Re: [PATCH 1/2] vhost: set dirty log bits without pinning pages on x86

From: Michael S. Tsirkin

Date: Thu Nov 12 2026 - 05:20:09 EST


On Thu, Nov 12, 2026 at 04:56:30PM +0800, zhaoguohan@xxxxxxxxxx wrote:
> From: Khan Zhao <zhaoguohan@xxxxxxxxxx>
>
> set_bit_to_user() pins the log page with pin_user_pages_fast(), maps it
> with kmap_atomic(), sets one bit and then unpins the page and marks it
> dirty. This runs for every bit set in the dirty log, e.g. on each used
> ring update while logging is enabled for live migration.
>
> log_write() always passes a byte address and a bit number in [0, 7], so
> the update can be done as a byte-sized cmpxchg on the user address. On
> x86, try that first using unsafe_try_cmpxchg_user() inside a
> user_access_begin()/user_access_end() section with page faults disabled.
> If the access faults (page not present, read-only or needing COW), or
> the cmpxchg loses to concurrent updates of the same byte four times in
> a row, fall back to the existing pin/kmap/set_bit path, which can fault
> the page in.
>
> unsafe_try_cmpxchg_user() is only implemented on x86, so other
> architectures keep using the existing path.
>
> On an x86_64 KVM guest (i7-12800HX host), a synthetic used-ring
> workload with logging enabled went from 243.60 to 171.11 ns per
> descriptor, averaged over five alternating runs. This is the
> end-to-end cost per descriptor, not live migration performance.
>
> Signed-off-by: Khan Zhao <zhaoguohan@xxxxxxxxxx>
> ---
> drivers/vhost/vhost.c | 50 ++++++++++++++++++++++++++++++++++++++-----
> 1 file changed, 45 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index e29f154706da..9c3045504709 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -2527,11 +2527,7 @@ long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
> }
> EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
>
> -/* TODO: This is really inefficient. We need something like get_user()
> - * (instruction directly accesses the data, with an exception table entry
> - * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
> - */
> -static int set_bit_to_user(int nr, void __user *addr)
> +static int set_bit_to_user_slow(int nr, void __user *addr)
> {
> unsigned long log = (unsigned long)addr;
> struct page *page;
> @@ -2550,6 +2546,50 @@ static int set_bit_to_user(int nr, void __user *addr)
> return 0;
> }
>
> +#ifdef CONFIG_X86

Wait this is x86 specific anyway? So let's just implement it
properly, why do we need all these hacks?

I run a quick test on this in a VM, feel free to reuse:

Signed-off-by: Michael S. Tsirkin <mst@xxxxxxxxxx>



diff --git a/arch/x86/include/asm/uaccess.h b/arch/x86/include/asm/uaccess.h
index 3a0dd3c2b233e..3ed49c91a7434 100644
--- a/arch/x86/include/asm/uaccess.h
+++ b/arch/x86/include/asm/uaccess.h
@@ -594,6 +594,23 @@ _label: \
__ret; \
})

+#define set_bit_user(nr, addr) \
+({ \
+ int __err = 0; \
+ __uaccess_begin(); \
+ asm volatile("\n" \
+ "1: " LOCK_PREFIX "btsb %[bit], %[ptr]\n" \
+ "2:\n" \
+ _ASM_EXTABLE_TYPE_REG(1b, 2b, EX_TYPE_EFAULT_REG, \
+ %[errout]) \
+ : [errout] "=r" (__err), \
+ [ptr] "+m" (*(u8 __user *)(addr)) \
+ : [bit] "r" ((unsigned int)(nr)) \
+ : "memory", "cc"); \
+ __uaccess_end(); \
+ __err ? -EFAULT : 0; \
+})
+
/*
* We want the unsafe accessors to always be inlined and use
* the error labels - thus the macro games.