Thanks for revising.

I've completely forgotten this changeset.

• Management to Interface

• sysctl -l to display the lenth of returned data.

• Bump the date, use "struct" as referenced in other parts of man sysctl, and give few examples of struct types.

In D34012#772496, @debdrup wrote:

Have you run mandoc -Tlint and igor on the manual page?

At the very least, the latter would hint that you need to bump .Dd :)

• Bump the date, use "struct" as referenced in other parts of man sysctl, and give few examples of struct types.

• Updated man page.

• systat: Added -F option to show "fmt"

• Update error description of show_info()

• Return error code from show_info()

I'm wondering if this change can be merged.
I have this change in local tree and haven't exposed to an issue since review was requested.

In D33495#760973, @mckusick wrote:

This looks like a useful change.

Do you want me to commit it?

I tested the new patch with 13.0-RELEASE and confirmed crash isn't happening.

I patched to 13.0-RELEASE.
I rebuilt and installed the new kernel to the specific machine.
I removed hint.agp.1.disabled="1" from /boot/loader.conf.
I verified the new kernel with the patch booted okay after removing the config.

I will test and report back.

There was a bug which resulted not calling new vm.objects_swap.
After fixing the bug and further testing, I no longer see long pause.
In other words, the kernel patch and new sysctl is helping performance improvement.

While you say that D31163 did not improve the performance issue with tmpfs files, did it improve performance on other tests? Or did it fail to improve performance on any of your tests?

• systat: Stop displaying total if there is one swap devise.
• systat: Implemented per-process swap display on -swap.
• Avoid double counting and address review comments.
• systat.1 updates
• D31163 - Add sysctl vm.objects_swap.
• Use new vm.objects_swap instead of vm.objects.

I started looking into https://reviews.freebsd.org/D31163.
I've been releng/13.0 based system (with cherrypick).
This time, the patch doesn't apply to 13 and I need to prepare main branch based environment.
I will need a bit time to build the latest.

I tested the 2 kernel changes both applied. In short, it helps since the 2nd refresh.

Update systat.1 based on suggestions.

In D29754#680269, @kib wrote:

Ok. Could you add the following to the kernel, please, and see if it makes any difference?

diff --git a/sys/vm/vm_object.c b/sys/vm/vm_object.c
index dbb09f67b4c1..cf8b3fcb8f9c 100644
--- a/sys/vm/vm_object.c
+++ b/sys/vm/vm_object.c
@@ -2478,7 +2478,7 @@ sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
 	struct vnode *vp;
 	struct vattr va;
 	vm_object_t obj;
-	vm_page_t m;
+//	vm_page_t m;
 	u_long sp;
 	int count, error;
 
@@ -2524,6 +2524,7 @@ sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
 		kvo->kvo_memattr = obj->memattr;
 		kvo->kvo_active = 0;
 		kvo->kvo_inactive = 0;
+#if 0

In D29754#677726, @kib wrote:

I mean, when asking whether my one-line patch helped, did interactivity of the system during vm.objects retrieval improved? Of course it cannot optimize the running time of the sysctl, but it should allow for other threads to run some more while sysctl is gathering data and copying it out.

Fix once more.

We have a duplicate review for this and are tracking changes in https://reviews.freebsd.org/D29754.
The comments are addressed in https://reviews.freebsd.org/D29754.

Adjust systat.1 from the other review.

Most likely it was sysctl vm.objects and not per-proc vmmap sysctl that caused the intermediate hang and CPU usage. As I and Mark discussed it earlier, some more fine-grained API to retrieve e.g. just the shadow chain for given object handle probably would help there.

But before I implement this sysctl, can you recheck, please, that the following patch solves the transient hang issue?

 commit fd2da9e1515bb50fb1c8fb747d9987ee9b8a314d
Author: Konstantin Belousov <kib@FreeBSD.org>
Date:   Sat May 8 01:13:29 2021 +0300

    sysctl vm.objects: yield if hog

diff --git a/sys/vm/vm_object.c b/sys/vm/vm_object.c
index 1c4e879d82ea..b1fa3ebdf9ab 100644
--- a/sys/vm/vm_object.c
+++ b/sys/vm/vm_object.c
@@ -2614,6 +2614,7 @@ sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
 		    sizeof(uint64_t));
 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
+		maybe_yield();
 		mtx_lock(&vm_object_list_mtx);
 		if (error)
 			break;

Thank you for reviews.

Remove extra comments.

another style fix.

Remove #include <devstat.h>

Style fixes based on review comments and recstored Kirk's systat.1 updates.

Thanks for the program. Indeed, I saw lots of duplicates from the java program.

• Avoid double counting and address review comments.

The algorithm has been adjusted proposed approach as below by Konstantin:

Kirk and I have been looking into this together.
Given the history of comments, this is a better one for tracking objective of this change.

I had mixed with another changeset and caused failures.

systat: Update to SPDX BSD-2-Clause - I've written zarc.c and sysput.c and University of California hasn't been involved

In D29754#667588, @kib wrote:

I saw this patch on a list, and I decided to avoid it.

It does not do what is advertised, i.e. the numbers displayed has nothing related to the 'process swap usage'. First, the swap usage per process is quite hard to define: due to the shadow chaining and sharing of the shadowing objects between siblings, a lot of swap objects is commonly used but simply not visible to the view from the process map level. Second, (and to be fair, it is mentioned in the man page update of this patch), swap object type itself does not imply that a swap is used, at all. Third, even if the top-level object is default, below it, the objects might have swap types.

Fix when the number of swap devise is 1.

Fix off-by-one error over 60% and stop extra +1 for percentage.

Fix style.

Add a , to .Xr ls

Update getbsize man page.

Remove extra spaces around a function argument.

Use bitmask to indicate which protocol versions to try.
Also remove duplicate check.

I make multiple read-only connection for installworld/installkernel from different machines.
Otherwise, I use ggate to export some ufs file systems.
I could use nfs for these as well.

Address some of problems pointed by code review.

The slowness issue existed back in a decade ago around 6.1-RELEASE and some years/releases thereafter.

Thank you for checking this again.
It was also my bad that the problem wasn't described well enough.

Stop using PR* formatter.
Tested on i386 and amd64.

ggate defines its offset in unsigned int 64 as below.
So, printing negative numbers from printf indicates using a wrong formatter.

Thank you for your support, Mark.
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=132845 is the corresponding bugzilla entry.

Verify paths are the same between handshakes for a single session.

Moved close() to connection_remove() and aligned with send/recv fds.

First of all, thank you for your quick response.

Added static to implementatoin; follow style of other existing static functions.

Drop unnecessary NULL check. This is a field on a struct and won't be NULL.