Skip to contentSkip to conversation

2.4.17rc1aa1

DISCUSSION Andrea Arcangeli 3 participants 4 messages Dec 14, 2001 – Dec 17, 2001

lore
Share & tools
lore
#0 lore raw

Marcelo, can you merge the loop-deadlock fix? (the others aren't trivially mergeable yet and for the tcp conntrack you'd need to ask Rusty first, the change is simple enough that I merged it for now)

Andrew, could you give it a spin to verify I/O performance is back? It is fast for me now.

URL:

	ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.17rc1aa1.bz2
	ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.17rc1aa1/

Only in 2.4.17rc1aa1: 00_block-highmem-all-18b-2.bz2 Only in 2.4.17pre4aa1: 00_block-highmem-all-18b.bz2 Only in 2.4.17pre4aa1: 00_flush_icache_range-1 Only in 2.4.17rc1aa1: 00_flush_icache_range-2 Only in 2.4.17pre4aa1: 00_silent-stack-overflow-12 Only in 2.4.17rc1aa1: 00_silent-stack-overflow-13 Only in 2.4.17pre4aa1: 00_spinlock-cacheline-2 Only in 2.4.17rc1aa1: 00_spinlock-cacheline-3 Only in 2.4.17pre4aa1: 10_numa-sched-14
Only in 2.4.17rc1aa1: 10_numa-sched-15
Only in 2.4.17pre4aa1: 60_atomic-alloc-6 Only in 2.4.17rc1aa1: 60_atomic-alloc-7

Rediffed.

Only in 2.4.17pre4aa1: 00_rwsem-fair-25-recursive-6 Only in 2.4.17rc1aa1: 00_rwsem-fair-25-recursive-7

Not used actively anymore for the coredump (we walk the pagetables by hand instead of using the MMU now).

Only in 2.4.17rc1aa1: 00_loop-deadlock-1

Fixed loop deadlock (balance_dirty() must not be called within the request_fn or it can deadlock or recurse way too much). It is enough to recall balance_dirty() at the highest layer so this won't introduce any stability problem.

Only in 2.4.17rc1aa1: 00_tcp-conntrack-fin-1

After upgrading my firewall/NAT to 2.4 (s/ipchains/iptables/ :) I started getting valid tcp packets dropped/logged (I'm allowing only RELATED and ESTABLISHED packets from the outside, with ipchains I was allowing all tcp packets that weren't syn and there weren't problems). Rusty sent me the above fix to see if my problem with the tcp connection tracking goes away (without adding again the rule that all non tcp syn can came in).

Only in 2.4.17pre4aa1: 10_vm-19
Only in 2.4.17rc1aa1: 10_vm-20

Should fix the I/O async flushing performance drop reported by Andrew. Cannot += BUF_LOCKED any longer in balance_dirty because we don't waste time any longer refiling clean buffers into the BUF_CLEAN now (sync has to pass all over the bh anyways so it doesn't matter for it either). Avoid collapsing the max_mapped during shrink_cache internal passes.

Only in 2.4.17pre4aa1: 50_uml-patch-2.4.15-3.bz2
Only in 2.4.17rc1aa1: 50_uml-patch-2.4.16-2.bz2

Latest update from Jeff at user-mode-linux.sourceforge.net.

Only in 2.4.17pre4aa1: 60_tux-2.4.16-final-C9.bz2
Only in 2.4.17rc1aa1: 60_tux-2.4.16-final-D5.bz2

Latest update from Ingo at www.redhat.com/~mingo/.

Andrea

#1 lore raw

Another loop bug? (Was: 2.4.17rc1aa1)

Quoted 7 lines

Only in 2.4.17rc1aa1: 00_loop-deadlock-1

Fixed loop deadlock (balance_dirty() must not be called within the request_fn or it can deadlock or recurse way too much). It is enough to recall balance_dirty() at the highest layer so this won't introduce any stability problem.

Maybe my problem is unrelated to this, but when I try to mount an ext2 image (from an ext3 filesystem, if that counts), loop0 starts to eat CPU, and brings up the load to around 2.90. After a short time (about a minute after mounting the image), everything that tries to access the mounted area hangs. I can't umount it, and can't kill the processes which are trying to access it.

I have this problem since 2.4.16, 2.4.14 (with ext3 0.9.15) works fine.

#2 lore raw
In reply to Andrea Arcangeli
Quoted 4 lines · Andrea Arcangeli

Marcelo, can you merge the loop-deadlock fix? (the others aren't trivially mergeable yet and for the tcp conntrack you'd need to ask Rusty first, the change is simple enough that I merged it for now)

I verified on two of my systems that the loop-deadlock fix patch works.

-Barry K. Nathan <[email protected]>

#3 lore raw
In reply to Andrea Arcangeli
Quoted 7 lines · Barry K. Nathan
Quoted 4 lines · Andrea Arcangeli

Marcelo, can you merge the loop-deadlock fix? (the others aren't trivially mergeable yet and for the tcp conntrack you'd need to ask Rusty first, the change is simple enough that I merged it for now)

I verified on two of my systems that the loop-deadlock fix patch works.

good, thanks.

Andrea