Re: mmap of a network buffer
I really do not know how to describe the problem. But a friend here asks me how to mmap a network buffer so that there is no need to copy the data from user space to kernel space. We are not sure whether FreeBSD can create a device file (mknod) for a network card, and if so, we can use the mmap() call to do so because mmap() requires a file descriptor. We assume that the file descriptor can be acquired by opening the network device. If this is infeasible, is there another way to accomplish the same goal? Use sendfile() for zero-copy file transmission; in all other cases it's necessary to copy data into the kernel. Memory-mapping a network buffer makes no sense if you just think about it for a moment... There's also very little need for this under real circumstances; some simple tests have demonstrated we can sustain about 800Mbps throughput (UDP), and the bottleneck here seems to be checksum calculations, not copyin/out. Oddly enough, I was just getting ready to implement something like this. Not because of copyin performance issues, but because async io for sockets could be done better if I didn't have to do a copyin. copyin has to have curproc==(proc with the buffer from which to copy) That's basically right. You have three options: - Switch to process context to access process data; this allows you to take page faults in controlled circumstances (eg. copyin). - Wire the process' pages into the kernel so you don't have to fault. - Copy the user data into kernel space in an efficient fashion. which means that I have to do a context switch for every socket buffer sized chunk (best case) or every io op (worst case). It sounds like your buffering is not efficient. My hope was to map the user's buffer into kernel space so that I could do event driven io on the socket without having to context switch to an aiod for every io operation. Is this really a bad idea? I am a little concerned about running out of kernel address space, but I don't think that's an immediate problem. If you map into the kernel, you still have to context switch unless you wire the data down. Excessive wiring can be expensive. Have a look at how physio() does it's thing. Such an implementation would lend itself to doing zero-copy writes async writes with some frobbing of the send routines. It would also bypass some of the messing around done to do socket buffers--that is, there would not be a limit per se on socket buffering for writes since they would be mapped user space. One might want to put arbitrary limits in place to ensure that an unreasonable amount of memory isn't locked. Thoughts? Sounds a lot like sendfile. See if you can't improve on it to do eg. sendmem(). -- \\ The mind's the standard \\ Mike Smith \\ of the man. \\ msm...@freebsd.org \\-- Joseph Merrick \\ msm...@cdrom.com To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message
Re: mmap of a network buffer
On Mon, 24 May 1999, Mike Smith wrote: There's also very little need for this under real circumstances; some simple tests have demonstrated we can sustain about 800Mbps throughput (UDP), and the bottleneck here seems to be checksum calculations, not copyin/out. Oddly enough, I was just getting ready to implement something like this. Not because of copyin performance issues, but because async io for sockets could be done better if I didn't have to do a copyin. copyin has to have curproc==(proc with the buffer from which to copy) That's basically right. You have three options: - Switch to process context to access process data; this allows you to take page faults in controlled circumstances (eg. copyin). - Wire the process' pages into the kernel so you don't have to fault. - Copy the user data into kernel space in an efficient fashion. Glad to know that my understanding wasn't too far off-base. which means that I have to do a context switch for every socket buffer sized chunk (best case) or every io op (worst case). It sounds like your buffering is not efficient. Well, I'd be happy if I could be convinced that were the problem, but protocols like HTTP and NNTP which have short, rapid-fire (sometimes lock-step) command sequences don't help the buffering any. This means that reading commands of an incoming connection causes many, many context switches between an aiod and the main process doing async io. On the outgoing side, in the optimal case of sending large blocks of data, I don't have control over the buffering--the aiod essentially impersonates my process, going to sleep in the socket write routines, context switching to it when a copyin is necessary. I could exert more control by metering my writes so that they fit into socket buffers and avoid the switches, but that increases the number of system calls, so I'm not sure how much of a win it ends up. Plus, I hope that I will get some added advantage out of constructing a zero-copy write (not that I've had any throughput troubles). My hope was to map the user's buffer into kernel space so that I could do event driven io on the socket without having to context switch to an aiod for every io operation. Is this really a bad idea? I am a little concerned about running out of kernel address space, but I don't think that's an immediate problem. If you map into the kernel, you still have to context switch unless you wire the data down. Excessive wiring can be expensive. Have a look at how physio() does it's thing. Will do. There's some of that code in the async io routines now, for dealing with raw io operations--I hoped to borrow from that to implement my stuff. Such an implementation would lend itself to doing zero-copy writes async writes with some frobbing of the send routines. It would also bypass some of the messing around done to do socket buffers--that is, there would not be a limit per se on socket buffering for writes since they would be mapped user space. One might want to put arbitrary limits in place to ensure that an unreasonable amount of memory isn't locked. Thoughts? Sounds a lot like sendfile. See if you can't improve on it to do eg. sendmem(). Yes. I'd like mine to be async rather than synchronous, though. I've considered creating an async sendfile too. (Actually, I've been thinking about extending the async io code to allow calling any syscall async, but there are other complexities there...). -Chris To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message
Re: mmap of a network buffer
I really do not know how to describe the problem. But a friend here asks me how to mmap a network buffer so that there is no need to copy the data from user space to kernel space. We are not sure whether FreeBSD can create a device file (mknod) for a network card, and if so, we can use the mmap() call to do so because mmap() requires a file descriptor. We assume that the file descriptor can be acquired by opening the network device. If this is infeasible, is there another way to accomplish the same goal? Use sendfile() for zero-copy file transmission; in all other cases it's necessary to copy data into the kernel. Memory-mapping a network buffer makes no sense if you just think about it for a moment... There's also very little need for this under real circumstances; some simple tests have demonstrated we can sustain about 800Mbps throughput (UDP), and the bottleneck here seems to be checksum calculations, not copyin/out. -- \\ The mind's the standard \\ Mike Smith \\ of the man. \\ msm...@freebsd.org \\-- Joseph Merrick \\ msm...@cdrom.com To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message
Re: mmap of a network buffer
On Fri, 21 May 1999, Mike Smith wrote: I really do not know how to describe the problem. But a friend here asks me how to mmap a network buffer so that there is no need to copy the data from user space to kernel space. We are not sure whether FreeBSD can create a device file (mknod) for a network card, and if so, we can use the mmap() call to do so because mmap() requires a file descriptor. We assume that the file descriptor can be acquired by opening the network device. If this is infeasible, is there another way to accomplish the same goal? Use sendfile() for zero-copy file transmission; in all other cases it's necessary to copy data into the kernel. Memory-mapping a network buffer makes no sense if you just think about it for a moment... There's also very little need for this under real circumstances; some simple tests have demonstrated we can sustain about 800Mbps throughput (UDP), and the bottleneck here seems to be checksum calculations, not copyin/out. Oddly enough, I was just getting ready to implement something like this. Not because of copyin performance issues, but because async io for sockets could be done better if I didn't have to do a copyin. copyin has to have curproc==(proc with the buffer from which to copy) which means that I have to do a context switch for every socket buffer sized chunk (best case) or every io op (worst case). In my testing, I've been doing 3000 context switches/second just in verification of a simple configuration--real load will be 20-100x what current load is, and I'm not sure that 60,000-300,000 context switches/sec is desirable. My hope was to map the user's buffer into kernel space so that I could do event driven io on the socket without having to context switch to an aiod for every io operation. Is this really a bad idea? I am a little concerned about running out of kernel address space, but I don't think that's an immediate problem. Such an implementation would lend itself to doing zero-copy writes async writes with some frobbing of the send routines. It would also bypass some of the messing around done to do socket buffers--that is, there would not be a limit per se on socket buffering for writes since they would be mapped user space. One might want to put arbitrary limits in place to ensure that an unreasonable amount of memory isn't locked. Thoughts? -Chris To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message
RE: mmap of a network buffer
My hope was to map the user's buffer into kernel space so that I could do event driven io on the socket without having to context switch to an aiod for every io operation. Is this really a bad idea? I am a little concerned about running out of kernel address space, but I don't think that's an immediate problem. Such an implementation would lend itself to doing zero-copy writes async writes with some frobbing of the send routines. It would also bypass some of the messing around done to do socket buffers--that is, there would not be a limit per se on socket buffering for writes since they would be mapped user space. One might want to put arbitrary limits in place to ensure that an unreasonable amount of memory isn't locked. Thoughts? In my view, the problem can be described like this. Some applications need to process data from their VA space, on some devices. If the data is going to/from a file, it looks perfectly well to copy it into kernel buffers, since the kernel does caching and improves disk I/O performance. However, there are cases when the kernel can't be concerned with the data. For example, I have an encryption/compression processor on PCI board. For each operation, this processor needs two separated data buffers and performs the busmaster DMA. The user program is supposed to prepare the buffers and communicate their location to the kernel mode driver via IOCTL. What is more efficient - copy the data to/from the locked kernel buffers or lock the user buffers in place and do processing? (In my case, I don't even need to _remap_ the buffers, I only need physical addresses). I'd prefer the later, but I don't have sufficient FreeBSD knowledge to insist that I think right. There may be some principles of this O/S that I don't currently see that I violate by doing this. It would be nice if somebody could give an analysis of the problem. Stan To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message
Re: RE: mmap of a network buffer
:In my view, the problem can be described like this. : :Some applications need to process data from their VA space, on some :devices. If the data is going to/from a file, it looks perfectly :well to copy it into kernel buffers, since the kernel does caching :and improves disk I/O performance. However, there are cases when the :kernel can't be concerned with the data. For example, I have an :encryption/compression processor on PCI board. For each operation, :this processor needs two separated data buffers and performs the :busmaster DMA. The user program is supposed to prepare the buffers :and communicate their location to the kernel mode driver via IOCTL. : :What is more efficient - copy the data to/from the locked kernel :buffers or lock the user buffers in place and do processing? :(In my case, I don't even need to _remap_ the buffers, I only need :physical addresses). : :I'd prefer the later, but I don't have sufficient FreeBSD knowledge :to insist that I think right. There may be some principles of this :O/S that I don't currently see that I violate by doing this. : :It would be nice if somebody could give an analysis of the problem. : :Stan Well, all the system buffer paradigm does is wire the pages and associate them with a struct buf. You do not have to map the pages into KVM. It also usually write-protects pages in user space for the duration of the I/O. Even if the pages are mapped into KVM, the overhead is virtually nil if you do not actually touch the associated KVM. I don't think you would notice the difference between using the existing buffer code and rolling something custom. -Matt Matthew Dillon dil...@backplane.com To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message