| 14 Mar 2026 |
 Mic92 | Okay but in my testing I had issues only with larger images | 19:44:09 |
 matthewcroughan | you don't need a particularly large image to make it happen | 19:44:10 |
| Mic92 | In any case, I know a project where this was triggered. | 19:45:14 |
| Mic92 | @lassulus:lassul.us: the dirty page sysctl doesn't sound like ads related to me | 19:46:39 |
| Mic92 | * @lassulus:lassul.us: the dirty page sysctl doesn't sound like zfs related to me | 19:46:56 |
 lassulus | is it only with zfs? | 19:48:13 |
| lassulus | I have a vague memory that it also happens on other FS | 19:48:25 |
| Mic92 | But not 100 percent sure. Would have to check the Kernel code | 19:49:02 |
| Mic92 | Also this should probably take into account how much memory we give to the vm | 19:49:58 |
| lassulus | probably, but we don't need any cache for paths we copy once from one fs to the other? | 19:50:21 |
| Mic92 | Well it might also hurt Io scheduling | 19:50:57 |
| Mic92 | If you flush more often | 19:51:09 |
| Mic92 | CP/rsync are probably doing direct I/o as well | 19:52:10 |
| Mic92 | This usually by passes page cache | 19:52:32 |
| lassulus | cp doesn't seem to do it, otherwise this would not happen? | 19:52:50 |
| lassulus | or it's a different issue | 19:53:08 |
| Mic92 | Has been to long since I did that. I with it did something like read into mmap? This would actually not be direct Io but don't remember the mmap flags it used | 19:54:21 |
| Mic92 | * Has been to long since I read that. I with it did something like read into mmap? This would actually not be direct Io but don't remember the mmap flags it used | 19:54:43 |
| Mic92 | * Has been to long since I read that. I think it did something like read into mmap? This would actually not be direct Io but don't remember the mmap flags it used | 19:55:19 |
| Mic92 | Some of them tell the Kernel to not bother with page cache | 19:56:06 |
| matthewcroughan | I think it is somehow dependent on the host | 20:04:40 |
| matthewcroughan | I have a reproducer now that only fails sometimes | 20:04:47 |
| matthewcroughan | Either that or it is dependent on some randomness | 20:05:42 |
| matthewcroughan | does xargs sort things on the way in? | 20:05:54 |
| lassulus | good question, maybe not | 20:06:20 |
| matthewcroughan | even if it was that, and you sorted it, that only means for the same closure you would get the same bug, rather than it failing randomly | 20:06:45 |
| matthewcroughan | but the root cause would still be there | 20:06:51 |
| lassulus | maybe we just create like 8x1gb files and copy that into the new image? for a reproducer | 20:07:02 |
| Mic92 | In reply to @lassulus:lassul.us
good question, maybe not Should be stable because closure info stays the same | 20:19:38 |
| Mic92 | The parallel flag might introduce some non-determinism | 20:20:07 |
