Source-Changes-HG archive
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]
[src/netbsd-8]: src Pull up following revision(s) (requested by knakahara in ...
details: https://anonhg.NetBSD.org/src/rev/83ddb50c4bd7
branches: netbsd-8
changeset: 851344:83ddb50c4bd7
user: martin <martin%NetBSD.org@localhost>
date: Mon Feb 05 15:07:30 2018 +0000
description:
Pull up following revision(s) (requested by knakahara in ticket #529):
sys/dev/pci/if_wm.c: revision 1.560
sys/dev/pci/if_wm.c: revision 1.561
sys/dev/pci/if_wm.c: revision 1.562
share/man/man4/wm.4: revision 1.37
share/man/man4/wm.4: revision 1.38
sys/dev/pci/if_wm.c: revision 1.551
sys/dev/pci/if_wm.c: revision 1.553
sys/dev/pci/if_wm.c: revision 1.554
sys/dev/pci/if_wm.c: revision 1.555
sys/dev/pci/if_wm.c: revision 1.556
sys/dev/pci/if_wm.c: revision 1.557
sys/dev/pci/if_wm.c: revision 1.558
sys/dev/pci/if_wm.c: revision 1.559
PR/52885 - Shinichi Doyashiki -- typo in comment
Fix legacy Tx descriptors printing when WM_DEBUG is enabled.
improve comments
Fix wm_watchdog_txq() lock region.
Not only wm_txeof() but also wm_watchdog_txq() itself requires txq_lock
as it reads Tx descriptor management variables such as "txq_free".
There is almost no influence on performance.
Fix duplicated "rxintr" evcnt counting. Pointed out by ozaki-r@n.o, thanks.
wm_txeof() can limit the loop count the same as wm_rxeof() now.
add WM_TX_PROCESS_LIMIT_DEFAULT and WM_TX_INTR_PROCESS_LIMIT_DEFAULT man.
More markup.
CID-1427779: Fix uninitialized variables
Fix 82574 MSI-X mode cannot receive packets after 82574 receives high rate traffic.
In short, 82574 MSI-X mode does not cause RXQ MSI-X vector when 82574's
phys FIFO overflows. I don't know why but 82574 causes not RXQ MSI-X vector
but OTHER MSI-X vector at the situation.
see:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v4.15-rc9&id=4aea7a5c5e940c1723add439f4088844cd26196d
advised by msaitoh@n.o, thanks.
Fix if_wm.c:r1.557 merge miss, sorry.
Fix unmatched return type. The return value of wm_txeof() is not useded yet.
Make wm(4) watchdog MP-safe. There is almost no influence on performance.
wm(4) does not use ifp->if_watchdog now, that is, it does not touch
ifp->if_timer.
It also uses own callout(wm_tick) as watchdog now. The watchdog uses
per-queue counter to check timeout. So, global lock is not required.
diffstat:
share/man/man4/wm.4 | 10 +-
sys/dev/pci/if_wm.c | 242 ++++++++++++++++++++++++++++++++++++++-------------
2 files changed, 185 insertions(+), 67 deletions(-)
diffs (truncated from 586 to 300 lines):
diff -r 6a0681c37a4c -r 83ddb50c4bd7 share/man/man4/wm.4
--- a/share/man/man4/wm.4 Mon Feb 05 14:55:15 2018 +0000
+++ b/share/man/man4/wm.4 Mon Feb 05 15:07:30 2018 +0000
@@ -1,4 +1,4 @@
-.\" $NetBSD: wm.4,v 1.36 2017/04/13 10:37:36 knakahara Exp $
+.\" $NetBSD: wm.4,v 1.36.4.1 2018/02/05 15:07:30 martin Exp $
.\"
.\" Copyright 2002, 2003 Wasabi Systems, Inc.
.\" All rights reserved.
@@ -33,7 +33,7 @@
.\" ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
.\" POSSIBILITY OF SUCH DAMAGE.
.\"
-.Dd March 22, 2017
+.Dd January 18, 2018
.Dt WM 4
.Os
.Sh NAME
@@ -182,6 +182,9 @@
The default value is 100.
When you increase this value, both the receive latency and
the receive throughput will increase.
+.It Dv WM_TX_PROCESS_LIMIT_DEFAULT
+Transmit side of
+.Dv WM_RX_PROCESS_LIMIT_DEFAULT .
.It Dv WM_RX_INTR_PROCESS_LIMIT_DEFAULT
The maximum number of received packets processed in each
hardware interrupt context.
@@ -191,6 +194,9 @@
The default value is 0.
When you increase this value, both the receive latency and
the receive throughput will decrease.
+.It Dv WM_TX_INTR_PROCESS_LIMIT_DEFAULT
+Transmit side of
+.Dv WM_RX_INTR_PROCESS_LIMIT_DEFAULT .
.It Dv WM_EVENT_COUNTERS
Enable many event counters such as each Tx drop counter and Rx interrupt
counter.
diff -r 6a0681c37a4c -r 83ddb50c4bd7 sys/dev/pci/if_wm.c
--- a/sys/dev/pci/if_wm.c Mon Feb 05 14:55:15 2018 +0000
+++ b/sys/dev/pci/if_wm.c Mon Feb 05 15:07:30 2018 +0000
@@ -1,4 +1,4 @@
-/* $NetBSD: if_wm.c,v 1.508.4.12 2018/01/13 21:42:45 snj Exp $ */
+/* $NetBSD: if_wm.c,v 1.508.4.13 2018/02/05 15:07:30 martin Exp $ */
/*
* Copyright (c) 2001, 2002, 2003, 2004 Wasabi Systems, Inc.
@@ -83,7 +83,7 @@
*/
#include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: if_wm.c,v 1.508.4.12 2018/01/13 21:42:45 snj Exp $");
+__KERNEL_RCSID(0, "$NetBSD: if_wm.c,v 1.508.4.13 2018/02/05 15:07:30 martin Exp $");
#ifdef _KERNEL_OPT
#include "opt_net_mpsafe.h"
@@ -183,6 +183,11 @@
int wm_disable_msi = WM_DISABLE_MSI;
int wm_disable_msix = WM_DISABLE_MSIX;
+#ifndef WM_WATCHDOG_TIMEOUT
+#define WM_WATCHDOG_TIMEOUT 5
+#endif
+static int wm_watchdog_timeout = WM_WATCHDOG_TIMEOUT;
+
/*
* Transmit descriptor list size. Due to errata, we can only have
* 256 hardware descriptors in the ring on < 82544, but we use 4096
@@ -213,6 +218,13 @@
#define WM_TXINTERQSIZE 256
+#ifndef WM_TX_PROCESS_LIMIT_DEFAULT
+#define WM_TX_PROCESS_LIMIT_DEFAULT 100U
+#endif
+#ifndef WM_TX_INTR_PROCESS_LIMIT_DEFAULT
+#define WM_TX_INTR_PROCESS_LIMIT_DEFAULT 0U
+#endif
+
/*
* Receive descriptor list size. We have one Rx buffer for normal
* sized packets. Jumbo packets consume 5 Rx buffers for a full-sized
@@ -356,6 +368,9 @@
bool txq_stopping;
+ bool txq_watchdog;
+ time_t txq_lastsent;
+
uint32_t txq_packets; /* for AIM */
uint32_t txq_bytes; /* for AIM */
#ifdef WM_EVENT_COUNTERS
@@ -417,6 +432,7 @@
uint32_t rxq_bytes; /* for AIM */
#ifdef WM_EVENT_COUNTERS
WM_Q_EVCNT_DEFINE(rxq, rxintr); /* Rx interrupts */
+ WM_Q_EVCNT_DEFINE(rxq, rxdefer); /* Rx deferred processing */
WM_Q_EVCNT_DEFINE(rxq, rxipsum); /* IP checksums checked in-bound */
WM_Q_EVCNT_DEFINE(rxq, rxtusum); /* TCP/UDP cksums checked in-bound */
@@ -518,6 +534,8 @@
int sc_nqueues;
struct wm_queue *sc_queue;
+ u_int sc_tx_process_limit; /* Tx processing repeat limit in softint */
+ u_int sc_tx_intr_process_limit; /* Tx processing repeat limit in H/W intr */
u_int sc_rx_process_limit; /* Rx processing repeat limit in softint */
u_int sc_rx_intr_process_limit; /* Rx processing repeat limit in H/W intr */
@@ -670,7 +688,8 @@
static bool wm_suspend(device_t, const pmf_qual_t *);
static bool wm_resume(device_t, const pmf_qual_t *);
static void wm_watchdog(struct ifnet *);
-static void wm_watchdog_txq(struct ifnet *, struct wm_txqueue *);
+static void wm_watchdog_txq(struct ifnet *, struct wm_txqueue *, uint16_t *);
+static void wm_watchdog_txq_locked(struct ifnet *, struct wm_txqueue *, uint16_t *);
static void wm_tick(void *);
static int wm_ifflags_cb(struct ethercom *);
static int wm_ioctl(struct ifnet *, u_long, void *);
@@ -756,7 +775,7 @@
static void wm_deferred_start_locked(struct wm_txqueue *);
static void wm_handle_queue(void *);
/* Interrupt */
-static int wm_txeof(struct wm_softc *, struct wm_txqueue *);
+static int wm_txeof(struct wm_txqueue *, u_int);
static void wm_rxeof(struct wm_rxqueue *, u_int);
static void wm_linkintr_gmii(struct wm_softc *, uint32_t);
static void wm_linkintr_tbi(struct wm_softc *, uint32_t);
@@ -2672,7 +2691,7 @@
if (wm_is_using_multiqueue(sc))
ifp->if_transmit = wm_transmit;
}
- ifp->if_watchdog = wm_watchdog;
+ /* wm(4) doest not use ifp->if_watchdog, use wm_tick as watchdog. */
ifp->if_init = wm_init;
ifp->if_stop = wm_stop;
IFQ_SET_MAXLEN(&ifp->if_snd, max(WM_IFQUEUELEN, IFQ_MAXLEN));
@@ -2762,6 +2781,8 @@
ifp->if_capabilities |= IFCAP_TSOv6;
}
+ sc->sc_tx_process_limit = WM_TX_PROCESS_LIMIT_DEFAULT;
+ sc->sc_tx_intr_process_limit = WM_TX_INTR_PROCESS_LIMIT_DEFAULT;
sc->sc_rx_process_limit = WM_RX_PROCESS_LIMIT_DEFAULT;
sc->sc_rx_intr_process_limit = WM_RX_INTR_PROCESS_LIMIT_DEFAULT;
@@ -2932,36 +2953,57 @@
{
int qid;
struct wm_softc *sc = ifp->if_softc;
+ uint16_t hang_queue = 0; /* Max queue number of wm(4) is 82576's 16. */
for (qid = 0; qid < sc->sc_nqueues; qid++) {
struct wm_txqueue *txq = &sc->sc_queue[qid].wmq_txq;
- wm_watchdog_txq(ifp, txq);
- }
-
- /* Reset the interface. */
- (void) wm_init(ifp);
-
- /*
- * There are still some upper layer processing which call
- * ifp->if_start(). e.g. ALTQ or one CPU system
- */
- /* Try to get more packets going. */
- ifp->if_start(ifp);
-}
-
-static void
-wm_watchdog_txq(struct ifnet *ifp, struct wm_txqueue *txq)
+ wm_watchdog_txq(ifp, txq, &hang_queue);
+ }
+
+ /*
+ * IF any of queues hanged up, reset the interface.
+ */
+ if (hang_queue != 0) {
+ (void) wm_init(ifp);
+
+ /*
+ * There are still some upper layer processing which call
+ * ifp->if_start(). e.g. ALTQ or one CPU system
+ */
+ /* Try to get more packets going. */
+ ifp->if_start(ifp);
+ }
+}
+
+
+static void
+wm_watchdog_txq(struct ifnet *ifp, struct wm_txqueue *txq, uint16_t *hang)
+{
+
+ mutex_enter(txq->txq_lock);
+ if (txq->txq_watchdog &&
+ time_uptime - txq->txq_lastsent > wm_watchdog_timeout) {
+ wm_watchdog_txq_locked(ifp, txq, hang);
+ }
+ mutex_exit(txq->txq_lock);
+}
+
+static void
+wm_watchdog_txq_locked(struct ifnet *ifp, struct wm_txqueue *txq, uint16_t *hang)
{
struct wm_softc *sc = ifp->if_softc;
+ struct wm_queue *wmq = container_of(txq, struct wm_queue, wmq_txq);
+
+ KASSERT(mutex_owned(txq->txq_lock));
/*
* Since we're using delayed interrupts, sweep up
* before we report an error.
*/
- mutex_enter(txq->txq_lock);
- wm_txeof(sc, txq);
- mutex_exit(txq->txq_lock);
+ wm_txeof(txq, UINT_MAX);
+ if (txq->txq_watchdog)
+ *hang |= __BIT(wmq->wmq_id);
if (txq->txq_free != WM_NTXDESC(txq)) {
#ifdef WM_DEBUG
@@ -2981,11 +3023,22 @@
i, txs->txs_firstdesc, txs->txs_lastdesc);
for (j = txs->txs_firstdesc; ;
j = WM_NEXTTX(txq, j)) {
- printf("\tdesc %d: 0x%" PRIx64 "\n", j,
- txq->txq_nq_descs[j].nqtx_data.nqtxd_addr);
- printf("\t %#08x%08x\n",
- txq->txq_nq_descs[j].nqtx_data.nqtxd_fields,
- txq->txq_nq_descs[j].nqtx_data.nqtxd_cmdlen);
+ if ((sc->sc_flags & WM_F_NEWQUEUE) != 0) {
+ printf("\tdesc %d: 0x%" PRIx64 "\n", j,
+ txq->txq_nq_descs[j].nqtx_data.nqtxd_addr);
+ printf("\t %#08x%08x\n",
+ txq->txq_nq_descs[j].nqtx_data.nqtxd_fields,
+ txq->txq_nq_descs[j].nqtx_data.nqtxd_cmdlen);
+ } else {
+ printf("\tdesc %d: 0x%" PRIx64 "\n", j,
+ (uint64_t)txq->txq_descs[j].wtx_addr.wa_high << 32 |
+ txq->txq_descs[j].wtx_addr.wa_low);
+ printf("\t %#04x%02x%02x%08x\n",
+ txq->txq_descs[j].wtx_fields.wtxu_vlan,
+ txq->txq_descs[j].wtx_fields.wtxu_options,
+ txq->txq_descs[j].wtx_fields.wtxu_status,
+ txq->txq_descs[j].wtx_cmdlen);
+ }
if (j == txs->txs_lastdesc)
break;
}
@@ -3011,8 +3064,13 @@
WM_CORE_LOCK(sc);
- if (sc->sc_core_stopping)
- goto out;
+ if (sc->sc_core_stopping) {
+ WM_CORE_UNLOCK(sc);
+#ifndef WM_MPSAFE
+ splx(s);
+#endif
+ return;
+ }
if (sc->sc_type >= WM_T_82542_2_1) {
WM_EVCNT_ADD(&sc->sc_ev_rx_xon, CSR_READ(sc, WMREG_XONRXC));
@@ -3050,12 +3108,11 @@
else
wm_tbi_tick(sc);
+ WM_CORE_UNLOCK(sc);
+
+ wm_watchdog(ifp);
+
callout_reset(&sc->sc_tick_ch, hz, wm_tick, sc);
-out:
- WM_CORE_UNLOCK(sc);
-#ifndef WM_MPSAFE
- splx(s);
-#endif
}
static int
@@ -4199,6 +4256,10 @@
wm_phy_post_reset(sc);
}
+/*
+ * Only used by WM_T_PCH_SPT which does not use multiqueue,
+ * so it is enough to check sc->sc_queue[0] only.
+ */
static void
wm_flush_desc_rings(struct wm_softc *sc)
{
@@ -5939,6 +6000,7 @@
struct wm_queue *wmq = &sc->sc_queue[qidx];
struct wm_txqueue *txq = &wmq->wmq_txq;
mutex_enter(txq->txq_lock);
+ txq->txq_watchdog = false; /* ensure watchdog disabled */
Home |
Main Index |
Thread Index |
Old Index