Re: swapctl: SWAP_STATS different to SWAP_NSWAP (2 != 3)

[email protected] (Thomas Klausner) writes:

>On Thu, Feb 12, 2026 at 06:14:41PM +0100, Matthew Green wrote:
>> > Hi!
>> >
>> > While trying to take a closer look at https://gnats.netbsd.org/56764
>> > I'm keeping 'swapctl -l; sleep 1' running in a loop next to a bulk
>> > build.
>> 
>> can you see if -current no longer has this problem?  i think i saw
>> what was wrong that would trigger this problem and commited a fix.

>I think that might have made it worse:

>#11 uvm_fault_internal (orig_map=orig_map@entry=0xffff9cd8c58d5ac0, vaddr=vaddr@entry=126083539185664, access_type=access_type@entry=2, fault_flag=fault_flag@entry=0) at /usr/src/sys/uvm/uvm_fault.c:946
>#12 0xffffffff8023c5d4 in trap (frame=0xffffa7a493785870) at /usr/src/sys/arch/amd64/amd64/trap.c:523
>#13 0xffffffff80234ad4 in alltraps ()
>#14 0xffffffff80234353 in copyout ()
>#15 0xffffffff80ddf279 in uvm_swap_stats (ptr=0x72ac1c90b000 <error: Cannot access memory at address 0x72ac1c90b000>, misc=2, f=f@entry=0x0, len=len@entry=1056, retval=retval@entry=0xffffa7a493785fb0)
>    at /usr/src/sys/uvm/uvm_swap.c:818


Holding a lock while running copyout that page-faults can be a problem.

Since the list is only rotated, it's sufficient to collect pointers
to the entries into a kernel buffer. After releasing the data
lock, the entries can then be copied out.