d/freebsd-skq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
8aa8e94d34
freebsd-skq/sys/net/bpf_zerocopy.c

603 lines
16 KiB
C
Raw Normal View History

Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
/*-
* Copyright (c) 2007 Seccuris Inc.
* All rights reserved.
*
* This sofware was developed by Robert N. M. Watson under contract to
* Seccuris Inc.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_bpf.h"
#include <sys/param.h>
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/mbuf.h>
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/sf_buf.h>
#include <sys/socket.h>
#include <sys/uio.h>
#include <machine/atomic.h>
#include <net/if.h>
#include <net/bpf.h>
#include <net/bpf_zerocopy.h>
Fix build with option BPF_JITTER.
2008-03-24 22:21:32 +00:00
#include <net/bpfdesc.h>
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
#include <vm/vm.h>
#include <vm/pmap.h>
#include <vm/vm_extern.h>
#include <vm/vm_map.h>
#include <vm/vm_page.h>
/*
* Zero-copy buffer scheme for BPF: user space "donates" two buffers, which
* are mapped into the kernel address space using sf_bufs and used directly
* by BPF. Memory is wired since page faults cannot be tolerated in the
* contexts where the buffers are copied to (locks held, interrupt context,
* etc). Access to shared memory buffers is synchronized using a header on
* each buffer, allowing the number of system calls to go to zero as BPF
* reaches saturation (buffers filled as fast as they can be drained by the
* user process). Full details of the protocol for communicating between the
* user process and BPF may be found in bpf(4).
*/
/*
* Maximum number of pages per buffer. Since all BPF devices use two, the
* maximum per device is 2*BPF_MAX_PAGES. Resource limits on the number of
* sf_bufs may be an issue, so do not set this too high. On older systems,
* kernel address space limits may also be an issue.
*/
#define BPF_MAX_PAGES 512
/*
* struct zbuf describes a memory buffer loaned by a user process to the
* kernel. We represent this as a series of pages managed using an array of
* sf_bufs. Even though the memory is contiguous in user space, it may not
* be mapped contiguously in the kernel (i.e., a set of physically
* non-contiguous pages in the direct map region) so we must implement
* scatter-gather copying. One significant mitigating factor is that on
* systems with a direct memory map, we can avoid TLB misses.
*
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
* At the front of the shared memory region is a bpf_zbuf_header, which
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
* contains shared control data to allow user space and the kernel to
* synchronize; this is included in zb_size, but not bpf_bufsize, so that BPF
* knows that the space is not available.
*/
struct zbuf {
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
vm_offset_t zb_uaddr; /* User address at time of setup. */
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
size_t zb_size; /* Size of buffer, incl. header. */
u_int zb_numpages; /* Number of pages. */
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
int zb_flags; /* Flags on zbuf. */
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
struct sf_buf **zb_pages; /* Pages themselves. */
struct bpf_zbuf_header *zb_header; /* Shared header. */
};
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
/*
* When a buffer has been assigned to userspace, flag it as such, as the
* buffer may remain in the store position as a result of the user process
* not yet having acknowledged the buffer in the hold position yet.
*/
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
#define ZBUF_FLAG_ASSIGNED 0x00000001 /* Set when owned by user. */
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
/*
* Release a page we've previously wired.
*/
static void
zbuf_page_free(vm_page_t pp)
{
This is the first step in transitioning responsibility for synchronizing access to the page's wire_count from the page queues lock to the page lock. Submitted by: kmacy
2010-05-03 05:41:50 +00:00
vm_page_lock(pp);
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
vm_page_unwire(pp, 0);
if (pp->wire_count == 0 && pp->object == NULL)
vm_page_free(pp);
This is the first step in transitioning responsibility for synchronizing access to the page's wire_count from the page queues lock to the page lock. Submitted by: kmacy
2010-05-03 05:41:50 +00:00
vm_page_unlock(pp);
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
}
/*
* Free an sf_buf with attached page.
*/
static void
zbuf_sfbuf_free(struct sf_buf *sf)
{
vm_page_t pp;
pp = sf_buf_page(sf);
sf_buf_free(sf);
zbuf_page_free(pp);
}
/*
* Free a zbuf, including its page array, sbufs, and pages. Allow partially
* allocated zbufs to be freed so that it may be used even during a zbuf
* setup.
*/
static void
zbuf_free(struct zbuf *zb)
{
int i;
for (i = 0; i < zb->zb_numpages; i++) {
if (zb->zb_pages[i] != NULL)
zbuf_sfbuf_free(zb->zb_pages[i]);
}
free(zb->zb_pages, M_BPF);
free(zb, M_BPF);
}
/*
* Given a user pointer to a page of user memory, return an sf_buf for the
* page. Because we may be requesting quite a few sf_bufs, prefer failure to
* deadlock and use SFB_NOWAIT.
*/
static struct sf_buf *
zbuf_sfbuf_get(struct vm_map *map, vm_offset_t uaddr)
{
struct sf_buf *sf;
vm_page_t pp;
if (vm_fault_quick((caddr_t) uaddr, VM_PROT_READ | VM_PROT_WRITE) <
0)
return (NULL);
pp = pmap_extract_and_hold(map->pmap, uaddr, VM_PROT_READ |
VM_PROT_WRITE);
if (pp == NULL)
return (NULL);
On Alan's advice, rather than do a wholesale conversion on a single architecture from page queue lock to a hashed array of page locks (based on a patch by Jeff Roberson), I've implemented page lock support in the MI code and have only moved vm_page's hold_count out from under page queue mutex to page lock. This changes pmap_extract_and_hold on all pmaps. Supported by: Bitgravity Inc. Discussed with: alc, jeffr, and kib
2010-04-30 00:46:43 +00:00
vm_page_lock(pp);
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
vm_page_wire(pp);
vm_page_unhold(pp);
On Alan's advice, rather than do a wholesale conversion on a single architecture from page queue lock to a hashed array of page locks (based on a patch by Jeff Roberson), I've implemented page lock support in the MI code and have only moved vm_page's hold_count out from under page queue mutex to page lock. This changes pmap_extract_and_hold on all pmaps. Supported by: Bitgravity Inc. Discussed with: alc, jeffr, and kib
2010-04-30 00:46:43 +00:00
vm_page_unlock(pp);
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
sf = sf_buf_alloc(pp, SFB_NOWAIT);
if (sf == NULL) {
zbuf_page_free(pp);
return (NULL);
}
return (sf);
}
/*
* Create a zbuf describing a range of user address space memory. Validate
* page alignment, size requirements, etc.
*/
static int
zbuf_setup(struct thread *td, vm_offset_t uaddr, size_t len,
struct zbuf **zbp)
{
struct zbuf *zb;
struct vm_map *map;
int error, i;
*zbp = NULL;
/*
* User address must be page-aligned.
*/
if (uaddr & PAGE_MASK)
return (EINVAL);
/*
* Length must be an integer number of full pages.
*/
if (len & PAGE_MASK)
return (EINVAL);
/*
* Length must not exceed per-buffer resource limit.
*/
if ((len / PAGE_SIZE) > BPF_MAX_PAGES)
return (EINVAL);
/*
* Allocate the buffer and set up each page with is own sf_buf.
*/
error = 0;
zb = malloc(sizeof(*zb), M_BPF, M_ZERO | M_WAITOK);
zb->zb_uaddr = uaddr;
zb->zb_size = len;
zb->zb_numpages = len / PAGE_SIZE;
zb->zb_pages = malloc(sizeof(struct sf_buf *) *
zb->zb_numpages, M_BPF, M_ZERO | M_WAITOK);
map = &td->td_proc->p_vmspace->vm_map;
for (i = 0; i < zb->zb_numpages; i++) {
zb->zb_pages[i] = zbuf_sfbuf_get(map,
uaddr + (i * PAGE_SIZE));
if (zb->zb_pages[i] == NULL) {
error = EFAULT;
goto error;
}
}
zb->zb_header =
(struct bpf_zbuf_header *)sf_buf_kva(zb->zb_pages[0]);
bzero(zb->zb_header, sizeof(*zb->zb_header));
*zbp = zb;
return (0);
error:
zbuf_free(zb);
return (error);
}
/*
* Copy bytes from a source into the specified zbuf. The caller is
* responsible for performing bounds checking, etc.
*/
void
bpf_zerocopy_append_bytes(struct bpf_d *d, caddr_t buf, u_int offset,
void *src, u_int len)
{
u_int count, page, poffset;
u_char *src_bytes;
struct zbuf *zb;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_append_bytes: not in zbuf mode"));
KASSERT(buf != NULL, ("bpf_zerocopy_append_bytes: NULL buf"));
src_bytes = (u_char *)src;
zb = (struct zbuf *)buf;
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
KASSERT((zb->zb_flags & ZBUF_FLAG_ASSIGNED) == 0,
("bpf_zerocopy_append_bytes: ZBUF_FLAG_ASSIGNED"));
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
/*
* Scatter-gather copy to user pages mapped into kernel address space
* using sf_bufs: copy up to a page at a time.
*/
offset += sizeof(struct bpf_zbuf_header);
page = offset / PAGE_SIZE;
poffset = offset % PAGE_SIZE;
while (len > 0) {
KASSERT(page < zb->zb_numpages, ("bpf_zerocopy_append_bytes:"
" page overflow (%d p %d np)\n", page, zb->zb_numpages));
count = min(len, PAGE_SIZE - poffset);
bcopy(src_bytes, ((u_char *)sf_buf_kva(zb->zb_pages[page])) +
poffset, count);
poffset += count;
if (poffset == PAGE_SIZE) {
poffset = 0;
page++;
}
KASSERT(poffset < PAGE_SIZE,
("bpf_zerocopy_append_bytes: page offset overflow (%d)",
poffset));
len -= count;
src_bytes += count;
}
}
/*
* Copy bytes from an mbuf chain to the specified zbuf: copying will be
* scatter-gather both from mbufs, which may be fragmented over memory, and
* to pages, which may not be contiguously mapped in kernel address space.
* As with bpf_zerocopy_append_bytes(), the caller is responsible for
* checking that this will not exceed the buffer limit.
*/
void
bpf_zerocopy_append_mbuf(struct bpf_d *d, caddr_t buf, u_int offset,
void *src, u_int len)
{
u_int count, moffset, page, poffset;
const struct mbuf *m;
struct zbuf *zb;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_append_mbuf not in zbuf mode"));
KASSERT(buf != NULL, ("bpf_zerocopy_append_mbuf: NULL buf"));
m = (struct mbuf *)src;
zb = (struct zbuf *)buf;
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
KASSERT((zb->zb_flags & ZBUF_FLAG_ASSIGNED) == 0,
("bpf_zerocopy_append_mbuf: ZBUF_FLAG_ASSIGNED"));
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
/*
* Scatter gather both from an mbuf chain and to a user page set
* mapped into kernel address space using sf_bufs. If we're lucky,
* each mbuf requires one copy operation, but if page alignment and
* mbuf alignment work out less well, we'll be doing two copies per
* mbuf.
*/
offset += sizeof(struct bpf_zbuf_header);
page = offset / PAGE_SIZE;
poffset = offset % PAGE_SIZE;
moffset = 0;
while (len > 0) {
KASSERT(page < zb->zb_numpages,
("bpf_zerocopy_append_mbuf: page overflow (%d p %d "
"np)\n", page, zb->zb_numpages));
KASSERT(m != NULL,
("bpf_zerocopy_append_mbuf: end of mbuf chain"));
count = min(m->m_len - moffset, len);
count = min(count, PAGE_SIZE - poffset);
bcopy(mtod(m, u_char *) + moffset,
((u_char *)sf_buf_kva(zb->zb_pages[page])) + poffset,
count);
poffset += count;
if (poffset == PAGE_SIZE) {
poffset = 0;
page++;
}
KASSERT(poffset < PAGE_SIZE,
("bpf_zerocopy_append_mbuf: page offset overflow (%d)",
poffset));
moffset += count;
if (moffset == m->m_len) {
m = m->m_next;
moffset = 0;
}
len -= count;
}
}
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
/*
* Notification from the BPF framework that a buffer in the store position is
* rejecting packets and may be considered full. We mark the buffer as
* immutable and assign to userspace so that it is immediately available for
* the user process to access.
*/
void
bpf_zerocopy_buffull(struct bpf_d *d)
{
struct zbuf *zb;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_buffull: not in zbuf mode"));
zb = (struct zbuf *)d->bd_sbuf;
KASSERT(zb != NULL, ("bpf_zerocopy_buffull: zb == NULL"));
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
if ((zb->zb_flags & ZBUF_FLAG_ASSIGNED) == 0) {
zb->zb_flags |= ZBUF_FLAG_ASSIGNED;
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
zb->zb_header->bzh_kernel_len = d->bd_slen;
atomic_add_rel_int(&zb->zb_header->bzh_kernel_gen, 1);
}
}
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
/*
* Notification from the BPF framework that a buffer has moved into the held
* slot on a descriptor. Zero-copy BPF will update the shared page to let
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
* the user process know and flag the buffer as assigned if it hasn't already
* been marked assigned due to filling while it was in the store position.
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
*
* Note: identical logic as in bpf_zerocopy_buffull(), except that we operate
* on bd_hbuf and bd_hlen.
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
*/
void
bpf_zerocopy_bufheld(struct bpf_d *d)
{
struct zbuf *zb;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_bufheld: not in zbuf mode"));
zb = (struct zbuf *)d->bd_hbuf;
KASSERT(zb != NULL, ("bpf_zerocopy_bufheld: zb == NULL"));
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
if ((zb->zb_flags & ZBUF_FLAG_ASSIGNED) == 0) {
zb->zb_flags |= ZBUF_FLAG_ASSIGNED;
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
zb->zb_header->bzh_kernel_len = d->bd_hlen;
atomic_add_rel_int(&zb->zb_header->bzh_kernel_gen, 1);
}
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
}
Make sure we are clearing the ZBUF_FLAG_IMMUTABLE any time a free buffer is reclaimed by the kernel. This fixes a bug resulted in the kernel over writing packet data while user-space was still processing it when zerocopy is enabled. (Or a panic if invariants was enabled). Discussed with: rwatson
2008-07-05 20:11:28 +00:00
/*
* Notification from the BPF framework that the free buffer has been been
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
* rotated out of the held position to the free position. This happens when
* the user acknowledges the held buffer.
Make sure we are clearing the ZBUF_FLAG_IMMUTABLE any time a free buffer is reclaimed by the kernel. This fixes a bug resulted in the kernel over writing packet data while user-space was still processing it when zerocopy is enabled. (Or a panic if invariants was enabled). Discussed with: rwatson
2008-07-05 20:11:28 +00:00
*/
void
bpf_zerocopy_buf_reclaimed(struct bpf_d *d)
{
struct zbuf *zb;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_reclaim_buf: not in zbuf mode"));
KASSERT(d->bd_fbuf != NULL,
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
("bpf_zerocopy_buf_reclaimed: NULL free buf"));
Make sure we are clearing the ZBUF_FLAG_IMMUTABLE any time a free buffer is reclaimed by the kernel. This fixes a bug resulted in the kernel over writing packet data while user-space was still processing it when zerocopy is enabled. (Or a panic if invariants was enabled). Discussed with: rwatson
2008-07-05 20:11:28 +00:00
zb = (struct zbuf *)d->bd_fbuf;
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
zb->zb_flags &= ~ZBUF_FLAG_ASSIGNED;
Make sure we are clearing the ZBUF_FLAG_IMMUTABLE any time a free buffer is reclaimed by the kernel. This fixes a bug resulted in the kernel over writing packet data while user-space was still processing it when zerocopy is enabled. (Or a panic if invariants was enabled). Discussed with: rwatson
2008-07-05 20:11:28 +00:00
}
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
/*
* Query from the BPF framework regarding whether the buffer currently in the
* held position can be moved to the free position, which can be indicated by
* the user process making their generation number equal to the kernel
* generation number.
*/
int
bpf_zerocopy_canfreebuf(struct bpf_d *d)
{
struct zbuf *zb;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_canfreebuf: not in zbuf mode"));
zb = (struct zbuf *)d->bd_hbuf;
if (zb == NULL)
return (0);
if (zb->zb_header->bzh_kernel_gen ==
atomic_load_acq_int(&zb->zb_header->bzh_user_gen))
return (1);
return (0);
}
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
/*
* Query from the BPF framework as to whether or not the buffer current in
* the store position can actually be written to. This may return false if
* the store buffer is assigned to userspace before the hold buffer is
* acknowledged.
*/
int
bpf_zerocopy_canwritebuf(struct bpf_d *d)
{
struct zbuf *zb;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_canwritebuf: not in zbuf mode"));
zb = (struct zbuf *)d->bd_sbuf;
KASSERT(zb != NULL, ("bpf_zerocopy_canwritebuf: bd_sbuf NULL"));
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
if (zb->zb_flags & ZBUF_FLAG_ASSIGNED)
Maintain and observe a ZBUF_FLAG_IMMUTABLE flag on zero-copy BPF buffer kernel descriptors, which is used to allow the buffer currently in the BPF "store" position to be assigned to userspace when it fills, even if userspace hasn't acknowledged the buffer in the "hold" position yet. To implement this, notify the buffer model when a buffer becomes full, and check that the store buffer is writable, not just for it being full, before trying to append new packet data. Shared memory buffers will be assigned to userspace at most once per fill, be it in the store or in the hold position. This removes the restriction that at most one shared memory can by owned by userspace, reducing the chances that userspace will need to call select() after acknowledging one buffer in order to wait for the next buffer when under high load. This more fully realizes the goal of zero system calls in order to process a high-speed packet stream from BPF. Update bpf.4 to reflect that both buffers may be owned by userspace at once; caution against assuming this.
2008-04-07 02:51:00 +00:00
return (0);
return (1);
}
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
/*
* Free zero copy buffers at request of descriptor.
*/
void
bpf_zerocopy_free(struct bpf_d *d)
{
struct zbuf *zb;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_free: not in zbuf mode"));
zb = (struct zbuf *)d->bd_sbuf;
if (zb != NULL)
zbuf_free(zb);
zb = (struct zbuf *)d->bd_hbuf;
if (zb != NULL)
zbuf_free(zb);
zb = (struct zbuf *)d->bd_fbuf;
if (zb != NULL)
zbuf_free(zb);
}
/*
* Ioctl to return the maximum buffer size.
*/
int
bpf_zerocopy_ioctl_getzmax(struct thread *td, struct bpf_d *d, size_t *i)
{
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_ioctl_getzmax: not in zbuf mode"));
*i = BPF_MAX_PAGES * PAGE_SIZE;
return (0);
}
/*
* Ioctl to force rotation of the two buffers, if there's any data available.
Clarify some comments, fix some types, and rename ZBUF_FLAG_IMMUTABLE to ZBUF_FLAG_ASSIGNED to make it clear why the buffer can't be written to: it is assigned to userspace.
2009-03-07 10:21:37 +00:00
* This can be used by user space to implement timeouts when waiting for a
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
* buffer to fill.
*/
int
bpf_zerocopy_ioctl_rotzbuf(struct thread *td, struct bpf_d *d,
struct bpf_zbuf *bz)
{
struct zbuf *bzh;
bzero(bz, sizeof(*bz));
BPFD_LOCK(d);
if (d->bd_hbuf == NULL && d->bd_slen != 0) {
ROTATE_BUFFERS(d);
bzh = (struct zbuf *)d->bd_hbuf;
bz->bz_bufa = (void *)bzh->zb_uaddr;
bz->bz_buflen = d->bd_hlen;
}
BPFD_UNLOCK(d);
return (0);
}
/*
* Ioctl to configure zero-copy buffers -- may be done only once.
*/
int
bpf_zerocopy_ioctl_setzbuf(struct thread *td, struct bpf_d *d,
struct bpf_zbuf *bz)
{
struct zbuf *zba, *zbb;
int error;
KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
("bpf_zerocopy_ioctl_setzbuf: not in zbuf mode"));
/*
* Must set both buffers. Cannot clear them.
*/
if (bz->bz_bufa == NULL || bz->bz_bufb == NULL)
return (EINVAL);
/*
* Buffers must have a size greater than 0. Alignment and other size
* validity checking is done in zbuf_setup().
*/
if (bz->bz_buflen == 0)
return (EINVAL);
/*
* Allocate new buffers.
*/
error = zbuf_setup(td, (vm_offset_t)bz->bz_bufa, bz->bz_buflen,
&zba);
if (error)
return (error);
error = zbuf_setup(td, (vm_offset_t)bz->bz_bufb, bz->bz_buflen,
&zbb);
if (error) {
zbuf_free(zba);
return (error);
}
/*
* We only allow buffers to be installed once, so atomically check
* that no buffers are currently installed and install new buffers.
*/
BPFD_LOCK(d);
if (d->bd_hbuf != NULL || d->bd_sbuf != NULL || d->bd_fbuf != NULL ||
d->bd_bif != NULL) {
BPFD_UNLOCK(d);
zbuf_free(zba);
zbuf_free(zbb);
return (EINVAL);
}
Add a comment explaining that we initialize the 'a' buffer for zero-copy to the store buffer position on the BPF descriptor, and the 'b' buffer as the free buffer in order to fill them in the order documented in bpf(4). MFC after: 4 months Suggested by: csjp
2008-03-26 21:29:13 +00:00
/*
* Point BPF descriptor at buffers; initialize sbuf as zba so that
* it is always filled first in the sequence, per bpf(4).
*/
Introduce support for zero-copy BPF buffering, which reduces the overhead of packet capture by allowing a user process to directly "loan" buffer memory to the kernel rather than using read(2) to explicitly copy data from kernel address space. The user process will issue new BPF ioctls to set the shared memory buffer mode and provide pointers to buffers and their size. The kernel then wires and maps the pages into kernel address space using sf_buf(9), which on supporting architectures will use the direct map region. The current "buffered" access mode remains the default, and support for zero-copy buffers must, for the time being, be explicitly enabled using a sysctl for the kernel to accept requests to use it. The kernel and user process synchronize use of the buffers with atomic operations, avoiding the need for system calls under load; the user process may use select()/poll()/kqueue() to manage blocking while waiting for network data if the user process is able to consume data faster than the kernel generates it. Patchs to libpcap are available to allow libpcap applications to transparently take advantage of this support. Detailed information on the new API may be found in bpf(4), including specific atomic operations and memory barriers required to synchronize buffer use safely. These changes modify the base BPF implementation to (roughly) abstrac the current buffer model, allowing the new shared memory model to be added, and add new monitoring statistics for netstat to print. The implementation, with the exception of some monitoring hanges that break the netstat monitoring ABI for BPF, will be MFC'd. Zerocopy bpf buffers are still considered experimental are disabled by default. To experiment with this new facility, adjust the net.bpf.zerocopy_enable sysctl variable to 1. Changes to libpcap will be made available as a patch for the time being, and further refinements to the implementation are expected. Sponsored by: Seccuris Inc. In collaboration with: rwatson Tested by: pwood, gallatin MFC after: 4 months [1] [1] Certain portions will probably not be MFCed, specifically things that can break the monitoring ABI.
2008-03-24 13:49:17 +00:00
d->bd_fbuf = (caddr_t)zbb;
d->bd_sbuf = (caddr_t)zba;
d->bd_slen = 0;
d->bd_hlen = 0;
/*
* We expose only the space left in the buffer after the size of the
* shared management region.
*/
d->bd_bufsize = bz->bz_buflen - sizeof(struct bpf_zbuf_header);
BPFD_UNLOCK(d);
return (0);
}
Reference in New Issue Copy Permalink