2011-03-15 00:06:18 +00:00
|
|
|
/*
|
|
|
|
* Back-end of the driver for virtual network devices. This portion of the
|
|
|
|
* driver exports a 'unified' network-device interface that can be accessed
|
|
|
|
* by any operating system that implements a compatible front end. A
|
|
|
|
* reference front-end implementation can be found in:
|
|
|
|
* drivers/net/xen-netfront.c
|
|
|
|
*
|
|
|
|
* Copyright (c) 2002-2005, K A Fraser
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public License version 2
|
|
|
|
* as published by the Free Software Foundation; or, when distributed
|
|
|
|
* separately from the Linux kernel or incorporated into other
|
|
|
|
* software packages, subject to the following license:
|
|
|
|
*
|
|
|
|
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
|
|
* of this source file (the "Software"), to deal in the Software without
|
|
|
|
* restriction, including without limitation the rights to use, copy, modify,
|
|
|
|
* merge, publish, distribute, sublicense, and/or sell copies of the Software,
|
|
|
|
* and to permit persons to whom the Software is furnished to do so, subject to
|
|
|
|
* the following conditions:
|
|
|
|
*
|
|
|
|
* The above copyright notice and this permission notice shall be included in
|
|
|
|
* all copies or substantial portions of the Software.
|
|
|
|
*
|
|
|
|
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
|
|
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
|
|
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
|
|
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
|
|
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
|
|
|
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
|
|
|
|
* IN THE SOFTWARE.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "common.h"
|
|
|
|
|
|
|
|
#include <linux/kthread.h>
|
|
|
|
#include <linux/if_vlan.h>
|
|
|
|
#include <linux/udp.h>
|
2014-03-06 21:48:29 +00:00
|
|
|
#include <linux/highmem.h>
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
#include <net/tcp.h>
|
|
|
|
|
2012-08-08 17:21:23 +00:00
|
|
|
#include <xen/xen.h>
|
2011-03-15 00:06:18 +00:00
|
|
|
#include <xen/events.h>
|
|
|
|
#include <xen/interface/memory.h>
|
|
|
|
|
|
|
|
#include <asm/xen/hypercall.h>
|
|
|
|
#include <asm/xen/page.h>
|
|
|
|
|
2013-05-22 06:34:45 +00:00
|
|
|
/* Provide an option to disable split event channels at load time as
|
|
|
|
* event channels are limited resource. Split event channels are
|
|
|
|
* enabled by default.
|
|
|
|
*/
|
|
|
|
bool separate_tx_rx_irq = 1;
|
|
|
|
module_param(separate_tx_rx_irq, bool, 0644);
|
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
/* The time that packets can stay on the guest Rx internal queue
|
|
|
|
* before they are dropped.
|
2014-03-06 21:48:30 +00:00
|
|
|
*/
|
|
|
|
unsigned int rx_drain_timeout_msecs = 10000;
|
|
|
|
module_param(rx_drain_timeout_msecs, uint, 0444);
|
|
|
|
|
2014-10-22 13:08:55 +00:00
|
|
|
/* The length of time before the frontend is considered unresponsive
|
|
|
|
* because it isn't providing Rx slots.
|
|
|
|
*/
|
2014-12-18 11:13:06 +00:00
|
|
|
unsigned int rx_stall_timeout_msecs = 60000;
|
2014-10-22 13:08:55 +00:00
|
|
|
module_param(rx_stall_timeout_msecs, uint, 0444);
|
|
|
|
|
2014-06-04 09:30:43 +00:00
|
|
|
unsigned int xenvif_max_queues;
|
|
|
|
module_param_named(max_queues, xenvif_max_queues, uint, 0644);
|
|
|
|
MODULE_PARM_DESC(max_queues,
|
|
|
|
"Maximum number of queues per virtual interface");
|
|
|
|
|
2013-04-22 02:20:42 +00:00
|
|
|
/*
|
|
|
|
* This is the maximum slots a skb can have. If a guest sends a skb
|
|
|
|
* which exceeds this limit it is considered malicious.
|
|
|
|
*/
|
2013-05-02 00:43:59 +00:00
|
|
|
#define FATAL_SKB_SLOTS_DEFAULT 20
|
|
|
|
static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
|
|
|
|
module_param(fatal_skb_slots, uint, 0444);
|
|
|
|
|
2014-11-05 10:50:22 +00:00
|
|
|
/* The amount to copy out of the first guest Tx slot into the skb's
|
|
|
|
* linear area. If the first slot has more data, it will be mapped
|
|
|
|
* and put into the first frag.
|
|
|
|
*
|
|
|
|
* This is sized to avoid pulling headers from the frags for most
|
|
|
|
* TCP/IP packets.
|
|
|
|
*/
|
|
|
|
#define XEN_NETBACK_TX_COPY_LEN 128
|
|
|
|
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
|
2013-08-26 11:59:39 +00:00
|
|
|
u8 status);
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void make_tx_response(struct xenvif_queue *queue,
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_tx_request *txp,
|
|
|
|
s8 st);
|
2013-08-26 11:59:38 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline int tx_work_todo(struct xenvif_queue *queue);
|
2013-08-26 11:59:38 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
|
2011-03-15 00:06:18 +00:00
|
|
|
u16 id,
|
|
|
|
s8 st,
|
|
|
|
u16 offset,
|
|
|
|
u16 size,
|
|
|
|
u16 flags);
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
|
2011-10-05 00:28:46 +00:00
|
|
|
u16 idx)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
return page_to_pfn(queue->mmap_pages[idx]);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
|
2011-10-05 00:28:46 +00:00
|
|
|
u16 idx)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-03-24 23:59:51 +00:00
|
|
|
#define callback_param(vif, pending_idx) \
|
|
|
|
(vif->pending_tx_info[pending_idx].callback_struct)
|
|
|
|
|
2014-03-06 21:48:26 +00:00
|
|
|
/* Find the containing VIF's structure from a pointer in pending_tx_info array
|
|
|
|
*/
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
|
2014-03-06 21:48:25 +00:00
|
|
|
{
|
2014-03-06 21:48:26 +00:00
|
|
|
u16 pending_idx = ubuf->desc;
|
|
|
|
struct pending_tx_info *temp =
|
|
|
|
container_of(ubuf, struct pending_tx_info, callback_struct);
|
|
|
|
return container_of(temp - pending_idx,
|
2014-06-04 09:30:42 +00:00
|
|
|
struct xenvif_queue,
|
2014-03-06 21:48:26 +00:00
|
|
|
pending_tx_info[0]);
|
2014-03-06 21:48:25 +00:00
|
|
|
}
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2011-10-05 00:28:46 +00:00
|
|
|
static u16 frag_get_pending_idx(skb_frag_t *frag)
|
|
|
|
{
|
|
|
|
return (u16)frag->page_offset;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
|
|
|
|
{
|
|
|
|
frag->page_offset = pending_idx;
|
|
|
|
}
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
static inline pending_ring_idx_t pending_index(unsigned i)
|
|
|
|
{
|
|
|
|
return i & (MAX_PENDING_REQS-1);
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue, int needed)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
RING_IDX prod, cons;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
do {
|
2014-06-04 09:30:42 +00:00
|
|
|
prod = queue->rx.sring->req_prod;
|
|
|
|
cons = queue->rx.req_cons;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
if (prod - cons >= needed)
|
|
|
|
return true;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->rx.sring->req_event = prod + 1;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
/* Make sure event is visible before we check prod
|
|
|
|
* again.
|
|
|
|
*/
|
|
|
|
mb();
|
2014-06-04 09:30:42 +00:00
|
|
|
} while (queue->rx.sring->req_prod != prod);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
return false;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
|
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
spin_lock_irqsave(&queue->rx_queue.lock, flags);
|
|
|
|
|
|
|
|
__skb_queue_tail(&queue->rx_queue, skb);
|
|
|
|
|
|
|
|
queue->rx_queue_len += skb->len;
|
|
|
|
if (queue->rx_queue_len > queue->rx_queue_max)
|
|
|
|
netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
|
|
|
|
|
|
|
|
spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb;
|
|
|
|
|
|
|
|
spin_lock_irq(&queue->rx_queue.lock);
|
|
|
|
|
|
|
|
skb = __skb_dequeue(&queue->rx_queue);
|
|
|
|
if (skb)
|
|
|
|
queue->rx_queue_len -= skb->len;
|
|
|
|
|
|
|
|
spin_unlock_irq(&queue->rx_queue.lock);
|
|
|
|
|
|
|
|
return skb;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
|
|
|
|
{
|
|
|
|
spin_lock_irq(&queue->rx_queue.lock);
|
|
|
|
|
|
|
|
if (queue->rx_queue_len < queue->rx_queue_max)
|
|
|
|
netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
|
|
|
|
|
|
|
|
spin_unlock_irq(&queue->rx_queue.lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb;
|
|
|
|
while ((skb = xenvif_rx_dequeue(queue)) != NULL)
|
|
|
|
kfree_skb(skb);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb;
|
|
|
|
|
|
|
|
for(;;) {
|
|
|
|
skb = skb_peek(&queue->rx_queue);
|
|
|
|
if (!skb)
|
|
|
|
break;
|
|
|
|
if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
|
|
|
|
break;
|
|
|
|
xenvif_rx_dequeue(queue);
|
|
|
|
kfree_skb(skb);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
struct netrx_pending_operations {
|
|
|
|
unsigned copy_prod, copy_cons;
|
|
|
|
unsigned meta_prod, meta_cons;
|
|
|
|
struct gnttab_copy *copy;
|
2013-08-26 11:59:38 +00:00
|
|
|
struct xenvif_rx_meta *meta;
|
2011-03-15 00:06:18 +00:00
|
|
|
int copy_off;
|
|
|
|
grant_ref_t copy_gref;
|
|
|
|
};
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
|
2013-08-26 11:59:38 +00:00
|
|
|
struct netrx_pending_operations *npo)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2013-08-26 11:59:38 +00:00
|
|
|
struct xenvif_rx_meta *meta;
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_rx_request *req;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
meta = npo->meta + npo->meta_prod++;
|
2013-10-16 16:50:32 +00:00
|
|
|
meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
|
2011-03-15 00:06:18 +00:00
|
|
|
meta->gso_size = 0;
|
|
|
|
meta->size = 0;
|
|
|
|
meta->id = req->id;
|
|
|
|
|
|
|
|
npo->copy_off = 0;
|
|
|
|
npo->copy_gref = req->gref;
|
|
|
|
|
|
|
|
return meta;
|
|
|
|
}
|
|
|
|
|
2013-10-08 09:54:21 +00:00
|
|
|
/*
|
|
|
|
* Set up the grant operations for this fragment. If it's a flipping
|
|
|
|
* interface, we also set up the unmap request from here.
|
|
|
|
*/
|
2014-06-04 09:30:42 +00:00
|
|
|
static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
|
2013-08-26 11:59:39 +00:00
|
|
|
struct netrx_pending_operations *npo,
|
|
|
|
struct page *page, unsigned long size,
|
2015-01-05 14:45:10 +00:00
|
|
|
unsigned long offset, int *head)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
struct gnttab_copy *copy_gop;
|
2013-08-26 11:59:38 +00:00
|
|
|
struct xenvif_rx_meta *meta;
|
2011-03-15 00:06:18 +00:00
|
|
|
unsigned long bytes;
|
2014-03-10 14:58:34 +00:00
|
|
|
int gso_type = XEN_NETIF_GSO_TYPE_NONE;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Data must not cross a page boundary. */
|
2012-10-10 03:48:42 +00:00
|
|
|
BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
meta = npo->meta + npo->meta_prod - 1;
|
|
|
|
|
2012-10-10 03:48:42 +00:00
|
|
|
/* Skip unused frames from start of page */
|
|
|
|
page += offset >> PAGE_SHIFT;
|
|
|
|
offset &= ~PAGE_MASK;
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
while (size > 0) {
|
2015-01-05 14:45:10 +00:00
|
|
|
struct xen_page_foreign *foreign;
|
|
|
|
|
2012-10-10 03:48:42 +00:00
|
|
|
BUG_ON(offset >= PAGE_SIZE);
|
2011-03-15 00:06:18 +00:00
|
|
|
BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
|
|
|
|
|
2015-01-20 14:49:52 +00:00
|
|
|
if (npo->copy_off == MAX_BUFFER_OFFSET)
|
|
|
|
meta = get_next_rx_buffer(queue, npo);
|
2012-10-10 03:48:42 +00:00
|
|
|
|
2015-01-20 14:49:52 +00:00
|
|
|
bytes = PAGE_SIZE - offset;
|
2012-10-10 03:48:42 +00:00
|
|
|
if (bytes > size)
|
|
|
|
bytes = size;
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
|
|
|
|
bytes = MAX_BUFFER_OFFSET - npo->copy_off;
|
|
|
|
|
|
|
|
copy_gop = npo->copy + npo->copy_prod++;
|
|
|
|
copy_gop->flags = GNTCOPY_dest_gref;
|
2013-08-26 11:59:38 +00:00
|
|
|
copy_gop->len = bytes;
|
|
|
|
|
2015-01-05 14:45:10 +00:00
|
|
|
foreign = xen_page_foreign(page);
|
|
|
|
if (foreign) {
|
|
|
|
copy_gop->source.domid = foreign->domid;
|
|
|
|
copy_gop->source.u.ref = foreign->gref;
|
2014-03-06 21:48:25 +00:00
|
|
|
copy_gop->flags |= GNTCOPY_source_gref;
|
|
|
|
} else {
|
|
|
|
copy_gop->source.domid = DOMID_SELF;
|
|
|
|
copy_gop->source.u.gmfn =
|
|
|
|
virt_to_mfn(page_address(page));
|
|
|
|
}
|
2011-03-15 00:06:18 +00:00
|
|
|
copy_gop->source.offset = offset;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
copy_gop->dest.domid = queue->vif->domid;
|
2011-03-15 00:06:18 +00:00
|
|
|
copy_gop->dest.offset = npo->copy_off;
|
|
|
|
copy_gop->dest.u.ref = npo->copy_gref;
|
|
|
|
|
|
|
|
npo->copy_off += bytes;
|
|
|
|
meta->size += bytes;
|
|
|
|
|
|
|
|
offset += bytes;
|
|
|
|
size -= bytes;
|
|
|
|
|
2012-10-10 03:48:42 +00:00
|
|
|
/* Next frame */
|
|
|
|
if (offset == PAGE_SIZE && size) {
|
|
|
|
BUG_ON(!PageCompound(page));
|
|
|
|
page++;
|
|
|
|
offset = 0;
|
|
|
|
}
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
/* Leave a gap for the GSO descriptor. */
|
2014-03-10 14:58:34 +00:00
|
|
|
if (skb_is_gso(skb)) {
|
|
|
|
if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
|
|
|
|
gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
|
|
|
|
else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
|
|
|
|
gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
|
|
|
|
}
|
2013-10-16 16:50:32 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if (*head && ((1 << gso_type) & queue->vif->gso_mask))
|
|
|
|
queue->rx.req_cons++;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2013-10-08 09:54:21 +00:00
|
|
|
*head = 0; /* There must be something in this buffer now. */
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Prepare an SKB to be transmitted to the frontend.
|
|
|
|
*
|
|
|
|
* This function is responsible for allocating grant operations, meta
|
|
|
|
* structures, etc.
|
|
|
|
*
|
|
|
|
* It returns the number of meta structures consumed. The number of
|
|
|
|
* ring slots used is always equal to the number of meta slots used
|
|
|
|
* plus the number of GSO descriptors used. Currently, we use either
|
|
|
|
* zero GSO descriptors (for non-GSO packets) or one descriptor (for
|
|
|
|
* frontend-side LRO).
|
|
|
|
*/
|
2013-08-26 11:59:39 +00:00
|
|
|
static int xenvif_gop_skb(struct sk_buff *skb,
|
2014-06-04 09:30:42 +00:00
|
|
|
struct netrx_pending_operations *npo,
|
|
|
|
struct xenvif_queue *queue)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
struct xenvif *vif = netdev_priv(skb->dev);
|
|
|
|
int nr_frags = skb_shinfo(skb)->nr_frags;
|
|
|
|
int i;
|
|
|
|
struct xen_netif_rx_request *req;
|
2013-08-26 11:59:38 +00:00
|
|
|
struct xenvif_rx_meta *meta;
|
2011-03-15 00:06:18 +00:00
|
|
|
unsigned char *data;
|
2013-10-08 09:54:21 +00:00
|
|
|
int head = 1;
|
2011-03-15 00:06:18 +00:00
|
|
|
int old_meta_prod;
|
2013-10-16 16:50:32 +00:00
|
|
|
int gso_type;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
old_meta_prod = npo->meta_prod;
|
|
|
|
|
2014-03-10 14:58:34 +00:00
|
|
|
gso_type = XEN_NETIF_GSO_TYPE_NONE;
|
|
|
|
if (skb_is_gso(skb)) {
|
|
|
|
if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
|
|
|
|
gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
|
|
|
|
else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
|
|
|
|
gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
|
2013-10-16 16:50:32 +00:00
|
|
|
}
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
/* Set up a GSO prefix descriptor, if necessary */
|
2013-12-12 14:20:13 +00:00
|
|
|
if ((1 << gso_type) & vif->gso_prefix_mask) {
|
2014-06-04 09:30:42 +00:00
|
|
|
req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
|
2011-03-15 00:06:18 +00:00
|
|
|
meta = npo->meta + npo->meta_prod++;
|
2013-10-16 16:50:32 +00:00
|
|
|
meta->gso_type = gso_type;
|
2014-03-10 14:58:34 +00:00
|
|
|
meta->gso_size = skb_shinfo(skb)->gso_size;
|
2011-03-15 00:06:18 +00:00
|
|
|
meta->size = 0;
|
|
|
|
meta->id = req->id;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
|
2011-03-15 00:06:18 +00:00
|
|
|
meta = npo->meta + npo->meta_prod++;
|
|
|
|
|
2013-10-16 16:50:32 +00:00
|
|
|
if ((1 << gso_type) & vif->gso_mask) {
|
|
|
|
meta->gso_type = gso_type;
|
2014-03-10 14:58:34 +00:00
|
|
|
meta->gso_size = skb_shinfo(skb)->gso_size;
|
2013-10-16 16:50:32 +00:00
|
|
|
} else {
|
|
|
|
meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
|
2011-03-15 00:06:18 +00:00
|
|
|
meta->gso_size = 0;
|
2013-10-16 16:50:32 +00:00
|
|
|
}
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
meta->size = 0;
|
|
|
|
meta->id = req->id;
|
|
|
|
npo->copy_off = 0;
|
|
|
|
npo->copy_gref = req->gref;
|
|
|
|
|
|
|
|
data = skb->data;
|
|
|
|
while (data < skb_tail_pointer(skb)) {
|
|
|
|
unsigned int offset = offset_in_page(data);
|
|
|
|
unsigned int len = PAGE_SIZE - offset;
|
|
|
|
|
|
|
|
if (data + len > skb_tail_pointer(skb))
|
|
|
|
len = skb_tail_pointer(skb) - data;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_gop_frag_copy(queue, skb, npo,
|
2015-01-05 14:45:10 +00:00
|
|
|
virt_to_page(data), len, offset, &head);
|
2011-03-15 00:06:18 +00:00
|
|
|
data += len;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; i < nr_frags; i++) {
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_gop_frag_copy(queue, skb, npo,
|
2013-08-26 11:59:39 +00:00
|
|
|
skb_frag_page(&skb_shinfo(skb)->frags[i]),
|
|
|
|
skb_frag_size(&skb_shinfo(skb)->frags[i]),
|
|
|
|
skb_shinfo(skb)->frags[i].page_offset,
|
2015-01-05 14:45:10 +00:00
|
|
|
&head);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
return npo->meta_prod - old_meta_prod;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2013-08-26 11:59:39 +00:00
|
|
|
* This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
|
2011-03-15 00:06:18 +00:00
|
|
|
* used to set up the operations on the top of
|
|
|
|
* netrx_pending_operations, which have since been done. Check that
|
|
|
|
* they didn't give any errors and advance over them.
|
|
|
|
*/
|
2013-08-26 11:59:39 +00:00
|
|
|
static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
|
|
|
|
struct netrx_pending_operations *npo)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
struct gnttab_copy *copy_op;
|
|
|
|
int status = XEN_NETIF_RSP_OKAY;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < nr_meta_slots; i++) {
|
|
|
|
copy_op = npo->copy + npo->copy_cons++;
|
|
|
|
if (copy_op->status != GNTST_okay) {
|
|
|
|
netdev_dbg(vif->dev,
|
|
|
|
"Bad status %d from copy to DOM%d.\n",
|
|
|
|
copy_op->status, vif->domid);
|
|
|
|
status = XEN_NETIF_RSP_ERROR;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
|
2013-08-26 11:59:39 +00:00
|
|
|
struct xenvif_rx_meta *meta,
|
|
|
|
int nr_meta_slots)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
int i;
|
|
|
|
unsigned long offset;
|
|
|
|
|
|
|
|
/* No fragments used */
|
|
|
|
if (nr_meta_slots <= 1)
|
|
|
|
return;
|
|
|
|
|
|
|
|
nr_meta_slots--;
|
|
|
|
|
|
|
|
for (i = 0; i < nr_meta_slots; i++) {
|
|
|
|
int flags;
|
|
|
|
if (i == nr_meta_slots - 1)
|
|
|
|
flags = 0;
|
|
|
|
else
|
|
|
|
flags = XEN_NETRXF_more_data;
|
|
|
|
|
|
|
|
offset = 0;
|
2014-06-04 09:30:42 +00:00
|
|
|
make_rx_response(queue, meta[i].id, status, offset,
|
2011-03-15 00:06:18 +00:00
|
|
|
meta[i].size, flags);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
void xenvif_kick_thread(struct xenvif_queue *queue)
|
2013-08-26 11:59:38 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
wake_up(&queue->wq);
|
2013-08-26 11:59:38 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void xenvif_rx_action(struct xenvif_queue *queue)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
s8 status;
|
2013-05-22 06:34:45 +00:00
|
|
|
u16 flags;
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_rx_response *resp;
|
|
|
|
struct sk_buff_head rxq;
|
|
|
|
struct sk_buff *skb;
|
|
|
|
LIST_HEAD(notify);
|
|
|
|
int ret;
|
|
|
|
unsigned long offset;
|
2014-01-08 12:41:58 +00:00
|
|
|
bool need_to_notify = false;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
struct netrx_pending_operations npo = {
|
2014-06-04 09:30:42 +00:00
|
|
|
.copy = queue->grant_copy_op,
|
|
|
|
.meta = queue->meta,
|
2011-03-15 00:06:18 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
skb_queue_head_init(&rxq);
|
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
while (xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX)
|
|
|
|
&& (skb = xenvif_rx_dequeue(queue)) != NULL) {
|
2014-03-28 11:39:07 +00:00
|
|
|
RING_IDX old_req_cons;
|
|
|
|
RING_IDX ring_slots_used;
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
|
2014-10-22 13:08:55 +00:00
|
|
|
queue->last_rx_time = jiffies;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
old_req_cons = queue->rx.req_cons;
|
|
|
|
XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
|
|
|
|
ring_slots_used = queue->rx.req_cons - old_req_cons;
|
2014-03-28 11:39:07 +00:00
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
__skb_queue_tail(&rxq, skb);
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
if (!npo.copy_prod)
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
goto done;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2013-12-23 09:27:17 +00:00
|
|
|
BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
|
2014-06-04 09:30:42 +00:00
|
|
|
gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
while ((skb = __skb_dequeue(&rxq)) != NULL) {
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if ((1 << queue->meta[npo.meta_cons].gso_type) &
|
|
|
|
queue->vif->gso_prefix_mask) {
|
|
|
|
resp = RING_GET_RESPONSE(&queue->rx,
|
|
|
|
queue->rx.rsp_prod_pvt++);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
resp->offset = queue->meta[npo.meta_cons].gso_size;
|
|
|
|
resp->id = queue->meta[npo.meta_cons].id;
|
2014-03-06 21:48:23 +00:00
|
|
|
resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
npo.meta_cons++;
|
2014-03-06 21:48:23 +00:00
|
|
|
XENVIF_RX_CB(skb)->meta_slots_used--;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->stats.tx_bytes += skb->len;
|
|
|
|
queue->stats.tx_packets++;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
status = xenvif_check_gop(queue->vif,
|
2014-03-06 21:48:23 +00:00
|
|
|
XENVIF_RX_CB(skb)->meta_slots_used,
|
|
|
|
&npo);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-03-06 21:48:23 +00:00
|
|
|
if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
|
2011-03-15 00:06:18 +00:00
|
|
|
flags = 0;
|
|
|
|
else
|
|
|
|
flags = XEN_NETRXF_more_data;
|
|
|
|
|
|
|
|
if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
|
|
|
|
flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
|
|
|
|
else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
|
|
|
|
/* remote but checksummed. */
|
|
|
|
flags |= XEN_NETRXF_data_validated;
|
|
|
|
|
|
|
|
offset = 0;
|
2014-06-04 09:30:42 +00:00
|
|
|
resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
|
2011-03-15 00:06:18 +00:00
|
|
|
status, offset,
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->meta[npo.meta_cons].size,
|
2011-03-15 00:06:18 +00:00
|
|
|
flags);
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if ((1 << queue->meta[npo.meta_cons].gso_type) &
|
|
|
|
queue->vif->gso_mask) {
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_extra_info *gso =
|
|
|
|
(struct xen_netif_extra_info *)
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_GET_RESPONSE(&queue->rx,
|
|
|
|
queue->rx.rsp_prod_pvt++);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
resp->flags |= XEN_NETRXF_extra_info;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
|
|
|
|
gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
|
2011-03-15 00:06:18 +00:00
|
|
|
gso->u.gso.pad = 0;
|
|
|
|
gso->u.gso.features = 0;
|
|
|
|
|
|
|
|
gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
|
|
|
|
gso->flags = 0;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_add_frag_responses(queue, status,
|
|
|
|
queue->meta + npo.meta_cons + 1,
|
2014-03-06 21:48:23 +00:00
|
|
|
XENVIF_RX_CB(skb)->meta_slots_used);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-01-08 12:41:58 +00:00
|
|
|
need_to_notify |= !!ret;
|
2013-08-26 11:59:38 +00:00
|
|
|
|
2014-03-06 21:48:23 +00:00
|
|
|
npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
|
2011-03-15 00:06:18 +00:00
|
|
|
dev_kfree_skb(skb);
|
|
|
|
}
|
|
|
|
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
done:
|
2013-08-26 11:59:38 +00:00
|
|
|
if (need_to_notify)
|
2014-06-04 09:30:42 +00:00
|
|
|
notify_remote_via_irq(queue->rx_irq);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
int more_to_do;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
if (more_to_do)
|
2014-06-04 09:30:42 +00:00
|
|
|
napi_schedule(&queue->napi);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void tx_add_credit(struct xenvif_queue *queue)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
unsigned long max_burst, max_credit;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Allow a burst big enough to transmit a jumbo packet of up to 128kB.
|
|
|
|
* Otherwise the interface can seize up due to insufficient credit.
|
|
|
|
*/
|
2014-06-04 09:30:42 +00:00
|
|
|
max_burst = RING_GET_REQUEST(&queue->tx, queue->tx.req_cons)->size;
|
2011-03-15 00:06:18 +00:00
|
|
|
max_burst = min(max_burst, 131072UL);
|
2014-06-04 09:30:42 +00:00
|
|
|
max_burst = max(max_burst, queue->credit_bytes);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Take care that adding a new chunk of credit doesn't wrap to zero. */
|
2014-06-04 09:30:42 +00:00
|
|
|
max_credit = queue->remaining_credit + queue->credit_bytes;
|
|
|
|
if (max_credit < queue->remaining_credit)
|
2011-03-15 00:06:18 +00:00
|
|
|
max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->remaining_credit = min(max_credit, max_burst);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void tx_credit_callback(unsigned long data)
|
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
struct xenvif_queue *queue = (struct xenvif_queue *)data;
|
|
|
|
tx_add_credit(queue);
|
|
|
|
xenvif_napi_schedule_or_enable_events(queue);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void xenvif_tx_err(struct xenvif_queue *queue,
|
2013-08-26 11:59:39 +00:00
|
|
|
struct xen_netif_tx_request *txp, RING_IDX end)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_IDX cons = queue->tx.req_cons;
|
2014-03-06 21:48:26 +00:00
|
|
|
unsigned long flags;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
do {
|
2014-06-04 09:30:42 +00:00
|
|
|
spin_lock_irqsave(&queue->response_lock, flags);
|
|
|
|
make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
|
|
|
|
spin_unlock_irqrestore(&queue->response_lock, flags);
|
2013-02-06 23:41:38 +00:00
|
|
|
if (cons == end)
|
2011-03-15 00:06:18 +00:00
|
|
|
break;
|
2014-06-04 09:30:42 +00:00
|
|
|
txp = RING_GET_REQUEST(&queue->tx, cons++);
|
2011-03-15 00:06:18 +00:00
|
|
|
} while (1);
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx.req_cons = cons;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2013-08-26 11:59:39 +00:00
|
|
|
static void xenvif_fatal_tx_err(struct xenvif *vif)
|
2013-02-06 23:41:35 +00:00
|
|
|
{
|
|
|
|
netdev_err(vif->dev, "fatal error; disabling device\n");
|
2014-04-01 11:46:12 +00:00
|
|
|
vif->disabled = true;
|
2014-06-04 09:30:42 +00:00
|
|
|
/* Disable the vif from queue 0's kthread */
|
|
|
|
if (vif->queues)
|
|
|
|
xenvif_kick_thread(&vif->queues[0]);
|
2013-02-06 23:41:35 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static int xenvif_count_requests(struct xenvif_queue *queue,
|
2013-08-26 11:59:39 +00:00
|
|
|
struct xen_netif_tx_request *first,
|
|
|
|
struct xen_netif_tx_request *txp,
|
|
|
|
int work_to_do)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_IDX cons = queue->tx.req_cons;
|
2013-04-22 02:20:42 +00:00
|
|
|
int slots = 0;
|
|
|
|
int drop_err = 0;
|
2013-05-02 00:43:58 +00:00
|
|
|
int more_data;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
if (!(first->flags & XEN_NETTXF_more_data))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
do {
|
2013-05-02 00:43:58 +00:00
|
|
|
struct xen_netif_tx_request dropped_tx = { 0 };
|
|
|
|
|
2013-04-22 02:20:42 +00:00
|
|
|
if (slots >= work_to_do) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2013-04-22 02:20:42 +00:00
|
|
|
"Asked for %d slots but exceeds this limit\n",
|
|
|
|
work_to_do);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_fatal_tx_err(queue->vif);
|
2013-02-14 03:18:57 +00:00
|
|
|
return -ENODATA;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2013-04-22 02:20:42 +00:00
|
|
|
/* This guest is really using too many slots and
|
|
|
|
* considered malicious.
|
|
|
|
*/
|
2013-05-02 00:43:59 +00:00
|
|
|
if (unlikely(slots >= fatal_skb_slots)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2013-04-22 02:20:42 +00:00
|
|
|
"Malicious frontend using %d slots, threshold %u\n",
|
2013-05-02 00:43:59 +00:00
|
|
|
slots, fatal_skb_slots);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_fatal_tx_err(queue->vif);
|
2013-02-14 03:18:57 +00:00
|
|
|
return -E2BIG;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2013-04-22 02:20:42 +00:00
|
|
|
/* Xen network protocol had implicit dependency on
|
2013-05-02 00:43:59 +00:00
|
|
|
* MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
|
|
|
|
* the historical MAX_SKB_FRAGS value 18 to honor the
|
|
|
|
* same behavior as before. Any packet using more than
|
|
|
|
* 18 slots but less than fatal_skb_slots slots is
|
|
|
|
* dropped
|
2013-04-22 02:20:42 +00:00
|
|
|
*/
|
2013-05-02 00:43:59 +00:00
|
|
|
if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
|
2013-04-22 02:20:42 +00:00
|
|
|
if (net_ratelimit())
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_dbg(queue->vif->dev,
|
2013-04-22 02:20:42 +00:00
|
|
|
"Too many slots (%d) exceeding limit (%d), dropping packet\n",
|
2013-05-02 00:43:59 +00:00
|
|
|
slots, XEN_NETBK_LEGACY_SLOTS_MAX);
|
2013-04-22 02:20:42 +00:00
|
|
|
drop_err = -E2BIG;
|
|
|
|
}
|
|
|
|
|
2013-05-02 00:43:58 +00:00
|
|
|
if (drop_err)
|
|
|
|
txp = &dropped_tx;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
memcpy(txp, RING_GET_REQUEST(&queue->tx, cons + slots),
|
2011-03-15 00:06:18 +00:00
|
|
|
sizeof(*txp));
|
2013-04-22 02:20:43 +00:00
|
|
|
|
|
|
|
/* If the guest submitted a frame >= 64 KiB then
|
|
|
|
* first->size overflowed and following slots will
|
|
|
|
* appear to be larger than the frame.
|
|
|
|
*
|
|
|
|
* This cannot be fatal error as there are buggy
|
|
|
|
* frontends that do this.
|
|
|
|
*
|
|
|
|
* Consume all slots and drop the packet.
|
|
|
|
*/
|
|
|
|
if (!drop_err && txp->size > first->size) {
|
|
|
|
if (net_ratelimit())
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_dbg(queue->vif->dev,
|
2013-04-22 02:20:43 +00:00
|
|
|
"Invalid tx request, slot size %u > remaining size %u\n",
|
|
|
|
txp->size, first->size);
|
|
|
|
drop_err = -EIO;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
first->size -= txp->size;
|
2013-04-22 02:20:42 +00:00
|
|
|
slots++;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
|
2011-03-15 00:06:18 +00:00
|
|
|
txp->offset, txp->size);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_fatal_tx_err(queue->vif);
|
2013-02-14 03:18:57 +00:00
|
|
|
return -EINVAL;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
2013-05-02 00:43:58 +00:00
|
|
|
|
|
|
|
more_data = txp->flags & XEN_NETTXF_more_data;
|
|
|
|
|
|
|
|
if (!drop_err)
|
|
|
|
txp++;
|
|
|
|
|
|
|
|
} while (more_data);
|
2013-04-22 02:20:42 +00:00
|
|
|
|
|
|
|
if (drop_err) {
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_tx_err(queue, first, cons + slots);
|
2013-04-22 02:20:42 +00:00
|
|
|
return drop_err;
|
|
|
|
}
|
|
|
|
|
|
|
|
return slots;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:23 +00:00
|
|
|
|
|
|
|
struct xenvif_tx_cb {
|
|
|
|
u16 pending_idx;
|
|
|
|
};
|
|
|
|
|
|
|
|
#define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
|
2014-04-02 17:04:57 +00:00
|
|
|
u16 pending_idx,
|
|
|
|
struct xen_netif_tx_request *txp,
|
|
|
|
struct gnttab_map_grant_ref *mop)
|
2014-03-06 21:48:26 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
|
|
|
|
gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
|
2014-03-06 21:48:26 +00:00
|
|
|
GNTMAP_host_map | GNTMAP_readonly,
|
2014-06-04 09:30:42 +00:00
|
|
|
txp->gref, queue->vif->domid);
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
memcpy(&queue->pending_tx_info[pending_idx].req, txp,
|
2014-03-06 21:48:26 +00:00
|
|
|
sizeof(*txp));
|
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:29 +00:00
|
|
|
static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb =
|
|
|
|
alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
|
|
|
|
GFP_ATOMIC | __GFP_NOWARN);
|
|
|
|
if (unlikely(skb == NULL))
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
/* Packets passed to netif_rx() must have some headroom. */
|
|
|
|
skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
|
|
|
|
|
|
|
|
/* Initialize it here to avoid later surprises */
|
|
|
|
skb_shinfo(skb)->destructor_arg = NULL;
|
|
|
|
|
|
|
|
return skb;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
|
2014-03-06 21:48:26 +00:00
|
|
|
struct sk_buff *skb,
|
|
|
|
struct xen_netif_tx_request *txp,
|
|
|
|
struct gnttab_map_grant_ref *gop)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
struct skb_shared_info *shinfo = skb_shinfo(skb);
|
|
|
|
skb_frag_t *frags = shinfo->frags;
|
2014-03-06 21:48:23 +00:00
|
|
|
u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
|
2014-03-06 21:48:27 +00:00
|
|
|
int start;
|
|
|
|
pending_ring_idx_t index;
|
2014-03-06 21:48:29 +00:00
|
|
|
unsigned int nr_slots, frag_overflow = 0;
|
2013-04-22 02:20:42 +00:00
|
|
|
|
|
|
|
/* At this point shinfo->nr_frags is in fact the number of
|
2013-05-02 00:43:59 +00:00
|
|
|
* slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
|
2013-04-22 02:20:42 +00:00
|
|
|
*/
|
2014-03-06 21:48:29 +00:00
|
|
|
if (shinfo->nr_frags > MAX_SKB_FRAGS) {
|
|
|
|
frag_overflow = shinfo->nr_frags - MAX_SKB_FRAGS;
|
|
|
|
BUG_ON(frag_overflow > MAX_SKB_FRAGS);
|
|
|
|
shinfo->nr_frags = MAX_SKB_FRAGS;
|
|
|
|
}
|
2013-04-22 02:20:42 +00:00
|
|
|
nr_slots = shinfo->nr_frags;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Skip first skb fragment if it is on same page as header fragment. */
|
2011-10-05 00:28:46 +00:00
|
|
|
start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-03-06 21:48:26 +00:00
|
|
|
for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
|
|
|
|
shinfo->nr_frags++, txp++, gop++) {
|
2014-06-04 09:30:42 +00:00
|
|
|
index = pending_index(queue->pending_cons++);
|
|
|
|
pending_idx = queue->pending_ring[index];
|
|
|
|
xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
|
2014-03-06 21:48:26 +00:00
|
|
|
frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:29 +00:00
|
|
|
if (frag_overflow) {
|
|
|
|
struct sk_buff *nskb = xenvif_alloc_skb(0);
|
|
|
|
if (unlikely(nskb == NULL)) {
|
|
|
|
if (net_ratelimit())
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2014-03-06 21:48:29 +00:00
|
|
|
"Can't allocate the frag_list skb.\n");
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
shinfo = skb_shinfo(nskb);
|
|
|
|
frags = shinfo->frags;
|
|
|
|
|
|
|
|
for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
|
|
|
|
shinfo->nr_frags++, txp++, gop++) {
|
2014-06-04 09:30:42 +00:00
|
|
|
index = pending_index(queue->pending_cons++);
|
|
|
|
pending_idx = queue->pending_ring[index];
|
|
|
|
xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
|
2014-03-06 21:48:29 +00:00
|
|
|
frag_set_pending_idx(&frags[shinfo->nr_frags],
|
|
|
|
pending_idx);
|
|
|
|
}
|
|
|
|
|
|
|
|
skb_shinfo(skb)->frag_list = nskb;
|
|
|
|
}
|
2013-04-22 02:20:42 +00:00
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
return gop;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
|
2014-03-06 21:48:26 +00:00
|
|
|
u16 pending_idx,
|
|
|
|
grant_handle_t handle)
|
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
if (unlikely(queue->grant_tx_handle[pending_idx] !=
|
2014-03-06 21:48:26 +00:00
|
|
|
NETBACK_INVALID_HANDLE)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2014-03-06 21:48:26 +00:00
|
|
|
"Trying to overwrite active handle! pending_idx: %x\n",
|
|
|
|
pending_idx);
|
|
|
|
BUG();
|
|
|
|
}
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->grant_tx_handle[pending_idx] = handle;
|
2014-03-06 21:48:26 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
|
2014-03-06 21:48:26 +00:00
|
|
|
u16 pending_idx)
|
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
if (unlikely(queue->grant_tx_handle[pending_idx] ==
|
2014-03-06 21:48:26 +00:00
|
|
|
NETBACK_INVALID_HANDLE)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2014-03-06 21:48:26 +00:00
|
|
|
"Trying to unmap invalid handle! pending_idx: %x\n",
|
|
|
|
pending_idx);
|
|
|
|
BUG();
|
|
|
|
}
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
|
2014-03-06 21:48:26 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static int xenvif_tx_check_gop(struct xenvif_queue *queue,
|
2013-08-26 11:59:39 +00:00
|
|
|
struct sk_buff *skb,
|
2014-04-02 17:04:58 +00:00
|
|
|
struct gnttab_map_grant_ref **gopp_map,
|
|
|
|
struct gnttab_copy **gopp_copy)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-04-02 17:04:57 +00:00
|
|
|
struct gnttab_map_grant_ref *gop_map = *gopp_map;
|
2014-03-06 21:48:23 +00:00
|
|
|
u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
|
2014-07-18 18:08:02 +00:00
|
|
|
/* This always points to the shinfo of the skb being checked, which
|
|
|
|
* could be either the first or the one on the frag_list
|
|
|
|
*/
|
2011-03-15 00:06:18 +00:00
|
|
|
struct skb_shared_info *shinfo = skb_shinfo(skb);
|
2014-07-18 18:08:02 +00:00
|
|
|
/* If this is non-NULL, we are currently checking the frag_list skb, and
|
|
|
|
* this points to the shinfo of the first one
|
|
|
|
*/
|
|
|
|
struct skb_shared_info *first_shinfo = NULL;
|
2011-03-15 00:06:18 +00:00
|
|
|
int nr_frags = shinfo->nr_frags;
|
2014-07-18 18:08:04 +00:00
|
|
|
const bool sharedslot = nr_frags &&
|
|
|
|
frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
|
2014-04-02 17:04:58 +00:00
|
|
|
int i, err;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Check status of header. */
|
2014-04-02 17:04:58 +00:00
|
|
|
err = (*gopp_copy)->status;
|
|
|
|
if (unlikely(err)) {
|
|
|
|
if (net_ratelimit())
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_dbg(queue->vif->dev,
|
2014-04-04 14:45:24 +00:00
|
|
|
"Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
|
2014-04-02 17:04:58 +00:00
|
|
|
(*gopp_copy)->status,
|
|
|
|
pending_idx,
|
|
|
|
(*gopp_copy)->source.u.ref);
|
2014-07-18 18:08:04 +00:00
|
|
|
/* The first frag might still have this slot mapped */
|
|
|
|
if (!sharedslot)
|
|
|
|
xenvif_idx_release(queue, pending_idx,
|
|
|
|
XEN_NETIF_RSP_ERROR);
|
2014-04-02 17:04:58 +00:00
|
|
|
}
|
2014-07-18 18:08:05 +00:00
|
|
|
(*gopp_copy)++;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-03-06 21:48:29 +00:00
|
|
|
check_frags:
|
2014-04-02 17:04:58 +00:00
|
|
|
for (i = 0; i < nr_frags; i++, gop_map++) {
|
2011-03-15 00:06:18 +00:00
|
|
|
int j, newerr;
|
|
|
|
|
2011-10-05 00:28:46 +00:00
|
|
|
pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Check error status: if okay then remember grant handle. */
|
2014-04-02 17:04:58 +00:00
|
|
|
newerr = gop_map->status;
|
2013-04-22 02:20:42 +00:00
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
if (likely(!newerr)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_grant_handle_set(queue,
|
2014-04-02 17:04:57 +00:00
|
|
|
pending_idx,
|
|
|
|
gop_map->handle);
|
2011-03-15 00:06:18 +00:00
|
|
|
/* Had a previous error? Invalidate this fragment. */
|
2014-07-18 18:08:04 +00:00
|
|
|
if (unlikely(err)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_idx_unmap(queue, pending_idx);
|
2014-07-18 18:08:04 +00:00
|
|
|
/* If the mapping of the first frag was OK, but
|
|
|
|
* the header's copy failed, and they are
|
|
|
|
* sharing a slot, send an error
|
|
|
|
*/
|
|
|
|
if (i == 0 && sharedslot)
|
|
|
|
xenvif_idx_release(queue, pending_idx,
|
|
|
|
XEN_NETIF_RSP_ERROR);
|
|
|
|
else
|
|
|
|
xenvif_idx_release(queue, pending_idx,
|
|
|
|
XEN_NETIF_RSP_OKAY);
|
|
|
|
}
|
2011-03-15 00:06:18 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Error on this fragment: respond to client with an error. */
|
2014-04-02 17:04:58 +00:00
|
|
|
if (net_ratelimit())
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_dbg(queue->vif->dev,
|
2014-04-04 14:45:24 +00:00
|
|
|
"Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
|
2014-04-02 17:04:58 +00:00
|
|
|
i,
|
|
|
|
gop_map->status,
|
|
|
|
pending_idx,
|
|
|
|
gop_map->ref);
|
2014-07-18 18:08:04 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Not the first error? Preceding frags already invalidated. */
|
|
|
|
if (err)
|
|
|
|
continue;
|
2014-07-18 18:08:04 +00:00
|
|
|
|
|
|
|
/* First error: if the header haven't shared a slot with the
|
|
|
|
* first frag, release it as well.
|
|
|
|
*/
|
|
|
|
if (!sharedslot)
|
|
|
|
xenvif_idx_release(queue,
|
|
|
|
XENVIF_TX_CB(skb)->pending_idx,
|
|
|
|
XEN_NETIF_RSP_OKAY);
|
|
|
|
|
|
|
|
/* Invalidate preceding fragments of this skb. */
|
2014-04-02 17:04:58 +00:00
|
|
|
for (j = 0; j < i; j++) {
|
2011-11-18 05:42:05 +00:00
|
|
|
pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_idx_unmap(queue, pending_idx);
|
2014-07-18 18:08:04 +00:00
|
|
|
xenvif_idx_release(queue, pending_idx,
|
|
|
|
XEN_NETIF_RSP_OKAY);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-07-18 18:08:02 +00:00
|
|
|
/* And if we found the error while checking the frag_list, unmap
|
|
|
|
* the first skb's frags
|
|
|
|
*/
|
|
|
|
if (first_shinfo) {
|
|
|
|
for (j = 0; j < first_shinfo->nr_frags; j++) {
|
|
|
|
pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
|
|
|
|
xenvif_idx_unmap(queue, pending_idx);
|
2014-07-18 18:08:04 +00:00
|
|
|
xenvif_idx_release(queue, pending_idx,
|
|
|
|
XEN_NETIF_RSP_OKAY);
|
2014-07-18 18:08:02 +00:00
|
|
|
}
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Remember the error: invalidate all subsequent fragments. */
|
|
|
|
err = newerr;
|
|
|
|
}
|
|
|
|
|
2014-07-18 18:08:02 +00:00
|
|
|
if (skb_has_frag_list(skb) && !first_shinfo) {
|
|
|
|
first_shinfo = skb_shinfo(skb);
|
|
|
|
shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
|
2014-03-06 21:48:29 +00:00
|
|
|
nr_frags = shinfo->nr_frags;
|
|
|
|
|
|
|
|
goto check_frags;
|
|
|
|
}
|
|
|
|
|
2014-04-02 17:04:58 +00:00
|
|
|
*gopp_map = gop_map;
|
2011-03-15 00:06:18 +00:00
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
struct skb_shared_info *shinfo = skb_shinfo(skb);
|
|
|
|
int nr_frags = shinfo->nr_frags;
|
|
|
|
int i;
|
2014-03-06 21:48:26 +00:00
|
|
|
u16 prev_pending_idx = INVALID_PENDING_IDX;
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
for (i = 0; i < nr_frags; i++) {
|
|
|
|
skb_frag_t *frag = shinfo->frags + i;
|
|
|
|
struct xen_netif_tx_request *txp;
|
2011-10-05 00:28:46 +00:00
|
|
|
struct page *page;
|
|
|
|
u16 pending_idx;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2011-10-05 00:28:46 +00:00
|
|
|
pending_idx = frag_get_pending_idx(frag);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-03-06 21:48:26 +00:00
|
|
|
/* If this is not the first frag, chain it to the previous*/
|
2014-04-02 17:04:58 +00:00
|
|
|
if (prev_pending_idx == INVALID_PENDING_IDX)
|
2014-03-06 21:48:26 +00:00
|
|
|
skb_shinfo(skb)->destructor_arg =
|
2014-06-04 09:30:42 +00:00
|
|
|
&callback_param(queue, pending_idx);
|
2014-04-02 17:04:58 +00:00
|
|
|
else
|
2014-06-04 09:30:42 +00:00
|
|
|
callback_param(queue, prev_pending_idx).ctx =
|
|
|
|
&callback_param(queue, pending_idx);
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
callback_param(queue, pending_idx).ctx = NULL;
|
2014-03-06 21:48:26 +00:00
|
|
|
prev_pending_idx = pending_idx;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
txp = &queue->pending_tx_info[pending_idx].req;
|
|
|
|
page = virt_to_page(idx_to_kaddr(queue, pending_idx));
|
2011-10-05 00:28:46 +00:00
|
|
|
__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
|
2011-03-15 00:06:18 +00:00
|
|
|
skb->len += txp->size;
|
|
|
|
skb->data_len += txp->size;
|
|
|
|
skb->truesize += txp->size;
|
|
|
|
|
2014-03-06 21:48:26 +00:00
|
|
|
/* Take an extra reference to offset network stack's put_page */
|
2014-06-04 09:30:42 +00:00
|
|
|
get_page(queue->mmap_pages[pending_idx]);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static int xenvif_get_extras(struct xenvif_queue *queue,
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_extra_info *extras,
|
|
|
|
int work_to_do)
|
|
|
|
{
|
|
|
|
struct xen_netif_extra_info extra;
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_IDX cons = queue->tx.req_cons;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
do {
|
|
|
|
if (unlikely(work_to_do-- <= 0)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev, "Missing extra info\n");
|
|
|
|
xenvif_fatal_tx_err(queue->vif);
|
2011-03-15 00:06:18 +00:00
|
|
|
return -EBADR;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
memcpy(&extra, RING_GET_REQUEST(&queue->tx, cons),
|
2011-03-15 00:06:18 +00:00
|
|
|
sizeof(extra));
|
|
|
|
if (unlikely(!extra.type ||
|
|
|
|
extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx.req_cons = ++cons;
|
|
|
|
netdev_err(queue->vif->dev,
|
2011-03-15 00:06:18 +00:00
|
|
|
"Invalid extra type: %d\n", extra.type);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_fatal_tx_err(queue->vif);
|
2011-03-15 00:06:18 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx.req_cons = ++cons;
|
2011-03-15 00:06:18 +00:00
|
|
|
} while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
|
|
|
|
|
|
|
|
return work_to_do;
|
|
|
|
}
|
|
|
|
|
2013-08-26 11:59:39 +00:00
|
|
|
static int xenvif_set_skb_gso(struct xenvif *vif,
|
|
|
|
struct sk_buff *skb,
|
|
|
|
struct xen_netif_extra_info *gso)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
if (!gso->u.gso.size) {
|
2013-02-06 23:41:35 +00:00
|
|
|
netdev_err(vif->dev, "GSO size must not be zero.\n");
|
2013-08-26 11:59:39 +00:00
|
|
|
xenvif_fatal_tx_err(vif);
|
2011-03-15 00:06:18 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2013-10-16 16:50:31 +00:00
|
|
|
switch (gso->u.gso.type) {
|
|
|
|
case XEN_NETIF_GSO_TYPE_TCPV4:
|
|
|
|
skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
|
|
|
|
break;
|
|
|
|
case XEN_NETIF_GSO_TYPE_TCPV6:
|
|
|
|
skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
|
|
|
|
break;
|
|
|
|
default:
|
2013-02-06 23:41:35 +00:00
|
|
|
netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
|
2013-08-26 11:59:39 +00:00
|
|
|
xenvif_fatal_tx_err(vif);
|
2011-03-15 00:06:18 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
skb_shinfo(skb)->gso_size = gso->u.gso.size;
|
2013-12-17 11:44:35 +00:00
|
|
|
/* gso_segs will be calculated later */
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
|
2013-10-16 16:50:29 +00:00
|
|
|
{
|
2014-01-09 10:02:47 +00:00
|
|
|
bool recalculate_partial_csum = false;
|
2013-10-16 16:50:29 +00:00
|
|
|
|
|
|
|
/* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
|
|
|
|
* peers can fail to set NETRXF_csum_blank when sending a GSO
|
|
|
|
* frame. In this case force the SKB to CHECKSUM_PARTIAL and
|
|
|
|
* recalculate the partial checksum.
|
|
|
|
*/
|
|
|
|
if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->stats.rx_gso_checksum_fixup++;
|
2013-10-16 16:50:29 +00:00
|
|
|
skb->ip_summed = CHECKSUM_PARTIAL;
|
2014-01-09 10:02:47 +00:00
|
|
|
recalculate_partial_csum = true;
|
2013-10-16 16:50:29 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
|
|
|
|
if (skb->ip_summed != CHECKSUM_PARTIAL)
|
|
|
|
return 0;
|
|
|
|
|
2014-01-09 10:02:47 +00:00
|
|
|
return skb_checksum_setup(skb, recalculate_partial_csum);
|
2013-10-16 16:50:29 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2013-10-28 12:07:57 +00:00
|
|
|
u64 now = get_jiffies_64();
|
2014-06-04 09:30:42 +00:00
|
|
|
u64 next_credit = queue->credit_window_start +
|
|
|
|
msecs_to_jiffies(queue->credit_usec / 1000);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Timer could already be pending in rare cases. */
|
2014-06-04 09:30:42 +00:00
|
|
|
if (timer_pending(&queue->credit_timeout))
|
2011-03-15 00:06:18 +00:00
|
|
|
return true;
|
|
|
|
|
|
|
|
/* Passed the point where we can replenish credit? */
|
2013-10-28 12:07:57 +00:00
|
|
|
if (time_after_eq64(now, next_credit)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->credit_window_start = now;
|
|
|
|
tx_add_credit(queue);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Still too big to send right now? Set a callback. */
|
2014-06-04 09:30:42 +00:00
|
|
|
if (size > queue->remaining_credit) {
|
|
|
|
queue->credit_timeout.data =
|
|
|
|
(unsigned long)queue;
|
|
|
|
queue->credit_timeout.function =
|
2011-03-15 00:06:18 +00:00
|
|
|
tx_credit_callback;
|
2014-06-04 09:30:42 +00:00
|
|
|
mod_timer(&queue->credit_timeout,
|
2011-03-15 00:06:18 +00:00
|
|
|
next_credit);
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->credit_window_start = next_credit;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void xenvif_tx_build_gops(struct xenvif_queue *queue,
|
2014-04-02 17:04:58 +00:00
|
|
|
int budget,
|
|
|
|
unsigned *copy_ops,
|
|
|
|
unsigned *map_ops)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
struct gnttab_map_grant_ref *gop = queue->tx_map_ops, *request_gop;
|
2011-03-15 00:06:18 +00:00
|
|
|
struct sk_buff *skb;
|
|
|
|
int ret;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
while (skb_queue_len(&queue->tx_queue) < budget) {
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_tx_request txreq;
|
2013-05-02 00:43:59 +00:00
|
|
|
struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
|
|
|
|
u16 pending_idx;
|
|
|
|
RING_IDX idx;
|
|
|
|
int work_to_do;
|
|
|
|
unsigned int data_len;
|
|
|
|
pending_ring_idx_t index;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if (queue->tx.sring->req_prod - queue->tx.req_cons >
|
2013-02-06 23:41:35 +00:00
|
|
|
XEN_NETIF_TX_RING_SIZE) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2013-02-06 23:41:35 +00:00
|
|
|
"Impossible number of requests. "
|
|
|
|
"req_prod %d, req_cons %d, size %ld\n",
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx.sring->req_prod, queue->tx.req_cons,
|
2013-02-06 23:41:35 +00:00
|
|
|
XEN_NETIF_TX_RING_SIZE);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_fatal_tx_err(queue->vif);
|
2014-04-01 11:46:12 +00:00
|
|
|
break;
|
2013-02-06 23:41:35 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
|
2013-08-26 11:59:38 +00:00
|
|
|
if (!work_to_do)
|
|
|
|
break;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
idx = queue->tx.req_cons;
|
2011-03-15 00:06:18 +00:00
|
|
|
rmb(); /* Ensure that we see the request before we copy it. */
|
2014-06-04 09:30:42 +00:00
|
|
|
memcpy(&txreq, RING_GET_REQUEST(&queue->tx, idx), sizeof(txreq));
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Credit-based scheduling. */
|
2014-06-04 09:30:42 +00:00
|
|
|
if (txreq.size > queue->remaining_credit &&
|
|
|
|
tx_credit_exceeded(queue, txreq.size))
|
2013-08-26 11:59:38 +00:00
|
|
|
break;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->remaining_credit -= txreq.size;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
work_to_do--;
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx.req_cons = ++idx;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
memset(extras, 0, sizeof(extras));
|
|
|
|
if (txreq.flags & XEN_NETTXF_extra_info) {
|
2014-06-04 09:30:42 +00:00
|
|
|
work_to_do = xenvif_get_extras(queue, extras,
|
2013-08-26 11:59:39 +00:00
|
|
|
work_to_do);
|
2014-06-04 09:30:42 +00:00
|
|
|
idx = queue->tx.req_cons;
|
2013-02-06 23:41:35 +00:00
|
|
|
if (unlikely(work_to_do < 0))
|
2013-08-26 11:59:38 +00:00
|
|
|
break;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
|
2013-02-06 23:41:35 +00:00
|
|
|
if (unlikely(ret < 0))
|
2013-08-26 11:59:38 +00:00
|
|
|
break;
|
2013-02-06 23:41:35 +00:00
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
idx += ret;
|
|
|
|
|
|
|
|
if (unlikely(txreq.size < ETH_HLEN)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_dbg(queue->vif->dev,
|
2011-03-15 00:06:18 +00:00
|
|
|
"Bad packet size: %d\n", txreq.size);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_tx_err(queue, &txreq, idx);
|
2013-08-26 11:59:38 +00:00
|
|
|
break;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* No crossing a page as the payload mustn't fragment. */
|
|
|
|
if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2011-03-15 00:06:18 +00:00
|
|
|
"txreq.offset: %x, size: %u, end: %lu\n",
|
|
|
|
txreq.offset, txreq.size,
|
|
|
|
(txreq.offset&~PAGE_MASK) + txreq.size);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_fatal_tx_err(queue->vif);
|
2013-08-26 11:59:38 +00:00
|
|
|
break;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
index = pending_index(queue->pending_cons);
|
|
|
|
pending_idx = queue->pending_ring[index];
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-11-05 10:50:22 +00:00
|
|
|
data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
|
2013-05-02 00:43:59 +00:00
|
|
|
ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
|
2014-11-05 10:50:22 +00:00
|
|
|
XEN_NETBACK_TX_COPY_LEN : txreq.size;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-03-06 21:48:29 +00:00
|
|
|
skb = xenvif_alloc_skb(data_len);
|
2011-03-15 00:06:18 +00:00
|
|
|
if (unlikely(skb == NULL)) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_dbg(queue->vif->dev,
|
2011-03-15 00:06:18 +00:00
|
|
|
"Can't allocate a skb in start_xmit.\n");
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_tx_err(queue, &txreq, idx);
|
2011-03-15 00:06:18 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
|
|
|
|
struct xen_netif_extra_info *gso;
|
|
|
|
gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
|
2013-08-26 11:59:39 +00:00
|
|
|
/* Failure in xenvif_set_skb_gso is fatal. */
|
2011-03-15 00:06:18 +00:00
|
|
|
kfree_skb(skb);
|
2013-08-26 11:59:38 +00:00
|
|
|
break;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:23 +00:00
|
|
|
XENVIF_TX_CB(skb)->pending_idx = pending_idx;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
__skb_put(skb, data_len);
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
|
|
|
|
queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
|
|
|
|
queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
|
2014-04-02 17:04:58 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
|
2014-04-02 17:04:58 +00:00
|
|
|
virt_to_mfn(skb->data);
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
|
|
|
|
queue->tx_copy_ops[*copy_ops].dest.offset =
|
2014-04-02 17:04:58 +00:00
|
|
|
offset_in_page(skb->data);
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx_copy_ops[*copy_ops].len = data_len;
|
|
|
|
queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
|
2014-04-02 17:04:58 +00:00
|
|
|
|
|
|
|
(*copy_ops)++;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
skb_shinfo(skb)->nr_frags = ret;
|
|
|
|
if (data_len < txreq.size) {
|
|
|
|
skb_shinfo(skb)->nr_frags++;
|
2011-10-05 00:28:46 +00:00
|
|
|
frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
|
|
|
|
pending_idx);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
|
2014-04-02 17:04:58 +00:00
|
|
|
gop++;
|
2011-03-15 00:06:18 +00:00
|
|
|
} else {
|
2011-10-05 00:28:46 +00:00
|
|
|
frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
|
|
|
|
INVALID_PENDING_IDX);
|
2014-06-04 09:30:42 +00:00
|
|
|
memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
|
2014-04-02 17:04:58 +00:00
|
|
|
sizeof(txreq));
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->pending_cons++;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
request_gop = xenvif_get_requests(queue, skb, txfrags, gop);
|
2011-03-15 00:06:18 +00:00
|
|
|
if (request_gop == NULL) {
|
|
|
|
kfree_skb(skb);
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_tx_err(queue, &txreq, idx);
|
2013-08-26 11:59:38 +00:00
|
|
|
break;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
gop = request_gop;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
__skb_queue_tail(&queue->tx_queue, skb);
|
2012-06-27 00:46:58 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx.req_cons = idx;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
|
|
|
|
(*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
|
2011-03-15 00:06:18 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
(*map_ops) = gop - queue->tx_map_ops;
|
2014-04-02 17:04:58 +00:00
|
|
|
return;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:29 +00:00
|
|
|
/* Consolidate skb with a frag_list into a brand new one with local pages on
|
|
|
|
* frags. Returns 0 or -ENOMEM if can't allocate new pages.
|
|
|
|
*/
|
2014-06-04 09:30:42 +00:00
|
|
|
static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
|
2014-03-06 21:48:29 +00:00
|
|
|
{
|
|
|
|
unsigned int offset = skb_headlen(skb);
|
|
|
|
skb_frag_t frags[MAX_SKB_FRAGS];
|
|
|
|
int i;
|
|
|
|
struct ubuf_info *uarg;
|
|
|
|
struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->stats.tx_zerocopy_sent += 2;
|
|
|
|
queue->stats.tx_frag_overflow++;
|
2014-03-06 21:48:29 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_fill_frags(queue, nskb);
|
2014-03-06 21:48:29 +00:00
|
|
|
/* Subtract frags size, we will correct it later */
|
|
|
|
skb->truesize -= skb->data_len;
|
|
|
|
skb->len += nskb->len;
|
|
|
|
skb->data_len += nskb->len;
|
|
|
|
|
|
|
|
/* create a brand new frags array and coalesce there */
|
|
|
|
for (i = 0; offset < skb->len; i++) {
|
|
|
|
struct page *page;
|
|
|
|
unsigned int len;
|
|
|
|
|
|
|
|
BUG_ON(i >= MAX_SKB_FRAGS);
|
2014-10-28 15:29:31 +00:00
|
|
|
page = alloc_page(GFP_ATOMIC);
|
2014-03-06 21:48:29 +00:00
|
|
|
if (!page) {
|
|
|
|
int j;
|
|
|
|
skb->truesize += skb->data_len;
|
|
|
|
for (j = 0; j < i; j++)
|
|
|
|
put_page(frags[j].page.p);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (offset + PAGE_SIZE < skb->len)
|
|
|
|
len = PAGE_SIZE;
|
|
|
|
else
|
|
|
|
len = skb->len - offset;
|
|
|
|
if (skb_copy_bits(skb, offset, page_address(page), len))
|
|
|
|
BUG();
|
|
|
|
|
|
|
|
offset += len;
|
|
|
|
frags[i].page.p = page;
|
|
|
|
frags[i].page_offset = 0;
|
|
|
|
skb_frag_size_set(&frags[i], len);
|
|
|
|
}
|
|
|
|
/* swap out with old one */
|
|
|
|
memcpy(skb_shinfo(skb)->frags,
|
|
|
|
frags,
|
|
|
|
i * sizeof(skb_frag_t));
|
|
|
|
skb_shinfo(skb)->nr_frags = i;
|
|
|
|
skb->truesize += i * PAGE_SIZE;
|
|
|
|
|
|
|
|
/* remove traces of mapped pages and frag_list */
|
|
|
|
skb_frag_list_init(skb);
|
|
|
|
uarg = skb_shinfo(skb)->destructor_arg;
|
2014-08-12 10:48:07 +00:00
|
|
|
/* increase inflight counter to offset decrement in callback */
|
|
|
|
atomic_inc(&queue->inflight_packets);
|
2014-03-06 21:48:29 +00:00
|
|
|
uarg->callback(uarg, true);
|
|
|
|
skb_shinfo(skb)->destructor_arg = NULL;
|
|
|
|
|
2014-08-12 10:48:07 +00:00
|
|
|
xenvif_skb_zerocopy_prepare(queue, nskb);
|
2014-03-06 21:48:29 +00:00
|
|
|
kfree_skb(nskb);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
2013-08-26 11:59:38 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static int xenvif_tx_submit(struct xenvif_queue *queue)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
|
|
|
|
struct gnttab_copy *gop_copy = queue->tx_copy_ops;
|
2011-03-15 00:06:18 +00:00
|
|
|
struct sk_buff *skb;
|
2013-08-26 11:59:38 +00:00
|
|
|
int work_done = 0;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_tx_request *txp;
|
|
|
|
u16 pending_idx;
|
|
|
|
unsigned data_len;
|
|
|
|
|
2014-03-06 21:48:23 +00:00
|
|
|
pending_idx = XENVIF_TX_CB(skb)->pending_idx;
|
2014-06-04 09:30:42 +00:00
|
|
|
txp = &queue->pending_tx_info[pending_idx].req;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
/* Check the remap error code. */
|
2014-06-04 09:30:42 +00:00
|
|
|
if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
|
2014-07-18 18:08:03 +00:00
|
|
|
/* If there was an error, xenvif_tx_check_gop is
|
|
|
|
* expected to release all the frags which were mapped,
|
|
|
|
* so kfree_skb shouldn't do it again
|
|
|
|
*/
|
2011-03-15 00:06:18 +00:00
|
|
|
skb_shinfo(skb)->nr_frags = 0;
|
2014-07-18 18:08:03 +00:00
|
|
|
if (skb_has_frag_list(skb)) {
|
|
|
|
struct sk_buff *nskb =
|
|
|
|
skb_shinfo(skb)->frag_list;
|
|
|
|
skb_shinfo(nskb)->nr_frags = 0;
|
|
|
|
}
|
2011-03-15 00:06:18 +00:00
|
|
|
kfree_skb(skb);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
data_len = skb->len;
|
2014-06-04 09:30:42 +00:00
|
|
|
callback_param(queue, pending_idx).ctx = NULL;
|
2011-03-15 00:06:18 +00:00
|
|
|
if (data_len < txp->size) {
|
|
|
|
/* Append the packet payload as a fragment. */
|
|
|
|
txp->offset += data_len;
|
|
|
|
txp->size -= data_len;
|
|
|
|
} else {
|
|
|
|
/* Schedule a response immediately. */
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_idx_release(queue, pending_idx,
|
2014-04-02 17:04:58 +00:00
|
|
|
XEN_NETIF_RSP_OKAY);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (txp->flags & XEN_NETTXF_csum_blank)
|
|
|
|
skb->ip_summed = CHECKSUM_PARTIAL;
|
|
|
|
else if (txp->flags & XEN_NETTXF_data_validated)
|
|
|
|
skb->ip_summed = CHECKSUM_UNNECESSARY;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_fill_frags(queue, skb);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-03-06 21:48:29 +00:00
|
|
|
if (unlikely(skb_has_frag_list(skb))) {
|
2014-06-04 09:30:42 +00:00
|
|
|
if (xenvif_handle_frag_list(queue, skb)) {
|
2014-03-06 21:48:29 +00:00
|
|
|
if (net_ratelimit())
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2014-03-06 21:48:29 +00:00
|
|
|
"Not enough memory to consolidate frag_list!\n");
|
2014-08-12 10:48:07 +00:00
|
|
|
xenvif_skb_zerocopy_prepare(queue, skb);
|
2014-03-06 21:48:29 +00:00
|
|
|
kfree_skb(skb);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
skb->dev = queue->vif->dev;
|
2011-03-15 00:06:18 +00:00
|
|
|
skb->protocol = eth_type_trans(skb, skb->dev);
|
2013-03-25 20:19:58 +00:00
|
|
|
skb_reset_network_header(skb);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if (checksum_setup(queue, skb)) {
|
|
|
|
netdev_dbg(queue->vif->dev,
|
2011-03-15 00:06:18 +00:00
|
|
|
"Can't setup checksum in net_tx_action\n");
|
2014-03-06 21:48:26 +00:00
|
|
|
/* We have to set this flag to trigger the callback */
|
|
|
|
if (skb_shinfo(skb)->destructor_arg)
|
2014-08-12 10:48:07 +00:00
|
|
|
xenvif_skb_zerocopy_prepare(queue, skb);
|
2011-03-15 00:06:18 +00:00
|
|
|
kfree_skb(skb);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2013-03-26 23:11:22 +00:00
|
|
|
skb_probe_transport_header(skb, 0);
|
2013-03-25 20:19:58 +00:00
|
|
|
|
2013-12-17 11:44:35 +00:00
|
|
|
/* If the packet is GSO then we will have just set up the
|
|
|
|
* transport header offset in checksum_setup so it's now
|
|
|
|
* straightforward to calculate gso_segs.
|
|
|
|
*/
|
|
|
|
if (skb_is_gso(skb)) {
|
|
|
|
int mss = skb_shinfo(skb)->gso_size;
|
|
|
|
int hdrlen = skb_transport_header(skb) -
|
|
|
|
skb_mac_header(skb) +
|
|
|
|
tcp_hdrlen(skb);
|
|
|
|
|
|
|
|
skb_shinfo(skb)->gso_segs =
|
|
|
|
DIV_ROUND_UP(skb->len - hdrlen, mss);
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->stats.rx_bytes += skb->len;
|
|
|
|
queue->stats.rx_packets++;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2013-08-26 11:59:38 +00:00
|
|
|
work_done++;
|
|
|
|
|
2014-03-06 21:48:26 +00:00
|
|
|
/* Set this flag right before netif_receive_skb, otherwise
|
|
|
|
* someone might think this packet already left netback, and
|
|
|
|
* do a skb_copy_ubufs while we are still in control of the
|
|
|
|
* skb. E.g. the __pskb_pull_tail earlier can do such thing.
|
|
|
|
*/
|
2014-03-06 21:48:28 +00:00
|
|
|
if (skb_shinfo(skb)->destructor_arg) {
|
2014-08-12 10:48:07 +00:00
|
|
|
xenvif_skb_zerocopy_prepare(queue, skb);
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->stats.tx_zerocopy_sent++;
|
2014-03-06 21:48:28 +00:00
|
|
|
}
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2013-08-26 11:59:38 +00:00
|
|
|
netif_receive_skb(skb);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
2013-08-26 11:59:38 +00:00
|
|
|
|
|
|
|
return work_done;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:25 +00:00
|
|
|
void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
|
|
|
|
{
|
2014-03-06 21:48:26 +00:00
|
|
|
unsigned long flags;
|
|
|
|
pending_ring_idx_t index;
|
2014-06-04 09:30:42 +00:00
|
|
|
struct xenvif_queue *queue = ubuf_to_queue(ubuf);
|
2014-03-06 21:48:26 +00:00
|
|
|
|
|
|
|
/* This is the only place where we grab this lock, to protect callbacks
|
|
|
|
* from each other.
|
|
|
|
*/
|
2014-06-04 09:30:42 +00:00
|
|
|
spin_lock_irqsave(&queue->callback_lock, flags);
|
2014-03-06 21:48:26 +00:00
|
|
|
do {
|
|
|
|
u16 pending_idx = ubuf->desc;
|
|
|
|
ubuf = (struct ubuf_info *) ubuf->ctx;
|
2014-06-04 09:30:42 +00:00
|
|
|
BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
|
2014-03-06 21:48:26 +00:00
|
|
|
MAX_PENDING_REQS);
|
2014-06-04 09:30:42 +00:00
|
|
|
index = pending_index(queue->dealloc_prod);
|
|
|
|
queue->dealloc_ring[index] = pending_idx;
|
2014-03-06 21:48:26 +00:00
|
|
|
/* Sync with xenvif_tx_dealloc_action:
|
|
|
|
* insert idx then incr producer.
|
|
|
|
*/
|
|
|
|
smp_wmb();
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->dealloc_prod++;
|
2014-03-06 21:48:26 +00:00
|
|
|
} while (ubuf);
|
2014-06-04 09:30:42 +00:00
|
|
|
wake_up(&queue->dealloc_wq);
|
|
|
|
spin_unlock_irqrestore(&queue->callback_lock, flags);
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2014-03-06 21:48:28 +00:00
|
|
|
if (likely(zerocopy_success))
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->stats.tx_zerocopy_success++;
|
2014-03-06 21:48:28 +00:00
|
|
|
else
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->stats.tx_zerocopy_fail++;
|
2014-08-12 10:48:07 +00:00
|
|
|
xenvif_skb_zerocopy_complete(queue);
|
2014-03-06 21:48:26 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
|
2014-03-06 21:48:26 +00:00
|
|
|
{
|
|
|
|
struct gnttab_unmap_grant_ref *gop;
|
|
|
|
pending_ring_idx_t dc, dp;
|
|
|
|
u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
|
|
|
|
unsigned int i = 0;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
dc = queue->dealloc_cons;
|
|
|
|
gop = queue->tx_unmap_ops;
|
2014-03-06 21:48:26 +00:00
|
|
|
|
|
|
|
/* Free up any grants we have finished using */
|
|
|
|
do {
|
2014-06-04 09:30:42 +00:00
|
|
|
dp = queue->dealloc_prod;
|
2014-03-06 21:48:26 +00:00
|
|
|
|
|
|
|
/* Ensure we see all indices enqueued by all
|
|
|
|
* xenvif_zerocopy_callback().
|
|
|
|
*/
|
|
|
|
smp_rmb();
|
|
|
|
|
|
|
|
while (dc != dp) {
|
2014-06-04 09:30:42 +00:00
|
|
|
BUG_ON(gop - queue->tx_unmap_ops > MAX_PENDING_REQS);
|
2014-03-06 21:48:26 +00:00
|
|
|
pending_idx =
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->dealloc_ring[pending_index(dc++)];
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
pending_idx_release[gop-queue->tx_unmap_ops] =
|
2014-03-06 21:48:26 +00:00
|
|
|
pending_idx;
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->pages_to_unmap[gop-queue->tx_unmap_ops] =
|
|
|
|
queue->mmap_pages[pending_idx];
|
2014-03-06 21:48:26 +00:00
|
|
|
gnttab_set_unmap_op(gop,
|
2014-06-04 09:30:42 +00:00
|
|
|
idx_to_kaddr(queue, pending_idx),
|
2014-03-06 21:48:26 +00:00
|
|
|
GNTMAP_host_map,
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->grant_tx_handle[pending_idx]);
|
|
|
|
xenvif_grant_handle_reset(queue, pending_idx);
|
2014-03-06 21:48:26 +00:00
|
|
|
++gop;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
} while (dp != queue->dealloc_prod);
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->dealloc_cons = dc;
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if (gop - queue->tx_unmap_ops > 0) {
|
2014-03-06 21:48:26 +00:00
|
|
|
int ret;
|
2014-06-04 09:30:42 +00:00
|
|
|
ret = gnttab_unmap_refs(queue->tx_unmap_ops,
|
2014-03-06 21:48:26 +00:00
|
|
|
NULL,
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->pages_to_unmap,
|
|
|
|
gop - queue->tx_unmap_ops);
|
2014-03-06 21:48:26 +00:00
|
|
|
if (ret) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tx ret %d\n",
|
|
|
|
gop - queue->tx_unmap_ops, ret);
|
|
|
|
for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
|
2014-03-06 21:48:26 +00:00
|
|
|
if (gop[i].status != GNTST_okay)
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2014-03-06 21:48:26 +00:00
|
|
|
" host_addr: %llx handle: %x status: %d\n",
|
|
|
|
gop[i].host_addr,
|
|
|
|
gop[i].handle,
|
|
|
|
gop[i].status);
|
|
|
|
}
|
|
|
|
BUG();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
|
|
|
|
xenvif_idx_release(queue, pending_idx_release[i],
|
2014-03-06 21:48:26 +00:00
|
|
|
XEN_NETIF_RSP_OKAY);
|
2014-03-06 21:48:25 +00:00
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
/* Called after netfront has transmitted */
|
2014-06-04 09:30:42 +00:00
|
|
|
int xenvif_tx_action(struct xenvif_queue *queue, int budget)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-04-02 17:04:58 +00:00
|
|
|
unsigned nr_mops, nr_cops = 0;
|
2014-03-06 21:48:26 +00:00
|
|
|
int work_done, ret;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if (unlikely(!tx_work_todo(queue)))
|
2013-08-26 11:59:38 +00:00
|
|
|
return 0;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-04-02 17:04:58 +00:00
|
|
|
if (nr_cops == 0)
|
2013-08-26 11:59:38 +00:00
|
|
|
return 0;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
|
2014-04-02 17:04:58 +00:00
|
|
|
if (nr_mops != 0) {
|
2014-06-04 09:30:42 +00:00
|
|
|
ret = gnttab_map_refs(queue->tx_map_ops,
|
2014-04-02 17:04:58 +00:00
|
|
|
NULL,
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->pages_to_map,
|
2014-04-02 17:04:58 +00:00
|
|
|
nr_mops);
|
|
|
|
BUG_ON(ret);
|
|
|
|
}
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
work_done = xenvif_tx_submit(queue);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2013-08-26 11:59:38 +00:00
|
|
|
return work_done;
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
|
2013-08-26 11:59:39 +00:00
|
|
|
u8 status)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
|
|
|
struct pending_tx_info *pending_tx_info;
|
2014-03-06 21:48:26 +00:00
|
|
|
pending_ring_idx_t index;
|
|
|
|
unsigned long flags;
|
2013-04-22 02:20:42 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
pending_tx_info = &queue->pending_tx_info[pending_idx];
|
|
|
|
spin_lock_irqsave(&queue->response_lock, flags);
|
|
|
|
make_tx_response(queue, &pending_tx_info->req, status);
|
|
|
|
index = pending_index(queue->pending_prod);
|
|
|
|
queue->pending_ring[index] = pending_idx;
|
2014-03-06 21:48:27 +00:00
|
|
|
/* TX shouldn't use the index before we give it back here */
|
|
|
|
mb();
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->pending_prod++;
|
|
|
|
spin_unlock_irqrestore(&queue->response_lock, flags);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2013-04-22 02:20:42 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static void make_tx_response(struct xenvif_queue *queue,
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_tx_request *txp,
|
|
|
|
s8 st)
|
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_IDX i = queue->tx.rsp_prod_pvt;
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_tx_response *resp;
|
|
|
|
int notify;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
resp = RING_GET_RESPONSE(&queue->tx, i);
|
2011-03-15 00:06:18 +00:00
|
|
|
resp->id = txp->id;
|
|
|
|
resp->status = st;
|
|
|
|
|
|
|
|
if (txp->flags & XEN_NETTXF_extra_info)
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->tx.rsp_prod_pvt = ++i;
|
|
|
|
RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
|
2011-03-15 00:06:18 +00:00
|
|
|
if (notify)
|
2014-06-04 09:30:42 +00:00
|
|
|
notify_remote_via_irq(queue->tx_irq);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
|
2011-03-15 00:06:18 +00:00
|
|
|
u16 id,
|
|
|
|
s8 st,
|
|
|
|
u16 offset,
|
|
|
|
u16 size,
|
|
|
|
u16 flags)
|
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
RING_IDX i = queue->rx.rsp_prod_pvt;
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_rx_response *resp;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
resp = RING_GET_RESPONSE(&queue->rx, i);
|
2011-03-15 00:06:18 +00:00
|
|
|
resp->offset = offset;
|
|
|
|
resp->flags = flags;
|
|
|
|
resp->id = id;
|
|
|
|
resp->status = (s16)size;
|
|
|
|
if (st < 0)
|
|
|
|
resp->status = (s16)st;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->rx.rsp_prod_pvt = ++i;
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
return resp;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
|
2014-03-06 21:48:26 +00:00
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
struct gnttab_unmap_grant_ref tx_unmap_op;
|
|
|
|
|
|
|
|
gnttab_set_unmap_op(&tx_unmap_op,
|
2014-06-04 09:30:42 +00:00
|
|
|
idx_to_kaddr(queue, pending_idx),
|
2014-03-06 21:48:26 +00:00
|
|
|
GNTMAP_host_map,
|
2014-06-04 09:30:42 +00:00
|
|
|
queue->grant_tx_handle[pending_idx]);
|
|
|
|
xenvif_grant_handle_reset(queue, pending_idx);
|
2014-03-06 21:48:26 +00:00
|
|
|
|
|
|
|
ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
|
2014-06-04 09:30:42 +00:00
|
|
|
&queue->mmap_pages[pending_idx], 1);
|
2014-03-24 23:59:51 +00:00
|
|
|
if (ret) {
|
2014-06-04 09:30:42 +00:00
|
|
|
netdev_err(queue->vif->dev,
|
2014-03-24 23:59:51 +00:00
|
|
|
"Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: %x status: %d\n",
|
|
|
|
ret,
|
|
|
|
pending_idx,
|
|
|
|
tx_unmap_op.host_addr,
|
|
|
|
tx_unmap_op.handle,
|
|
|
|
tx_unmap_op.status);
|
|
|
|
BUG();
|
|
|
|
}
|
2014-03-06 21:48:26 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline int tx_work_todo(struct xenvif_queue *queue)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
|
2011-03-15 00:06:18 +00:00
|
|
|
return 1;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
|
2014-03-06 21:48:26 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
return queue->dealloc_cons != queue->dealloc_prod;
|
2014-03-06 21:48:26 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
if (queue->tx.sring)
|
|
|
|
xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
|
|
|
|
queue->tx.sring);
|
|
|
|
if (queue->rx.sring)
|
|
|
|
xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
|
|
|
|
queue->rx.sring);
|
2011-03-15 00:06:18 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
int xenvif_map_frontend_rings(struct xenvif_queue *queue,
|
2013-08-26 11:59:39 +00:00
|
|
|
grant_ref_t tx_ring_ref,
|
|
|
|
grant_ref_t rx_ring_ref)
|
2011-03-15 00:06:18 +00:00
|
|
|
{
|
2011-09-29 15:53:31 +00:00
|
|
|
void *addr;
|
2011-03-15 00:06:18 +00:00
|
|
|
struct xen_netif_tx_sring *txs;
|
|
|
|
struct xen_netif_rx_sring *rxs;
|
|
|
|
|
|
|
|
int err = -ENOMEM;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
|
2011-09-29 15:53:31 +00:00
|
|
|
tx_ring_ref, &addr);
|
|
|
|
if (err)
|
2011-03-15 00:06:18 +00:00
|
|
|
goto err;
|
|
|
|
|
2011-09-29 15:53:31 +00:00
|
|
|
txs = (struct xen_netif_tx_sring *)addr;
|
2014-06-04 09:30:42 +00:00
|
|
|
BACK_RING_INIT(&queue->tx, txs, PAGE_SIZE);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
|
2011-09-29 15:53:31 +00:00
|
|
|
rx_ring_ref, &addr);
|
|
|
|
if (err)
|
2011-03-15 00:06:18 +00:00
|
|
|
goto err;
|
|
|
|
|
2011-09-29 15:53:31 +00:00
|
|
|
rxs = (struct xen_netif_rx_sring *)addr;
|
2014-06-04 09:30:42 +00:00
|
|
|
BACK_RING_INIT(&queue->rx, rxs, PAGE_SIZE);
|
2011-03-15 00:06:18 +00:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
err:
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_unmap_frontend_rings(queue);
|
2011-03-15 00:06:18 +00:00
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2014-10-22 13:08:55 +00:00
|
|
|
static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
{
|
2014-10-22 13:08:55 +00:00
|
|
|
struct xenvif *vif = queue->vif;
|
|
|
|
|
|
|
|
queue->stalled = true;
|
|
|
|
|
|
|
|
/* At least one queue has stalled? Disable the carrier. */
|
|
|
|
spin_lock(&vif->lock);
|
|
|
|
if (vif->stalled_queues++ == 0) {
|
|
|
|
netdev_info(vif->dev, "Guest Rx stalled");
|
|
|
|
netif_carrier_off(vif->dev);
|
|
|
|
}
|
|
|
|
spin_unlock(&vif->lock);
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
}
|
|
|
|
|
2014-10-22 13:08:55 +00:00
|
|
|
static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
{
|
2014-10-22 13:08:55 +00:00
|
|
|
struct xenvif *vif = queue->vif;
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
|
2014-10-22 13:08:55 +00:00
|
|
|
queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
|
|
|
|
queue->stalled = false;
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
|
2014-10-22 13:08:55 +00:00
|
|
|
/* All queues are ready? Enable the carrier. */
|
|
|
|
spin_lock(&vif->lock);
|
|
|
|
if (--vif->stalled_queues == 0) {
|
|
|
|
netdev_info(vif->dev, "Guest Rx ready");
|
|
|
|
netif_carrier_on(vif->dev);
|
|
|
|
}
|
|
|
|
spin_unlock(&vif->lock);
|
|
|
|
}
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
|
2014-10-22 13:08:55 +00:00
|
|
|
static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
|
|
|
|
{
|
|
|
|
RING_IDX prod, cons;
|
|
|
|
|
|
|
|
prod = queue->rx.sring->req_prod;
|
|
|
|
cons = queue->rx.req_cons;
|
|
|
|
|
|
|
|
return !queue->stalled
|
|
|
|
&& prod - cons < XEN_NETBK_RX_SLOTS_MAX
|
|
|
|
&& time_after(jiffies,
|
2014-12-18 11:13:06 +00:00
|
|
|
queue->last_rx_time + queue->vif->stall_timeout);
|
2014-10-22 13:08:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
|
|
|
|
{
|
|
|
|
RING_IDX prod, cons;
|
|
|
|
|
|
|
|
prod = queue->rx.sring->req_prod;
|
|
|
|
cons = queue->rx.req_cons;
|
|
|
|
|
|
|
|
return queue->stalled
|
|
|
|
&& prod - cons >= XEN_NETBK_RX_SLOTS_MAX;
|
|
|
|
}
|
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
static bool xenvif_have_rx_work(struct xenvif_queue *queue)
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
{
|
2014-10-22 13:08:54 +00:00
|
|
|
return (!skb_queue_empty(&queue->rx_queue)
|
|
|
|
&& xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX))
|
2014-12-18 11:13:06 +00:00
|
|
|
|| (queue->vif->stall_timeout &&
|
|
|
|
(xenvif_rx_queue_stalled(queue)
|
|
|
|
|| xenvif_rx_queue_ready(queue)))
|
2014-10-22 13:08:54 +00:00
|
|
|
|| kthread_should_stop()
|
|
|
|
|| queue->vif->disabled;
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
}
|
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
{
|
2014-10-22 13:08:54 +00:00
|
|
|
struct sk_buff *skb;
|
|
|
|
long timeout;
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
skb = skb_peek(&queue->rx_queue);
|
|
|
|
if (!skb)
|
|
|
|
return MAX_SCHEDULE_TIMEOUT;
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
timeout = XENVIF_RX_CB(skb)->expires - jiffies;
|
|
|
|
return timeout < 0 ? 0 : timeout;
|
|
|
|
}
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
/* Wait until the guest Rx thread has work.
|
|
|
|
*
|
|
|
|
* The timeout needs to be adjusted based on the current head of the
|
|
|
|
* queue (and not just the head at the beginning). In particular, if
|
|
|
|
* the queue is initially empty an infinite timeout is used and this
|
|
|
|
* needs to be reduced when a skb is queued.
|
|
|
|
*
|
|
|
|
* This cannot be done with wait_event_timeout() because it only
|
|
|
|
* calculates the timeout once.
|
|
|
|
*/
|
|
|
|
static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
|
|
|
|
{
|
|
|
|
DEFINE_WAIT(wait);
|
|
|
|
|
|
|
|
if (xenvif_have_rx_work(queue))
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (;;) {
|
|
|
|
long ret;
|
|
|
|
|
|
|
|
prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
|
|
|
|
if (xenvif_have_rx_work(queue))
|
|
|
|
break;
|
|
|
|
ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
|
|
|
|
if (!ret)
|
|
|
|
break;
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
}
|
2014-10-22 13:08:54 +00:00
|
|
|
finish_wait(&queue->wq, &wait);
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:24 +00:00
|
|
|
int xenvif_kthread_guest_rx(void *data)
|
2013-08-26 11:59:38 +00:00
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
struct xenvif_queue *queue = data;
|
2014-10-22 13:08:54 +00:00
|
|
|
struct xenvif *vif = queue->vif;
|
2013-08-26 11:59:38 +00:00
|
|
|
|
2014-12-18 11:13:06 +00:00
|
|
|
if (!vif->stall_timeout)
|
|
|
|
xenvif_queue_carrier_on(queue);
|
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
for (;;) {
|
|
|
|
xenvif_wait_for_rx_work(queue);
|
2014-04-01 11:46:12 +00:00
|
|
|
|
xen-netback: Turn off the carrier if the guest is not able to receive
Currently when the guest is not able to receive more packets, qdisc layer starts
a timer, and when it goes off, qdisc is started again to deliver a packet again.
This is a very slow way to drain the queues, consumes unnecessary resources and
slows down other guests shutdown.
This patch change the behaviour by turning the carrier off when that timer
fires, so all the packets are freed up which were stucked waiting for that vif.
Instead of the rx_queue_purge bool it uses the VIF_STATUS_RX_PURGE_EVENT bit to
signal the thread that either the timeout happened or an RX interrupt arrived,
so the thread can check what it should do. It also disables NAPI, so the guest
can't transmit, but leaves the interrupts on, so it can resurrect.
Only the queues which brought down the interface can enable it again, the bit
QUEUE_STATUS_RX_STALLED makes sure of that.
Signed-off-by: Zoltan Kiss <zoltan.kiss@citrix.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: xen-devel@lists.xenproject.org
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-04 15:20:58 +00:00
|
|
|
if (kthread_should_stop())
|
|
|
|
break;
|
|
|
|
|
2014-04-01 11:46:12 +00:00
|
|
|
/* This frontend is found to be rogue, disable it in
|
|
|
|
* kthread context. Currently this is only set when
|
|
|
|
* netback finds out frontend sends malformed packet,
|
|
|
|
* but we cannot disable the interface in softirq
|
2014-06-04 09:30:42 +00:00
|
|
|
* context so we defer it here, if this thread is
|
|
|
|
* associated with queue 0.
|
2014-04-01 11:46:12 +00:00
|
|
|
*/
|
2014-10-22 13:08:54 +00:00
|
|
|
if (unlikely(vif->disabled && queue->id == 0)) {
|
|
|
|
xenvif_carrier_off(vif);
|
2015-02-02 16:57:51 +00:00
|
|
|
break;
|
2014-03-06 21:48:30 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
if (!skb_queue_empty(&queue->rx_queue))
|
|
|
|
xenvif_rx_action(queue);
|
2013-08-26 11:59:38 +00:00
|
|
|
|
2014-10-22 13:08:55 +00:00
|
|
|
/* If the guest hasn't provided any Rx slots for a
|
|
|
|
* while it's probably not responsive, drop the
|
|
|
|
* carrier so packets are dropped earlier.
|
|
|
|
*/
|
2014-12-18 11:13:06 +00:00
|
|
|
if (vif->stall_timeout) {
|
|
|
|
if (xenvif_rx_queue_stalled(queue))
|
|
|
|
xenvif_queue_carrier_off(queue);
|
|
|
|
else if (xenvif_rx_queue_ready(queue))
|
|
|
|
xenvif_queue_carrier_on(queue);
|
|
|
|
}
|
2014-10-22 13:08:55 +00:00
|
|
|
|
2014-10-22 13:08:54 +00:00
|
|
|
/* Queued packets may have foreign pages from other
|
|
|
|
* domains. These cannot be queued indefinitely as
|
|
|
|
* this would starve guests of grant refs and transmit
|
|
|
|
* slots.
|
|
|
|
*/
|
|
|
|
xenvif_rx_queue_drop_expired(queue);
|
|
|
|
|
|
|
|
xenvif_rx_queue_maybe_wake(queue);
|
|
|
|
|
2013-08-26 11:59:38 +00:00
|
|
|
cond_resched();
|
|
|
|
}
|
|
|
|
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
/* Bin any remaining skbs */
|
2014-10-22 13:08:54 +00:00
|
|
|
xenvif_rx_queue_purge(queue);
|
xen-netback: improve guest-receive-side flow control
The way that flow control works without this patch is that, in start_xmit()
the code uses xenvif_count_skb_slots() to predict how many slots
xenvif_gop_skb() will consume and then adds this to a 'req_cons_peek'
counter which it then uses to determine if the shared ring has that amount
of space available by checking whether 'req_prod' has passed that value.
If the ring doesn't have space the tx queue is stopped.
xenvif_gop_skb() will then consume slots and update 'req_cons' and issue
responses, updating 'rsp_prod' as it goes. The frontend will consume those
responses and post new requests, by updating req_prod. So, req_prod chases
req_cons which chases rsp_prod, and can never exceed that value. Thus if
xenvif_count_skb_slots() ever returns a number of slots greater than
xenvif_gop_skb() uses, req_cons_peek will get to a value that req_prod cannot
possibly achieve (since it's limited by the 'real' req_cons) and, if this
happens enough times, req_cons_peek gets more than a ring size ahead of
req_cons and the tx queue then remains stopped forever waiting for an
unachievable amount of space to become available in the ring.
Having two routines trying to calculate the same value is always going to be
fragile, so this patch does away with that. All we essentially need to do is
make sure that we have 'enough stuff' on our internal queue without letting
it build up uncontrollably. So start_xmit() makes a cheap optimistic check
of how much space is needed for an skb and only turns the queue off if that
is unachievable. net_rx_action() is the place where we could do with an
accurate predicition but, since that has proven tricky to calculate, a cheap
worse-case (but not too bad) estimate is all we really need since the only
thing we *must* prevent is xenvif_gop_skb() consuming more slots than are
available.
Without this patch I can trivially stall netback permanently by just doing
a large guest to guest file copy between two Windows Server 2008R2 VMs on a
single host.
Patch tested with frontends in:
- Windows Server 2008R2
- CentOS 6.0
- Debian Squeeze
- Debian Wheezy
- SLES11
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Annie Li <annie.li@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-12-06 16:36:07 +00:00
|
|
|
|
2013-08-26 11:59:38 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-08-12 10:48:07 +00:00
|
|
|
static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
|
|
|
|
{
|
|
|
|
/* Dealloc thread must remain running until all inflight
|
|
|
|
* packets complete.
|
|
|
|
*/
|
|
|
|
return kthread_should_stop() &&
|
|
|
|
!atomic_read(&queue->inflight_packets);
|
|
|
|
}
|
|
|
|
|
2014-03-06 21:48:26 +00:00
|
|
|
int xenvif_dealloc_kthread(void *data)
|
|
|
|
{
|
2014-06-04 09:30:42 +00:00
|
|
|
struct xenvif_queue *queue = data;
|
2014-03-06 21:48:26 +00:00
|
|
|
|
2014-08-12 10:48:07 +00:00
|
|
|
for (;;) {
|
2014-06-04 09:30:42 +00:00
|
|
|
wait_event_interruptible(queue->dealloc_wq,
|
|
|
|
tx_dealloc_work_todo(queue) ||
|
2014-08-12 10:48:07 +00:00
|
|
|
xenvif_dealloc_kthread_should_stop(queue));
|
|
|
|
if (xenvif_dealloc_kthread_should_stop(queue))
|
2014-03-06 21:48:26 +00:00
|
|
|
break;
|
|
|
|
|
2014-06-04 09:30:42 +00:00
|
|
|
xenvif_tx_dealloc_action(queue);
|
2014-03-06 21:48:26 +00:00
|
|
|
cond_resched();
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Unmap anything remaining*/
|
2014-06-04 09:30:42 +00:00
|
|
|
if (tx_dealloc_work_todo(queue))
|
|
|
|
xenvif_tx_dealloc_action(queue);
|
2014-03-06 21:48:26 +00:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
static int __init netback_init(void)
|
|
|
|
{
|
|
|
|
int rc = 0;
|
|
|
|
|
2011-12-14 20:12:13 +00:00
|
|
|
if (!xen_domain())
|
2011-03-15 00:06:18 +00:00
|
|
|
return -ENODEV;
|
|
|
|
|
2014-06-04 09:30:43 +00:00
|
|
|
/* Allow as many queues as there are CPUs, by default */
|
|
|
|
xenvif_max_queues = num_online_cpus();
|
|
|
|
|
2013-05-02 00:43:59 +00:00
|
|
|
if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
|
2013-06-28 04:57:49 +00:00
|
|
|
pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
|
|
|
|
fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
|
2013-05-02 00:43:59 +00:00
|
|
|
fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
|
2013-04-22 02:20:42 +00:00
|
|
|
}
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
rc = xenvif_xenbus_init();
|
|
|
|
if (rc)
|
|
|
|
goto failed_init;
|
|
|
|
|
2014-07-08 18:49:14 +00:00
|
|
|
#ifdef CONFIG_DEBUG_FS
|
|
|
|
xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
|
|
|
|
if (IS_ERR_OR_NULL(xen_netback_dbg_root))
|
|
|
|
pr_warn("Init of debugfs returned %ld!\n",
|
|
|
|
PTR_ERR(xen_netback_dbg_root));
|
|
|
|
#endif /* CONFIG_DEBUG_FS */
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
failed_init:
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
module_init(netback_init);
|
|
|
|
|
2013-05-16 23:26:11 +00:00
|
|
|
static void __exit netback_fini(void)
|
|
|
|
{
|
2014-07-08 18:49:14 +00:00
|
|
|
#ifdef CONFIG_DEBUG_FS
|
|
|
|
if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
|
|
|
|
debugfs_remove_recursive(xen_netback_dbg_root);
|
|
|
|
#endif /* CONFIG_DEBUG_FS */
|
2013-05-16 23:26:11 +00:00
|
|
|
xenvif_xenbus_fini();
|
|
|
|
}
|
|
|
|
module_exit(netback_fini);
|
|
|
|
|
2011-03-15 00:06:18 +00:00
|
|
|
MODULE_LICENSE("Dual BSD/GPL");
|
2011-06-30 18:19:09 +00:00
|
|
|
MODULE_ALIAS("xen-backend:vif");
|