Skip to contentSkip to conversation

vfio/pci: Support 8-byte PCI loads and stores

PATCH Superseded Gerd Bayer 3 participants 7 messages Apr 22, 2024 – Apr 23, 2024

Subsystems (3)

v2 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Gerd Bayer

author
#0 lore raw

PATCH v2 1/1 vfio/pci: Support 8-byte PCI loads and stores

From: Ben Segal <[email protected]>

Many PCI adapters can benefit or even require full 64bit read and write access to their registers. In order to enable work on user-space drivers for these devices add two new variations vfio_pci_core_io{read|write}64 of the existing access methods when the architecture supports 64-bit ioreads and iowrites.

Since these access methods are instantiated on 64bit architectures, only, their use in vfio_pci_core_do_io_rw() is restricted by conditional compiles to these architectures.

---
Hi all,

we've successfully used this patch with a user-mode driver for a PCI device that requires 64bit register read/writes on s390. A quick grep showed that there are several other drivers for PCI devices in the kernel that use readq/writeq and eventually could use this, too.
So we decided to propose this for general inclusion.

Thank you,
Gerd Bayer

Changes v1 -> v2:
- On non 64bit architecture use at most 32bit accesses in vfio_pci_core_do_io_rw and describe that in the commit message.
- Drop the run-time error on 32bit architectures.
- The #endif splitting the "else if" is not really fortunate, but I'm open to suggestions.

 drivers/vfio/pci/vfio_pci_rdwr.c | 28 ++++++++++++++++++++++++++++
 include/linux/vfio_pci_core.h    |  3 +++
 2 files changed, 31 insertions(+)
drivers/vfio/pci/vfio_pci_rdwr.c +28 -0
index 03b8f7ada1ac..d83cb0bb7aa5 100644
@@ -89,6 +89,9 @@ EXPORT_SYMBOL_GPL(vfio_pci_core_ioread##size);
 VFIO_IOREAD(8)
 VFIO_IOREAD(16)
 VFIO_IOREAD(32)
+#ifdef ioread64
+VFIO_IOREAD(64)
+#endif
 
 /*
  * Read or write from an __iomem region (MMIO or I/O port) with an excluded
@@ -114,6 +117,31 @@ ssize_t vfio_pci_core_do_io_rw(struct vfio_pci_core_device *vdev, bool test_mem,
 		else
 			fillable = 0;
 
+#if defined(ioread64) && defined(iowrite64)
+		if (fillable >= 8 && !(off % 8)) {
+			u64 val;
+
+			if (iswrite) {
+				if (copy_from_user(&val, buf, 8))
+					return -EFAULT;
+
+				ret = vfio_pci_core_iowrite64(vdev, test_mem,
+							 val, io + off);
+				if (ret)
+					return ret;
+			} else {
+				ret = vfio_pci_core_ioread64(vdev, test_mem,
+							&val, io + off);
+				if (ret)
+					return ret;
+
+				if (copy_to_user(buf, &val, 8))
+					return -EFAULT;
+			}
+
+			filled = 8;
+		} else
+#endif /* defined(ioread64) && defined(iowrite64) */
 		if (fillable >= 4 && !(off % 4)) {
 			u32 val;
 
include/linux/vfio_pci_core.h +3 -0
index a2c8b8bba711..f4cf5fd2350c 100644
@@ -157,5 +157,8 @@ int vfio_pci_core_ioread##size(struct vfio_pci_core_device *vdev,	\
 VFIO_IOREAD_DECLATION(8)
 VFIO_IOREAD_DECLATION(16)
 VFIO_IOREAD_DECLATION(32)
+#ifdef ioread64
+VFIO_IOREAD_DECLATION(64)
+#endif
 
 #endif /* VFIO_PCI_CORE_H */
Signature · 1 line
2.44.0
#1 lore raw
Quoted 27 lines · Gerd Bayer

From: Ben Segal <[email protected]>

Many PCI adapters can benefit or even require full 64bit read and write access to their registers. In order to enable work on user-space drivers for these devices add two new variations vfio_pci_core_io{read|write}64 of the existing access methods when the architecture supports 64-bit ioreads and iowrites.

Since these access methods are instantiated on 64bit architectures, only, their use in vfio_pci_core_do_io_rw() is restricted by conditional compiles to these architectures.

---
Hi all,

we've successfully used this patch with a user-mode driver for a PCI device that requires 64bit register read/writes on s390. A quick grep showed that there are several other drivers for PCI devices in the kernel that use readq/writeq and eventually could use this, too.
So we decided to propose this for general inclusion.

Thank you,
Gerd Bayer

Quoted 6 lines

Changes v1 -> v2:
- On non 64bit architecture use at most 32bit accesses in vfio_pci_core_do_io_rw and describe that in the commit message.
- Drop the run-time error on 32bit architectures.
- The #endif splitting the "else if" is not really fortunate, but I'm open to suggestions.

Provide a iowrite64() that does back to back writes for 32 bit?

Jason

#2 lore raw
Quoted 28 lines · On Mon, 22 Apr 2024 14:43:05 -0300 Jason Gunthorpe <[email protected]> wrote:
Quoted 27 lines · Gerd Bayer

From: Ben Segal <[email protected]>

Many PCI adapters can benefit or even require full 64bit read and write access to their registers. In order to enable work on user-space drivers for these devices add two new variations vfio_pci_core_io{read|write}64 of the existing access methods when the architecture supports 64-bit ioreads and iowrites.

Since these access methods are instantiated on 64bit architectures, only, their use in vfio_pci_core_do_io_rw() is restricted by conditional compiles to these architectures.

---
Hi all,

we've successfully used this patch with a user-mode driver for a PCI device that requires 64bit register read/writes on s390. A quick grep showed that there are several other drivers for PCI devices in the kernel that use readq/writeq and eventually could use this, too.
So we decided to propose this for general inclusion.

Thank you,
Gerd Bayer

Quoted 8 lines
Quoted 6 lines

Changes v1 -> v2:
- On non 64bit architecture use at most 32bit accesses in vfio_pci_core_do_io_rw and describe that in the commit message.
- Drop the run-time error on 32bit architectures.
- The #endif splitting the "else if" is not really fortunate, but I'm open to suggestions.

Provide a iowrite64() that does back to back writes for 32 bit?

I was curious what this looked like. If we want to repeat the 4 byte access then I think we need to refactor all the read/write accesses into macros to avoid duplicating code, which results in something like
[1] below. But also once we refactor to macros, the #ifdef within the function as originally proposed gets a lot more bearable too [2].

I'd probably just go with something like [2] unless you want to further macro-ize these branches out of existence in the main function. Feel free to grab any of this you like, the VFIO_IORDWR macro should probably be it's own patch. Thanks,

Alex

[1]

 drivers/vfio/pci/vfio_pci_rdwr.c |  145 +++++++++++++++++--------------
 include/linux/vfio_pci_core.h    |    3 
 2 files changed, 87 insertions(+), 61 deletions(-)
drivers/vfio/pci/vfio_pci_rdwr.c +84 -61
index 03b8f7ada1ac..84bd1d670086 100644
@@ -89,6 +89,71 @@ EXPORT_SYMBOL_GPL(vfio_pci_core_ioread##size);
 VFIO_IOREAD(8)
 VFIO_IOREAD(16)
 VFIO_IOREAD(32)
+#ifdef ioread64
+VFIO_IOREAD(64)
+#endif
+
+#define VFIO_IORDWR(size)						\
+static int vfio_pci_core_iordwr##size(struct vfio_pci_core_device *vdev,\
+				bool iswrite, bool test_mem,		\
+				void __iomem *io, char __user *buf,	\
+				loff_t off, size_t *filled)		\
+{									\
+	u##size val;							\
+	int ret;							\
+									\
+	if (iswrite) {							\
+		if (copy_from_user(&val, buf, sizeof(val)))		\
+			return -EFAULT;					\
+									\
+		ret = vfio_pci_core_iowrite##size(vdev, test_mem,	\
+						  val, io + off);	\
+		if (ret)						\
+			return ret;					\
+	} else {							\
+		ret = vfio_pci_core_ioread##size(vdev, test_mem,	\
+						 &val, io + off);	\
+		if (ret)						\
+			return ret;					\
+									\
+		if (copy_to_user(buf, &val, sizeof(val)))		\
+			return -EFAULT;					\
+	}								\
+									\
+	*filled = sizeof(val);						\
+	return 0;							\
+}									\
+
+VFIO_IORDWR(8)
+VFIO_IORDWR(16)
+VFIO_IORDWR(32)
+#if defined(ioread64) && defined(iowrite64)
+VFIO_IORDWR(64)
+#else
+static int vfio_pci_core_iordwr64(struct vfio_pci_core_device *vdev,
+				  bool iswrite, bool test_mem,
+				  void __iomem *io, char __user *buf,
+				  loff_t off, size_t *filled)
+{
+	int ret;
+	size_t cnt;
+
+	ret = vfio_pci_core_iordwr32(vdev, iswrite, test_mem,
+				     io, buf, off, &cnt);
+	if (ret)
+		return ret;
+
+	ret = vfio_pci_core_iordwr32(vdev, iswrite, test_mem,
+				     io, buf + cnt, off + cnt,
+				     filled);
+	if (ret)
+		return ret;
+
+	*filled += cnt;
+
+	return 0;
+}
+#endif
 
 /*
  * Read or write from an __iomem region (MMIO or I/O port) with an excluded
@@ -114,72 +179,30 @@ ssize_t vfio_pci_core_do_io_rw(struct vfio_pci_core_device *vdev, bool test_mem,
 		else
 			fillable = 0;
 
-		if (fillable >= 4 && !(off % 4)) {
-			u32 val;
-
-			if (iswrite) {
-				if (copy_from_user(&val, buf, 4))
-					return -EFAULT;
-
-				ret = vfio_pci_core_iowrite32(vdev, test_mem,
-							      val, io + off);
-				if (ret)
-					return ret;
-			} else {
-				ret = vfio_pci_core_ioread32(vdev, test_mem,
-							     &val, io + off);
-				if (ret)
-					return ret;
-
-				if (copy_to_user(buf, &val, 4))
-					return -EFAULT;
-			}
+		if (fillable >= 8 && !(off % 8)) {
+			ret = vfio_pci_core_iordwr64(vdev, iswrite, test_mem,
+						     io, buf, off, &filled);
+			if (ret)
+				return ret;
+
+		} else if (fillable >= 4 && !(off % 4)) {
+			ret = vfio_pci_core_iordwr32(vdev, iswrite, test_mem,
+						     io, buf, off, &filled);
+			if (ret)
+				return ret;
 
-			filled = 4;
 		} else if (fillable >= 2 && !(off % 2)) {
-			u16 val;
-
-			if (iswrite) {
-				if (copy_from_user(&val, buf, 2))
-					return -EFAULT;
-
-				ret = vfio_pci_core_iowrite16(vdev, test_mem,
-							      val, io + off);
-				if (ret)
-					return ret;
-			} else {
-				ret = vfio_pci_core_ioread16(vdev, test_mem,
-							     &val, io + off);
-				if (ret)
-					return ret;
-
-				if (copy_to_user(buf, &val, 2))
-					return -EFAULT;
-			}
+			ret = vfio_pci_core_iordwr16(vdev, iswrite, test_mem,
+						     io, buf, off, &filled);
+			if (ret)
+				return ret;
 
-			filled = 2;
 		} else if (fillable) {
-			u8 val;
-
-			if (iswrite) {
-				if (copy_from_user(&val, buf, 1))
-					return -EFAULT;
-
-				ret = vfio_pci_core_iowrite8(vdev, test_mem,
-							     val, io + off);
-				if (ret)
-					return ret;
-			} else {
-				ret = vfio_pci_core_ioread8(vdev, test_mem,
-							    &val, io + off);
-				if (ret)
-					return ret;
-
-				if (copy_to_user(buf, &val, 1))
-					return -EFAULT;
-			}
+			ret = vfio_pci_core_iordwr8(vdev, iswrite, test_mem,
+						    io, buf, off, &filled);
+			if (ret)
+				return ret;
 
-			filled = 1;
 		} else {
 			/* Fill reads with -1, drop writes */
 			filled = min(count, (size_t)(x_end - off));
include/linux/vfio_pci_core.h +3 -0
index a2c8b8bba711..f4cf5fd2350c 100644
@@ -157,5 +157,8 @@ int vfio_pci_core_ioread##size(struct vfio_pci_core_device *vdev,	\
 VFIO_IOREAD_DECLATION(8)
 VFIO_IOREAD_DECLATION(16)
 VFIO_IOREAD_DECLATION(32)
+#ifdef ioread64
+VFIO_IOREAD_DECLATION(64)
+#endif
 
 #endif /* VFIO_PCI_CORE_H */

--------------------------------------------------------------------------------------

[2]

 drivers/vfio/pci/vfio_pci_rdwr.c |  122 +++++++++++++++----------------
 include/linux/vfio_pci_core.h    |    3 
 2 files changed, 65 insertions(+), 60 deletions(-)

commit 05aa3342b4f9e249f3bb370e2fdf66e95de556b8

Author: Alex Williamson <[email protected]>
Date:   Mon Apr 22 15:58:48 2024 -0600

test

drivers/vfio/pci/vfio_pci_rdwr.c +62 -60

140 lines not on this page · raw message

include/linux/vfio_pci_core.h +3 -0
index a2c8b8bba711..f4cf5fd2350c 100644
@@ -157,5 +157,8 @@ int vfio_pci_core_ioread##size(struct vfio_pci_core_device *vdev,	\
 VFIO_IOREAD_DECLATION(8)
 VFIO_IOREAD_DECLATION(16)
 VFIO_IOREAD_DECLATION(32)
+#ifdef ioread64
+VFIO_IOREAD_DECLATION(64)
+#endif
 
 #endif /* VFIO_PCI_CORE_H */
#3 lore raw
Quoted 1 line · On Mon, 22 Apr 2024 16:33:53 -0600 Alex Williamson…

[2]

..

drivers/vfio/pci/vfio_pci_rdwr.c ssize_t vfio_pci_core_do_io_rw(struct vfio_pci_core_device *vdev, bool test_mem, · line 166

Quoted 13 lines
drivers/vfio/pci/vfio_pci_rdwr.c +9 -0
@@ -114,72 +155,33 @@ ssize_t vfio_pci_core_do_io_rw(struct vfio_pci_core_device *vdev, bool test_mem,
 		else
 			fillable = 0;
 
+#if defined(ioread64) && defined(iowrite64)
+		if (fillable >= 8 && !(off % 8)) {
+			ret = vfio_pci_core_iordwr64(vdev, iswrite, test_mem,
+						     io, buf, off, &filled);
+			if (ret)
+				return ret;
+
+		} else
+#endif

This could also just be:

#ifdef vfio_pci_core_iordwr64

at this point. BTW, all this is only compile tested. Thanks,

Alex

author
#4 lore raw
In reply to Alex Williamson
Quoted 53 lines · Alex Williamson
Quoted 41 lines · On Mon, 22 Apr 2024 14:43:05 -0300 Jason Gunthorpe <[email protected]> wrote:
Quoted 32 lines · Gerd Bayer

From: Ben Segal <[email protected]>

Many PCI adapters can benefit or even require full 64bit read and write access to their registers. In order to enable work on user-space drivers for these devices add two new variations vfio_pci_core_io{read|write}64 of the existing access methods when the architecture supports 64-bit ioreads and iowrites.

Since these access methods are instantiated on 64bit architectures,
only, their use in vfio_pci_core_do_io_rw() is restricted by conditional
compiles to these architectures.

---
Hi all,

we've successfully used this patch with a user-mode driver for a PCI
device that requires 64bit register read/writes on s390. A quick grep
showed that there are several other drivers for PCI devices in the kernel
that use readq/writeq and eventually could use this, too. So we decided to propose this for general inclusion.

Thank you,
Gerd Bayer

Quoted 7 lines

Changes v1 -> v2:
- On non 64bit architecture use at most 32bit accesses in vfio_pci_core_do_io_rw and describe that in the commit message.
- Drop the run-time error on 32bit architectures.
- The #endif splitting the "else if" is not really fortunate, but I'm
  open to suggestions.

Quoted 1 line

Provide a iowrite64() that does back to back writes for 32 bit?

I was curious what this looked like.  If we want to repeat the 4 byte access then I think we need to refactor all the read/write accesses into macros to avoid duplicating code, which results in something like [1] below.  But also once we refactor to macros, the #ifdef within the function as originally proposed gets a lot more bearable too [2].

Quoted 6 lines

I'd probably just go with something like [2] unless you want to further macro-ize these branches out of existence in the main function. Feel free to grab any of this you like, the VFIO_IORDWR macro should probably be it's own patch.  Thanks,

Alex

Hi Alex,

thanks for your suggestions, I like your VFIO_IORDWR macro in [1]. As I just explained to Jason, I don't think that the back-to-back 32bit accesses are a safe emulation of 64bit accesses in general, though. So I'd rather leave that out.

However, I'm working on an idea - extending on the VFIO_IORDWR macro - to convert the if - else if - chain into a switch/case construct, where I can more easily #ifdef out the 64bit case if not available.

Now I "just" need to test this ;)

Thanks,
Gerd

#5 lore raw
Quoted 59 lines · Gerd Bayer
Quoted 58 lines · Alex Williamson
Quoted 41 lines · On Mon, 22 Apr 2024 14:43:05 -0300 Jason Gunthorpe <[email protected]> wrote:
Quoted 32 lines · Gerd Bayer

From: Ben Segal <[email protected]>

Many PCI adapters can benefit or even require full 64bit read and write access to their registers. In order to enable work on user-space drivers for these devices add two new variations vfio_pci_core_io{read|write}64 of the existing access methods when the architecture supports 64-bit ioreads and iowrites.

Since these access methods are instantiated on 64bit architectures,
only, their use in vfio_pci_core_do_io_rw() is restricted by conditional
compiles to these architectures.

---
Hi all,

we've successfully used this patch with a user-mode driver for a PCI
device that requires 64bit register read/writes on s390. A quick grep
showed that there are several other drivers for PCI devices in the kernel
that use readq/writeq and eventually could use this, too. So we decided to propose this for general inclusion.

Thank you,
Gerd Bayer

Quoted 7 lines

Changes v1 -> v2:
- On non 64bit architecture use at most 32bit accesses in vfio_pci_core_do_io_rw and describe that in the commit message.
- Drop the run-time error on 32bit architectures.
- The #endif splitting the "else if" is not really fortunate, but I'm
  open to suggestions.

Quoted 1 line

Provide a iowrite64() that does back to back writes for 32 bit?

I was curious what this looked like.  If we want to repeat the 4 byte access then I think we need to refactor all the read/write accesses into macros to avoid duplicating code, which results in something like [1] below.  But also once we refactor to macros, the #ifdef within the function as originally proposed gets a lot more bearable too [2].

I'd probably just go with something like [2] unless you want to further macro-ize these branches out of existence in the main function. Feel free to grab any of this you like, the VFIO_IORDWR macro should probably be it's own patch.  Thanks,

Quoted 8 lines
Quoted 1 line

Alex

Hi Alex,

thanks for your suggestions, I like your VFIO_IORDWR macro in [1]. As I just explained to Jason, I don't think that the back-to-back 32bit accesses are a safe emulation of 64bit accesses in general, though. So I'd rather leave that out.

It is though, it is exactly what the code does now.

This function is not 'do exactly XX byte store' it is actually 'memcpy to io' with some occasional support for making contiguous TLPs in common cases..

Jason

author
#6 lore raw
In reply to Jason Gunthorpe
Quoted 40 lines · Jason Gunthorpe
Quoted 31 lines · Gerd Bayer

From: Ben Segal <[email protected]>

Many PCI adapters can benefit or even require full 64bit read and write access to their registers. In order to enable work on user-space drivers for these devices add two new variations vfio_pci_core_io{read|write}64 of the existing access methods when the architecture supports 64-bit ioreads and iowrites.

Since these access methods are instantiated on 64bit architectures, only, their use in vfio_pci_core_do_io_rw() is restricted by conditional
compiles to these architectures.

---
Hi all,

we've successfully used this patch with a user-mode driver for a PCI
device that requires 64bit register read/writes on s390. A quick grep
showed that there are several other drivers for PCI devices in the kernel
that use readq/writeq and eventually could use this, too. So we decided to propose this for general inclusion.

Thank you,
Gerd Bayer

Quoted 7 lines

Changes v1 -> v2:
- On non 64bit architecture use at most 32bit accesses in vfio_pci_core_do_io_rw and describe that in the commit message.
- Drop the run-time error on 32bit architectures.
- The #endif splitting the "else if" is not really fortunate, but I'm
  open to suggestions.

Quoted 1 line

Provide a iowrite64() that does back to back writes for 32 bit?

Hi Jason,

unfortunately, the nomenclature in vfio_pci_rdwr.c is not very clear... vfio_io{read|write}64 are mapped to io{read|write}64 as defined in include/asm-generic/io.h prior to my change already. OTOH, looks like vfio_io{read|write}64 are consumed only by the vfio_pci_core_io{read|write}64 functions. This however is an exported symbol - that seems to be used only as vfio_pci_core_io{read|write}16, so far.

vfio_pci_core_io{read|write}X is also used by vfio_pci_core_do_io_rw() which does "bulk" reads/writes using the largest suitable access size. I think there, we can live without 64bit accesses as the while loop there will use 32bit read/writes back-to-back as applicable.

So I think 64bit accesses on 32bit architectures through VFIO are somewhat uncharted territory - and I'm not sure that back-to-back 32bit accesses are the right thing to do. If the device defined 64bit registers, you could trigger side-effects in the wrong order (or not at all).

Somewhat overwhelmed,
Gerd