Re: [Patch] Optimize dropping of relation buffers using dlist

On Thu, Oct 22, 2020 at 8:32 PM tsunakawa(dot)takay(at)fujitsu(dot)com
<tsunakawa(dot)takay(at)fujitsu(dot)com> wrote:
> I'm probably being silly, but can't we avoid the problem by using fstat() instead of lseek(SEEK_END)? Would they return the same value from the i-node?

Amazingly, st_size can disagree with SEEK_END when using the Linux NFS
client, but its behaviour is worse. Here's a sequence from a Linux
NFS client talking to a Linux NFS server with no free space. This
time, I also replaced the fsync() with sleep(60), just to make it
clear that SEEK_END offset can move at any time due to asynchronous
activity in kernel threads:

lseek(..., SEEK_END) = 9670656
fstat(...) = 0, st_size = 9670656

write(...) = 8192
lseek(..., SEEK_END) = 9678848
fstat(...) = 0, st_size = 9670656 (*1)

sleep(...) = 0

lseek(..., SEEK_END) = 9670656 (*2)
fstat(...) = 0, st_size = 9670656

fsync(...) = -1
lseek(..., SEEK_END) = 9670656
fstat(...) = 0, st_size = 9670656
fsync(...) = 0

However, I'm not entirely sure which phenomena visible here to blame
on which subsystems, and therefore which things to expect on local
filesystems, or on other operating systems. I can say that with a
FreeBSD NFS client and the same Linux NFS server, I don't see
phenomenon *1 (unsurprising) but I do see phenomenon *2 (surprising to
me).

> Or, can't we just try to do BufTableLookup() one block after what smgrnblocks() returns?

Unfortunately the problem isn't limited to one block.