IRC channel logs

<youpi>diegonc: I sent a heads-up to the debian-hurd mailing list some time ago

<damo22>yeah im getting 5.8MB/s over ssh lan to the hurd box with rumpnet

<damo22>but when i hit the internet i get < 100kB/s

<damo22>there is a tunable in rumpnet

<damo22>#define BPF_TIMEOUT_MS 10

<damo22>BIOCSRTIMEOUT

<damo22>which is the timeout for when the buffer should be flushed in bpf even if there is still room in the buffer for more packets

<damo22>maybe 0.5MB is too big for the buffer?

<damo22>maybe sending is fast but receiving is slow

<damo22>yup

<damo22>youpi: i am using blocking mode with a timeout in rumpnet for rcv, should i be using nonblocking with select?

<damo22>heh i think i fixed it

<damo22>received at 5MB/s

<damo22>Pellescours: rumpnet is fixed

<damo22>i just need to make it support more than one nic

<damo22>done

<damo22>receiving a large file hangs in smp

<damo22>but starts at 5MB/s

<damo22>otherwise it works fine in UP

<damo22>receiving a large file on full smp hangs gnumach

<damo22>i have a dynamic rumpnet driver running fine in UP

<damo22>youpi: what is on IRQ 1?

<damo22>keybd?

<damo22>ICR is 0xfa logical edge assert no-shorthand

<damo22>and IRR has 48 49 59

<damo22>2 cores spinning hard

<damo22>gnumach unresponsive

<damo22>it keeps using IPI CALL_AST_CHECK

<youpi>damo22: (i am using blocking mode with a timeout in rumpnet for rcv, should i be using nonblocking with select?): I don't know how your rumpnet is set up in general, so I have no idea. In principle there is no reason why non-blocking should be better, if you have two threads, one for receiving packets, and one for sending packets

<youpi>irq 1 is the keyboard

<damo22>blocking mode works fine with a thread

<damo22>but full smp + rcv rumpnet hangs gnumach

<damo22>with a large file

<damo22>youpi: intel SDM volume 3 has incomplete info on sending IPI with apic, it just says when ICR low is written to, it sends an IPI

<damo22>so i dont know if we need to assert edge then deassert edge or not

<damo22>linux just waits for pending then asserts edge

<damo22>we are asserting first, then waiting, then deasserting, then waiting again

<damo22>youpi: should i mail in patch for rumpnet driver that works in UP?

<damo22># renice -n 0 -p `pidof rumpnet`

<damo22>that fixed it on smp

<damo22>well not completely

<Pellescours>htop (3.4.0+) compilation fails on hurd due to header <linux/capability.h> not present (for obvious reasons)

IRC channel logs

2025-07-06.log