Skip to contentSkip to conversation

CONFIG_PACKET_MMAP revisited

DISCUSSION [email protected] 5 participants 6 messages Oct 29, 2003 – Nov 6, 2003

lore
Share & tools
lore
#0 lore raw

I've been looking into faster ways to do packet captures and I stumbled on the following discussion on the Linux Kernel mailing list:

http://www.ussg.iu.edu/hypermail/linux/kernel/0202.2/1173.html

In that discussion Jamie Lokier suggested having a memory buffer that's shared between user and kernel space and having the NIC do DMA transfers directly to that buffer as an alternative to using Alexy's shared ring buffer stuff. The argument was that this would avoid the memory copy that the kernel does from the DMA buffer to the memory mapped ring buffer. However, Alan Cox pointed out that the main cost of the memory copy is getting the data from system memory (where the NIC put it via DMA) into the L1 cache (DMA doesn't do any cache coherence so it can't go there directly). The memory copy (presumably from L1 cache to L1 cache) is insignificant compared to this cost and since you'll need to get the data into L1 cache to use it anyway, the memory copy is virtually free.

I'm wondering if this takes all of the costs into account. If I understand how this works the user space application can't get at the packet without a context switch so that the kernel can first copy the packet to the shared buffer. The cost of the context switch is pretty high and this seems to me to be the main bottleneck. I believe that in normal operation each packet (or with NICs that do interrupt coalescing, every n packets) causes an interrupt which causes a context switch, the kernel then copies the data from the DMA buffer to the shared buffer and does a RETI. That's fairly expensive. If, on the other hand, data could be copied directly to user-space accessible memory the NIC wouldn't need to generate any interrupts and the kernel wouldn't need to get involved at all (this assumes a NIC that can be configured not to generate any interrupts). The user space application could then poll the shared buffer and process packets as fast as possible (some synchronization mechanism is clearly needed here, but I think some of the high-end programmable NICs could do this). Would this not be significantly more efficient then the current implementation?

Thank,
Oliver

PS: I'm not a mailing list subscriber so CCs on responses would be appreciated.

--------------------------------------------------------------------
mail2web - Check your email from the web at
http://mail2web.com/ .

#1 lore raw
Quoted 7 lines · [email protected] wrote:

Alan Cox pointed out that the main cost of the memory copy is getting the data from system memory (where the NIC put it via DMA) into the L1 cache (DMA doesn't do any cache coherence so it can't go there directly). The memory copy (presumably from L1 cache to L1 cache) is insignificant compared to this cost and since you'll need to get the data into L1 cache to use it anyway, the memory copy is virtually free.

After the data is copied, it's likely to be written from L1 to L2 (at least) before that position in the ring buffer is rewritten, so that's another overhead.

-- Jamie

#2 lore raw
In reply to [email protected]
Quoted 5 lines · On Wed, 2003-10-29 at 05:09, [email protected] wrote:

I believe that in normal operation each packet
(or with NICs that do interrupt coalescing, every n packets) causes an interrupt which causes a context switch, the kernel then copies the data from the DMA buffer to the shared buffer and does a RETI. That's fairly expensive.

The cost of handling that interrupt and doing an iret is unavoidable (ignoring NAPI). The main point you are missing with the ring buffer is that if packets come in at a fast enough rate, the usermode task never context switches, because there is always data in the ring buffer, so it loops in usermode forever.

The problem could be the packets are coming in just too slow to allow the ring buffer to work properly and causing the application to sleep on poll(2) every time. This would kill performance at pathelogical packet rates I guess.

You could work around this by spinning for a few thousand spins before calling poll(2) (or even indefinately for that matter, and allow the kernel to preempt you if need be).

Signature · 3 lines
// Gianni Tedesco (gianni at scaramanga dot co dot uk)
lynx --source www.scaramanga.co.uk/gianni-at-ecsc.asc | gpg --import
8646BE7D: 6D9F 2287 870E A2C9 8F60 3A3C 91B5 7669 8646 BE7D
#3 lore raw
Quoted 7 lines · Gianni Tedesco
Quoted 5 lines · On Wed, 2003-10-29 at 05:09, [email protected] wrote:

I believe that in normal operation each packet
(or with NICs that do interrupt coalescing, every n packets) causes an interrupt which causes a context switch, the kernel then copies the data from the DMA buffer to the shared buffer and does a RETI. That's fairly expensive.

Quoted 5 lines

The cost of handling that interrupt and doing an iret is unavoidable (ignoring NAPI). The main point you are missing with the ring buffer is that if packets come in at a fast enough rate, the usermode task never context switches, because there is always data in the ring buffer, so it loops in usermode forever.

It seems to me that it can't loop in user mode forever. There is no way to get data into user mode (the ring buffer) witout going through the kernel. My understanding is that the NIC doesn't transfer directly to the user mode ring buffer, but rather to a different DMA buffer. The kernel must copy it from the DMA buffer to the ring buffer. Thus once the user mode app has processed all the data in the ring buffer the kenel _must_ get involved to get more data to user space. Currently the data gets there because the NIC produces an interrupt for each packet (or for every few packets) and when the kernel handles these the data is copied to user space. Then, as you point out, the cost of the RETI can't be avoided.

NAPI tries to solve this problem. I don't know much about NAPI, but as I understand it, the idea is this: The cost of the RETI's and context switches (which occur on each interrupt) can be reduced if the NIC doesn't produce an interrupt for every packet but instead does interrupt coalescing, but this only goes so far. If too many packets are coalesced the data copied by the kernel will no longer fit in the L1 cache and we'll pay the price of moving it there twice (once when the kernel copies the data from main memory to the ring buffer and once when the user mode application reads it out of the ring), the latency may become a problem, we've still got a context switch every time the user mode application has processed everything in the ring buffer (and perhaps more often), and we're still paying the price of copying data from the DMA buffer to the ring.

However, if the NIC could transfer the data directly to user space it wouldn't need to cause an interrupt and the cost of the RETI and the context switch is avoided. The user mode app really could process forever without sleeping at that point.

Quoted 8 lines

The problem could be the packets are coming in just too slow to allow the ring buffer to work properly and causing the application to sleep on poll(2) every time. This would kill performance at pathelogical packet rates I guess.

You could work around this by spinning for a few thousand spins before calling poll(2) (or even indefinately for that matter, and allow the kernel to preempt you if need be).

#4 lore raw
Quoted 10 lines · Oliver Dain

It seems to me that it can't loop in user mode forever. There is no way to get data into user mode (the ring buffer) witout going through the kernel. My understanding is that the NIC doesn't transfer directly to the user mode ring buffer, but rather to a different DMA buffer. The kernel must copy it from the DMA buffer to the ring buffer. Thus once the user mode app has processed all the data in the ring buffer the kenel _must_ get involved to get more data to user space. Currently the data gets there because the NIC produces an interrupt for each packet (or for every few packets) and when the kernel handles these the data is copied to user space. Then, as you point out, the cost of the RETI can't be avoided.

yes, in interrupt context. My point is that that *task* will never go in to kernel mode, it will always be running in user mode.

Quoted 4 lines

However, if the NIC could transfer the data directly to user space it wouldn't need to cause an interrupt and the cost of the RETI and the context switch is avoided. The user mode app really could process forever without sleeping at that point.

it would need to cause an interrupt to notify of the packet, unless the program communicated directly with the NIC in polling mode. This the point of mmap packet socket, provides a *portable* ring buffer structure so that userspace doesn't have to reimplement drtivers.

Signature · 3 lines
// Gianni Tedesco (gianni at scaramanga dot co dot uk)
lynx --source www.scaramanga.co.uk/gianni-at-ecsc.asc | gpg --import
8646BE7D: 6D9F 2287 870E A2C9 8F60 3A3C 91B5 7669 8646 BE7D
#5 lore raw
Quoted 7 lines · Gianni Tedesco
Quoted 1 line · Oliver Dain

It seems to me that it can't loop in user mode forever. There is no way to get data into user mode (the ring buffer) witout going through the kernel. My understanding is that the NIC doesn't transfer directly to the user mode ring buffer, but rather to a different DMA buffer. The kernel must copy it from the DMA buffer to the ring buffer. Thus once the user mode app has processed all the data in the ring buffer the kenel _must_ get involved to get more data to user space. Currently the data gets there because the NIC produces an interrupt for each packet (or for every few packets) and when the kernel handles these the data is copied to user space. Then, as you point out, the cost of the RETI can't be avoided.

yes, in interrupt context. My point is that that *task* will never go in to kernel mode, it will always be running in user mode.

In my experience (PIII 1.2GHz, i815, e100, NAPI), user mode would read at most 7 packets at a time, even when artificial busy loops insterted. The max packet rate acheived was around 120Kpps, but that was limited at the driver level. Most of the CPU was consumed while doing this (measured with cyclesoak, especially required since NAPI was used).

Pádraig.