: There is a Sashiko finding that PCI Store Block needs 8 byte
alignment and we don't ensure that off is 8 byte aligned here. While
generally true the ISM device has relaxed alignment rules and this
requirement does not apply. That said we should set PAGE_SIZE alignment
on the kmem_cache such that data is guaranteed PAGE_SIZE aligned and
the page crossing check works correctly.

> +	if (ret)
> +		goto out_free;
> +
> +	ret = count;
> +
> +out_free:
> +	kmem_cache_free(ivpcd->store_block_cache, data);
> +	return ret;
> +}
> +
--- snip ---

> +
> +static int ism_vfio_pci_init_dev(struct vfio_device *core_vdev)
> +{
> +	struct zpci_dev *zdev = to_zpci(to_pci_dev(core_vdev->dev));
> +	struct ism_vfio_pci_core_device *ivpcd;
> +	char cache_name[20];
> +	int ret;
> +
> +	ivpcd = container_of(core_vdev, struct ism_vfio_pci_core_device,
> +			     core_device.vdev);
> +
> +	snprintf(cache_name, sizeof(cache_name), "ism_sb_fid_%08x", zdev->fid);
> +	ivpcd->store_block_cache =
> +		kmem_cache_create(cache_name, zdev->maxstbl, 0, 0, NULL);

Sashiko notes here that the cache_name is stack allocated and says that
kmem_cache_create() doesn't copy it. Looking at
__kmem_cache_create_args() however this seems to be not true as that
does krstrdup_const() on the passed-in name.

> +	if (!ivpcd->store_block_cache)
> +		return -ENOMEM;
> +
> +	ret = vfio_pci_core_init_dev(core_vdev);
> +	if (ret)
> +		kmem_cache_destroy(ivpcd->store_block_cache);
> +
> +	return ret;
> +}
> +
--- snip ---

Re: [PATCH v6 2/3] vfio/ism: Implement vfio_pci driver for ISM devices

[Julian Ruess]

On Fri Mar 20, 2026 at 1:54 PM CET, Niklas Schnelle wrote:
> On Thu, 2026-03-19 at 16:42 +0100, Julian Ruess wrote:
>> Add a vfio_pci variant driver for the s390-specific Internal Shared
>> Memory (ISM) devices used for inter-VM communication.
>> 
>> This enables the development of vfio-pci-based user space drivers for
>> ISM devices.
>> 
>> On s390, kernel primitives such as ioread() and iowrite() are switched
>> over from function handle based PCI load/stores instructions to PCI
>> memory-I/O (MIO) loads/stores when these are available and not
>> explicitly disabled. Since these instructions cannot be used with ISM
>> devices, ensure that classic function handle-based PCI instructions are
>> used instead.
>> 
>> The driver is still required even when MIO instructions are disabled, as
>> the ISM device relies on the PCI store block (PCISTB) instruction to
>> perform write operations.
>> 
>> Stores are not fragmented, therefore one ioctl corresponds to exactly
>> one PCISTB instruction. User space must ensure to not write more than
>> 4096 bytes at once to an ISM BAR which is the maximum payload of the
>> PCISTB instruction.
>> 
>> Signed-off-by: Julian Ruess <julianr@linux.ibm.com>
>> ---
>>  drivers/vfio/pci/Kconfig      |   2 +
>>  drivers/vfio/pci/Makefile     |   2 +
>>  drivers/vfio/pci/ism/Kconfig  |  10 ++
>>  drivers/vfio/pci/ism/Makefile |   3 +
>>  drivers/vfio/pci/ism/main.c   | 400 ++++++++++++++++++++++++++++++++++++++++++
>>  5 files changed, 417 insertions(+)
>> 
>
> As a general note, ISM devices do not support any low power/sleep modes
> which is why there is no PM handling done.
>
>> +
>> +/*
>> + * ism_vfio_pci_do_io_w()
>> + *
>> + * Ensure that the PCI store block (PCISTB) instruction is used as required by the
>> + * ISM device. The ISM device also uses a 256 TiB BAR 0 for write operations,
>> + * which requires a 48bit region address space (ISM_VFIO_PCI_OFFSET_SHIFT).
>> + */
>> +static ssize_t ism_vfio_pci_do_io_w(struct vfio_pci_core_device *vdev,
>> +				    char __user *buf, loff_t off, size_t count,
>> +				    int bar)
>> +{
>> +	struct zpci_dev *zdev = to_zpci(vdev->pdev);
>> +	struct ism_vfio_pci_core_device *ivpcd;
>> +	ssize_t ret;
>> +	void *data;
>> +	u64 req;
>> +
>> +	if (count > zdev->maxstbl)
>> +		return -EINVAL;
>> +	if (((off % PAGE_SIZE) + count) > PAGE_SIZE)
>> +		return -EINVAL;
>> +
>> +	ivpcd = container_of(vdev, struct ism_vfio_pci_core_device,
>> +			     core_device);
>> +	data = kmem_cache_alloc(ivpcd->store_block_cache, GFP_KERNEL);
>> +	if (!data)
>> +		return -ENOMEM;
>> +
>> +	if (copy_from_user(data, buf, count)) {
>> +		ret = -EFAULT;
>> +		goto out_free;
>> +	}
>> +
>> +	req = ZPCI_CREATE_REQ(READ_ONCE(zdev->fh), bar, count);
>> +	ret = __zpci_store_block(data, req, off);
>
> Note: There is a Sashiko finding that PCI Store Block needs 8 byte
> alignment and we don't ensure that off is 8 byte aligned here. While
> generally true the ISM device has relaxed alignment rules and this
> requirement does not apply. That said we should set PAGE_SIZE alignment
> on the kmem_cache such that data is guaranteed PAGE_SIZE aligned and
> the page crossing check works correctly.
>
>> +	if (ret)
>> +		goto out_free;
>> +
>> +	ret = count;
>> +
>> +out_free:
>> +	kmem_cache_free(ivpcd->store_block_cache, data);
>> +	return ret;
>> +}
>> +
> --- snip ---
>
>> +
>> +static int ism_vfio_pci_init_dev(struct vfio_device *core_vdev)
>> +{
>> +	struct zpci_dev *zdev = to_zpci(to_pci_dev(core_vdev->dev));
>> +	struct ism_vfio_pci_core_device *ivpcd;
>> +	char cache_name[20];
>> +	int ret;
>> +
>> +	ivpcd = container_of(core_vdev, struct ism_vfio_pci_core_device,
>> +			     core_device.vdev);
>> +
>> +	snprintf(cache_name, sizeof(cache_name), "ism_sb_fid_%08x", zdev->fid);
>> +	ivpcd->store_block_cache =
>> +		kmem_cache_create(cache_name, zdev->maxstbl, 0, 0, NULL);
>
> Sashiko notes here that the cache_name is stack allocated and says that
> kmem_cache_create() doesn't copy it. Looking at
> __kmem_cache_create_args() however this seems to be not true as that
> does krstrdup_const() on the passed-in name.

I think we should change this to:

ivpcd->store_block_cache = kmem_cache_create(
       cache_name, zdev->maxstbl, PAGE_SIZE,
       (SLAB_RECLAIM_ACCOUNT | SLAB_ACCOUNT | SLAB_PANIC), NULL);

The alignment ensures that we do not cross the integral boundary, and as
Sashiko said, memcg accounting is also useful here.

-- snip --

Re: [PATCH v6 2/3] vfio/ism: Implement vfio_pci driver for ISM devices

[Niklas Schnelle]

On Fri, 2026-03-20 at 14:37 +0100, Julian Ruess wrote:
> On Fri Mar 20, 2026 at 1:54 PM CET, Niklas Schnelle wrote:
> > On Thu, 2026-03-19 at 16:42 +0100, Julian Ruess wrote:
> > > Add a vfio_pci variant driver for the s390-specific Internal Shared
> > > Memory (ISM) devices used for inter-VM communication.
> > > 
> > > This enables the development of vfio-pci-based user space drivers for
> > > ISM devices.
> > > 
> > > On s390, kernel primitives such as ioread() and iowrite() are switched
> > > over from function handle based PCI load/stores instructions to PCI
> > > memory-I/O (MIO) loads/stores when these are available and not
> > > explicitly disabled. Since these instructions cannot be used with ISM
> > > devices, ensure that classic function handle-based PCI instructions are
> > > used instead.
> > > 
> > > The driver is still required even when MIO instructions are disabled, as
> > > the ISM device relies on the PCI store block (PCISTB) instruction to
> > > perform write operations.
> > > 
> > > Stores are not fragmented, therefore one ioctl corresponds to exactly
> > > one PCISTB instruction. User space must ensure to not write more than
> > > 4096 bytes at once to an ISM BAR which is the maximum payload of the
> > > PCISTB instruction.
> > > 
> > > Signed-off-by: Julian Ruess <julianr@linux.ibm.com>
> > > ---
> > >  drivers/vfio/pci/Kconfig      |   2 +
> > >  drivers/vfio/pci/Makefile     |   2 +
> > >  drivers/vfio/pci/ism/Kconfig  |  10 ++
> > >  drivers/vfio/pci/ism/Makefile |   3 +
> > >  drivers/vfio/pci/ism/main.c   | 400 ++++++++++++++++++++++++++++++++++++++++++
> > >  5 files changed, 417 insertions(+)
> > > 
> > 
> > As a general note, ISM devices do not support any low power/sleep modes
> > which is why there is no PM handling done.
> > 
> > > +
> > > +/*
> > > + * ism_vfio_pci_do_io_w()
> > > + *
> > > + * Ensure that the PCI store block (PCISTB) instruction is used as required by the
> > > + * ISM device. The ISM device also uses a 256 TiB BAR 0 for write operations,
> > > + * which requires a 48bit region address space (ISM_VFIO_PCI_OFFSET_SHIFT).
> > > + */
> > > +static ssize_t ism_vfio_pci_do_io_w(struct vfio_pci_core_device *vdev,
> > > +				    char __user *buf, loff_t off, size_t count,
> > > +				    int bar)
> > > +{
> > > +	struct zpci_dev *zdev = to_zpci(vdev->pdev);
> > > +	struct ism_vfio_pci_core_device *ivpcd;
> > > +	ssize_t ret;
> > > +	void *data;
> > > +	u64 req;
> > > +
> > > +	if (count > zdev->maxstbl)
> > > +		return -EINVAL;
> > > +	if (((off % PAGE_SIZE) + count) > PAGE_SIZE)
> > > +		return -EINVAL;
> > > +
> > > +	ivpcd = container_of(vdev, struct ism_vfio_pci_core_device,
> > > +			     core_device);
> > > +	data = kmem_cache_alloc(ivpcd->store_block_cache, GFP_KERNEL);
> > > +	if (!data)
> > > +		return -ENOMEM;
> > > +
> > > +	if (copy_from_user(data, buf, count)) {
> > > +		ret = -EFAULT;
> > > +		goto out_free;
> > > +	}
> > > +
> > > +	req = ZPCI_CREATE_REQ(READ_ONCE(zdev->fh), bar, count);
> > > +	ret = __zpci_store_block(data, req, off);
> > 
> > Note: There is a Sashiko finding that PCI Store Block needs 8 byte
> > alignment and we don't ensure that off is 8 byte aligned here. While
> > generally true the ISM device has relaxed alignment rules and this
> > requirement does not apply. That said we should set PAGE_SIZE alignment
> > on the kmem_cache such that data is guaranteed PAGE_SIZE aligned and
> > the page crossing check works correctly.
> > 
> > > +	if (ret)
> > > +		goto out_free;
> > > +
> > > +	ret = count;
> > > +
> > > +out_free:
> > > +	kmem_cache_free(ivpcd->store_block_cache, data);
> > > +	return ret;
> > > +}
> > > +
> > --- snip ---
> > 
> > > +
> > > +static int ism_vfio_pci_init_dev(struct vfio_device *core_vdev)
> > > +{
> > > +	struct zpci_dev *zdev = to_zpci(to_pci_dev(core_vdev->dev));
> > > +	struct ism_vfio_pci_core_device *ivpcd;
> > > +	char cache_name[20];
> > > +	int ret;
> > > +
> > > +	ivpcd = container_of(core_vdev, struct ism_vfio_pci_core_device,
> > > +			     core_device.vdev);
> > > +
> > > +	snprintf(cache_name, sizeof(cache_name), "ism_sb_fid_%08x", zdev->fid);
> > > +	ivpcd->store_block_cache =
> > > +		kmem_cache_create(cache_name, zdev->maxstbl, 0, 0, NULL);
> > 
> > Sashiko notes here that the cache_name is stack allocated and says that
> > kmem_cache_create() doesn't copy it. Looking at
> > __kmem_cache_create_args() however this seems to be not true as that
> > does krstrdup_const() on the passed-in name.
> 
> I think we should change this to:
> 
> ivpcd->store_block_cache = kmem_cache_create(
>        cache_name, zdev->maxstbl, PAGE_SIZE,
>        (SLAB_RECLAIM_ACCOUNT | SLAB_ACCOUNT | SLAB_PANIC), NULL);
> 
> The alignment ensures that we do not cross the integral boundary, and as
> Sashiko said, memcg accounting is also useful here.

We definitely do *not* want to panic when kmem_cache_create() fails,
that would take down the whole kernel just for a PCI device! The others
make sense though I'm not sure if SLAB_RECLAIM_ACCOUNT needs other
setup.

Thanks,
Niklas