Re: [PATCH] iommu: riscv: Split 8-byte accesses on 32 bit I/O bus platform

From: Robin Murphy
Date: Wed Apr 02 2025 - 07:29:08 EST


On 2025-03-25 2:42 pm, Xu Lu wrote:
Introduce a new configuration CONFIG_RISCV_IOMMU_32BIT to enable
splitting 8-byte access into 4-byte transactions for hardware platform
whose I/O bus limits access to 4-byte transfers.

Signed-off-by: Xu Lu <luxu.kernel@xxxxxxxxxxxxx>
---
drivers/iommu/riscv/Kconfig | 9 +++++++++
drivers/iommu/riscv/iommu.h | 28 +++++++++++++++++++++++-----
2 files changed, 32 insertions(+), 5 deletions(-)

diff --git a/drivers/iommu/riscv/Kconfig b/drivers/iommu/riscv/Kconfig
index c071816f59a6..b7c9ea22d969 100644
--- a/drivers/iommu/riscv/Kconfig
+++ b/drivers/iommu/riscv/Kconfig
@@ -18,3 +18,12 @@ config RISCV_IOMMU_PCI
def_bool y if RISCV_IOMMU && PCI_MSI
help
Support for the PCIe implementation of RISC-V IOMMU architecture.
+
+config RISCV_IOMMU_32BIT
+ bool "Support 4-Byte Accesses on RISC-V IOMMU Registers"
+ depends on RISCV_IOMMU
+ default n
+ help
+ Support hardware platform whose I/O bus limits access to 4-byte
+ transfers. When enabled, all accesses to IOMMU registers will be
+ split into 4-byte accesses.
diff --git a/drivers/iommu/riscv/iommu.h b/drivers/iommu/riscv/iommu.h
index 46df79dd5495..0e3552a8142d 100644
--- a/drivers/iommu/riscv/iommu.h
+++ b/drivers/iommu/riscv/iommu.h
@@ -14,6 +14,10 @@
#include <linux/iommu.h>
#include <linux/types.h>
#include <linux/iopoll.h>
+#ifdef CONFIG_RISCV_IOMMU_32BIT
+#include <linux/io-64-nonatomic-hi-lo.h>
+#include <linux/io-64-nonatomic-lo-hi.h>
+#endif
#include "iommu-bits.h"
@@ -69,21 +73,35 @@ void riscv_iommu_disable(struct riscv_iommu_device *iommu);
#define riscv_iommu_readl(iommu, addr) \
readl_relaxed((iommu)->reg + (addr))
-#define riscv_iommu_readq(iommu, addr) \
- readq_relaxed((iommu)->reg + (addr))
-
#define riscv_iommu_writel(iommu, addr, val) \
writel_relaxed((val), (iommu)->reg + (addr))
+#define riscv_iommu_readl_timeout(iommu, addr, val, cond, delay_us, timeout_us) \
+ readx_poll_timeout(readl_relaxed, (iommu)->reg + (addr), val, cond, \
+ delay_us, timeout_us)
+
+#ifndef CONFIG_RISCV_IOMMU_32BIT
+#define riscv_iommu_readq(iommu, addr) \
+ readq_relaxed((iommu)->reg + (addr))
+
#define riscv_iommu_writeq(iommu, addr, val) \
writeq_relaxed((val), (iommu)->reg + (addr))
#define riscv_iommu_readq_timeout(iommu, addr, val, cond, delay_us, timeout_us) \
readx_poll_timeout(readq_relaxed, (iommu)->reg + (addr), val, cond, \
delay_us, timeout_us)
+#else /* CONFIG_RISCV_IOMMU_32BIT */
+#define riscv_iommu_readq(iommu, addr) \
+ hi_lo_readq_relaxed((iommu)->reg + (addr))
-#define riscv_iommu_readl_timeout(iommu, addr, val, cond, delay_us, timeout_us) \
- readx_poll_timeout(readl_relaxed, (iommu)->reg + (addr), val, cond, \
+#define riscv_iommu_writeq(iommu, addr, val) \
+ ((addr == RISCV_IOMMU_REG_IOHPMCYCLES) ? \
+ lo_hi_writeq_relaxed((val), (iommu)->reg + (addr)) : \
+ hi_lo_writeq_relaxed((val), (iommu)->reg + (addr)))

Echoing Jason's comment, what is this even trying to achieve? Nothing in the spec suggests that the cycle counter register is functionally different from the other PMU counter registers (other than its self-contained overflow bit).

It is not, in general, safe to do a split write to a running counter either way - low-high vs. high-low just moves the problem around, changing *which* combinations of values are problematic and capable of overflowing into each other between the writes. If the PMU driver can't write counters atomically, it will need to ensure that it only ever write them while stopped (at which point the order surely shouldn't matter). Conversely, though, reading from running counters is a bit more reasonable, but it needs more than just hi_lo_readq to guarantee it's not got a torn result.

Thanks,
Robin.

+
+#define riscv_iommu_readq_timeout(iommu, addr, val, cond, delay_us, timeout_us) \
+ readx_poll_timeout(hi_lo_readq_relaxed, (iommu)->reg + (addr), val, cond, \
delay_us, timeout_us)
+#endif /* CONFIG_RISCV_IOMMU_32BIT */
#endif