summaryrefslogtreecommitdiffstats
path: root/drivers/infiniband/hw/qib
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/infiniband/hw/qib')
-rw-r--r--drivers/infiniband/hw/qib/qib.h15
-rw-r--r--drivers/infiniband/hw/qib/qib_driver.c20
-rw-r--r--drivers/infiniband/hw/qib/qib_file_ops.c2
-rw-r--r--drivers/infiniband/hw/qib/qib_iba6120.c2
-rw-r--r--drivers/infiniband/hw/qib/qib_iba7220.c2
-rw-r--r--drivers/infiniband/hw/qib/qib_iba7322.c135
-rw-r--r--drivers/infiniband/hw/qib/qib_init.c8
-rw-r--r--drivers/infiniband/hw/qib/qib_qp.c90
-rw-r--r--drivers/infiniband/hw/qib/qib_qsfp.c25
-rw-r--r--drivers/infiniband/hw/qib/qib_qsfp.h3
-rw-r--r--drivers/infiniband/hw/qib/qib_rc.c36
-rw-r--r--drivers/infiniband/hw/qib/qib_ruc.c7
-rw-r--r--drivers/infiniband/hw/qib/qib_srq.c5
-rw-r--r--drivers/infiniband/hw/qib/qib_sysfs.c3
-rw-r--r--drivers/infiniband/hw/qib/qib_uc.c25
-rw-r--r--drivers/infiniband/hw/qib/qib_verbs.c36
-rw-r--r--drivers/infiniband/hw/qib/qib_verbs.h5
17 files changed, 276 insertions, 143 deletions
diff --git a/drivers/infiniband/hw/qib/qib.h b/drivers/infiniband/hw/qib/qib.h
index c9624ea8720..b881bdc401f 100644
--- a/drivers/infiniband/hw/qib/qib.h
+++ b/drivers/infiniband/hw/qib/qib.h
@@ -171,7 +171,9 @@ struct qib_ctxtdata {
/* how many alloc_pages() chunks in rcvegrbuf_pages */
u32 rcvegrbuf_chunks;
/* how many egrbufs per chunk */
- u32 rcvegrbufs_perchunk;
+ u16 rcvegrbufs_perchunk;
+ /* ilog2 of above */
+ u16 rcvegrbufs_perchunk_shift;
/* order for rcvegrbuf_pages */
size_t rcvegrbuf_size;
/* rcvhdrq size (for freeing) */
@@ -221,6 +223,9 @@ struct qib_ctxtdata {
/* ctxt rcvhdrq head offset */
u32 head;
u32 pkt_count;
+ /* lookaside fields */
+ struct qib_qp *lookaside_qp;
+ u32 lookaside_qpn;
/* QPs waiting for context processing */
struct list_head qp_wait_list;
};
@@ -807,6 +812,10 @@ struct qib_devdata {
* supports, less gives more pio bufs/ctxt, etc.
*/
u32 cfgctxts;
+ /*
+ * number of ctxts available for PSM open
+ */
+ u32 freectxts;
/*
* hint that we should update pioavailshadow before
@@ -936,7 +945,9 @@ struct qib_devdata {
/* chip address space used by 4k pio buffers */
u32 align4k;
/* size of each rcvegrbuffer */
- u32 rcvegrbufsize;
+ u16 rcvegrbufsize;
+ /* log2 of above */
+ u16 rcvegrbufsize_shift;
/* localbus width (1, 2,4,8,16,32) from config space */
u32 lbus_width;
/* localbus speed in MHz */
diff --git a/drivers/infiniband/hw/qib/qib_driver.c b/drivers/infiniband/hw/qib/qib_driver.c
index 23e584f4c36..9a9047f385a 100644
--- a/drivers/infiniband/hw/qib/qib_driver.c
+++ b/drivers/infiniband/hw/qib/qib_driver.c
@@ -279,10 +279,10 @@ bail:
*/
static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
{
- const u32 chunk = etail / rcd->rcvegrbufs_perchunk;
- const u32 idx = etail % rcd->rcvegrbufs_perchunk;
+ const u32 chunk = etail >> rcd->rcvegrbufs_perchunk_shift;
+ const u32 idx = etail & ((u32)rcd->rcvegrbufs_perchunk - 1);
- return rcd->rcvegrbuf[chunk] + idx * rcd->dd->rcvegrbufsize;
+ return rcd->rcvegrbuf[chunk] + (idx << rcd->dd->rcvegrbufsize_shift);
}
/*
@@ -310,7 +310,6 @@ static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
u32 opcode;
u32 psn;
int diff;
- unsigned long flags;
/* Sanity check packet */
if (tlen < 24)
@@ -365,7 +364,6 @@ static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
switch (qp->ibqp.qp_type) {
case IB_QPT_RC:
- spin_lock_irqsave(&qp->s_lock, flags);
ruc_res =
qib_ruc_check_hdr(
ibp, hdr,
@@ -373,11 +371,8 @@ static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
qp,
be32_to_cpu(ohdr->bth[0]));
if (ruc_res) {
- spin_unlock_irqrestore(&qp->s_lock,
- flags);
goto unlock;
}
- spin_unlock_irqrestore(&qp->s_lock, flags);
/* Only deal with RDMA Writes for now */
if (opcode <
@@ -547,6 +542,15 @@ move_along:
updegr = 0;
}
}
+ /*
+ * Notify qib_destroy_qp() if it is waiting
+ * for lookaside_qp to finish.
+ */
+ if (rcd->lookaside_qp) {
+ if (atomic_dec_and_test(&rcd->lookaside_qp->refcount))
+ wake_up(&rcd->lookaside_qp->wait);
+ rcd->lookaside_qp = NULL;
+ }
rcd->head = l;
rcd->pkt_count += i;
diff --git a/drivers/infiniband/hw/qib/qib_file_ops.c b/drivers/infiniband/hw/qib/qib_file_ops.c
index 26253039d2c..77633666f81 100644
--- a/drivers/infiniband/hw/qib/qib_file_ops.c
+++ b/drivers/infiniband/hw/qib/qib_file_ops.c
@@ -1284,6 +1284,7 @@ static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
strlcpy(rcd->comm, current->comm, sizeof(rcd->comm));
ctxt_fp(fp) = rcd;
qib_stats.sps_ctxts++;
+ dd->freectxts++;
ret = 0;
goto bail;
@@ -1792,6 +1793,7 @@ static int qib_close(struct inode *in, struct file *fp)
if (dd->pageshadow)
unlock_expected_tids(rcd);
qib_stats.sps_ctxts--;
+ dd->freectxts--;
}
mutex_unlock(&qib_mutex);
diff --git a/drivers/infiniband/hw/qib/qib_iba6120.c b/drivers/infiniband/hw/qib/qib_iba6120.c
index d8ca0a0b970..781a802a321 100644
--- a/drivers/infiniband/hw/qib/qib_iba6120.c
+++ b/drivers/infiniband/hw/qib/qib_iba6120.c
@@ -3273,6 +3273,8 @@ static int init_6120_variables(struct qib_devdata *dd)
/* we always allocate at least 2048 bytes for eager buffers */
ret = ib_mtu_enum_to_int(qib_ibmtu);
dd->rcvegrbufsize = ret != -1 ? max(ret, 2048) : QIB_DEFAULT_MTU;
+ BUG_ON(!is_power_of_2(dd->rcvegrbufsize));
+ dd->rcvegrbufsize_shift = ilog2(dd->rcvegrbufsize);
qib_6120_tidtemplate(dd);
diff --git a/drivers/infiniband/hw/qib/qib_iba7220.c b/drivers/infiniband/hw/qib/qib_iba7220.c
index e1f947446c2..3f1d562ba89 100644
--- a/drivers/infiniband/hw/qib/qib_iba7220.c
+++ b/drivers/infiniband/hw/qib/qib_iba7220.c
@@ -4085,6 +4085,8 @@ static int qib_init_7220_variables(struct qib_devdata *dd)
/* we always allocate at least 2048 bytes for eager buffers */
ret = ib_mtu_enum_to_int(qib_ibmtu);
dd->rcvegrbufsize = ret != -1 ? max(ret, 2048) : QIB_DEFAULT_MTU;
+ BUG_ON(!is_power_of_2(dd->rcvegrbufsize));
+ dd->rcvegrbufsize_shift = ilog2(dd->rcvegrbufsize);
qib_7220_tidtemplate(dd);
diff --git a/drivers/infiniband/hw/qib/qib_iba7322.c b/drivers/infiniband/hw/qib/qib_iba7322.c
index 5ea9ece23b3..efd0a110091 100644
--- a/drivers/infiniband/hw/qib/qib_iba7322.c
+++ b/drivers/infiniband/hw/qib/qib_iba7322.c
@@ -2310,12 +2310,15 @@ static int qib_7322_bringup_serdes(struct qib_pportdata *ppd)
val = ppd->cpspec->ibcctrl_a | (QLOGIC_IB_IBCC_LINKINITCMD_DISABLE <<
QLOGIC_IB_IBCC_LINKINITCMD_SHIFT);
+ ppd->cpspec->ibcctrl_a = val;
/*
* Reset the PCS interface to the serdes (and also ibc, which is still
* in reset from above). Writes new value of ibcctrl_a as last step.
*/
qib_7322_mini_pcs_reset(ppd);
qib_write_kreg(dd, kr_scratch, 0ULL);
+ /* clear the linkinit cmds */
+ ppd->cpspec->ibcctrl_a &= ~SYM_MASK(IBCCtrlA_0, LinkInitCmd);
if (!ppd->cpspec->ibcctrl_b) {
unsigned lse = ppd->link_speed_enabled;
@@ -2387,11 +2390,6 @@ static int qib_7322_bringup_serdes(struct qib_pportdata *ppd)
qib_write_kreg_port(ppd, krp_rcvctrl, ppd->p_rcvctrl);
spin_unlock_irqrestore(&dd->cspec->rcvmod_lock, flags);
- /* Hold the link state machine for mezz boards */
- if (IS_QMH(dd) || IS_QME(dd))
- qib_set_ib_7322_lstate(ppd, 0,
- QLOGIC_IB_IBCC_LINKINITCMD_DISABLE);
-
/* Also enable IBSTATUSCHG interrupt. */
val = qib_read_kreg_port(ppd, krp_errmask);
qib_write_kreg_port(ppd, krp_errmask,
@@ -2853,9 +2851,8 @@ static irqreturn_t qib_7322intr(int irq, void *data)
for (i = 0; i < dd->first_user_ctxt; i++) {
if (ctxtrbits & rmask) {
ctxtrbits &= ~rmask;
- if (dd->rcd[i]) {
+ if (dd->rcd[i])
qib_kreceive(dd->rcd[i], NULL, &npkts);
- }
}
rmask <<= 1;
}
@@ -5230,6 +5227,8 @@ static int qib_7322_ib_updown(struct qib_pportdata *ppd, int ibup, u64 ibcs)
QIBL_IB_AUTONEG_INPROG)))
set_7322_ibspeed_fast(ppd, ppd->link_speed_enabled);
if (!(ppd->lflags & QIBL_IB_AUTONEG_INPROG)) {
+ struct qib_qsfp_data *qd =
+ &ppd->cpspec->qsfp_data;
/* unlock the Tx settings, speed may change */
qib_write_kreg_port(ppd, krp_tx_deemph_override,
SYM_MASK(IBSD_TX_DEEMPHASIS_OVERRIDE_0,
@@ -5237,6 +5236,12 @@ static int qib_7322_ib_updown(struct qib_pportdata *ppd, int ibup, u64 ibcs)
qib_cancel_sends(ppd);
/* on link down, ensure sane pcs state */
qib_7322_mini_pcs_reset(ppd);
+ /* schedule the qsfp refresh which should turn the link
+ off */
+ if (ppd->dd->flags & QIB_HAS_QSFP) {
+ qd->t_insert = get_jiffies_64();
+ schedule_work(&qd->work);
+ }
spin_lock_irqsave(&ppd->sdma_lock, flags);
if (__qib_sdma_running(ppd))
__qib_sdma_process_event(ppd,
@@ -5587,43 +5592,79 @@ static void qsfp_7322_event(struct work_struct *work)
struct qib_qsfp_data *qd;
struct qib_pportdata *ppd;
u64 pwrup;
+ unsigned long flags;
int ret;
u32 le2;
qd = container_of(work, struct qib_qsfp_data, work);
ppd = qd->ppd;
- pwrup = qd->t_insert + msecs_to_jiffies(QSFP_PWR_LAG_MSEC);
+ pwrup = qd->t_insert +
+ msecs_to_jiffies(QSFP_PWR_LAG_MSEC - QSFP_MODPRS_LAG_MSEC);
- /*
- * Some QSFP's not only do not respond until the full power-up
- * time, but may behave badly if we try. So hold off responding
- * to insertion.
- */
- while (1) {
- u64 now = get_jiffies_64();
- if (time_after64(now, pwrup))
- break;
- msleep(20);
- }
- ret = qib_refresh_qsfp_cache(ppd, &qd->cache);
- /*
- * Need to change LE2 back to defaults if we couldn't
- * read the cable type (to handle cable swaps), so do this
- * even on failure to read cable information. We don't
- * get here for QME, so IS_QME check not needed here.
- */
- if (!ret && !ppd->dd->cspec->r1) {
- if (QSFP_IS_ACTIVE_FAR(qd->cache.tech))
- le2 = LE2_QME;
- else if (qd->cache.atten[1] >= qib_long_atten &&
- QSFP_IS_CU(qd->cache.tech))
- le2 = LE2_5m;
- else
+ /* Delay for 20 msecs to allow ModPrs resistor to setup */
+ mdelay(QSFP_MODPRS_LAG_MSEC);
+
+ if (!qib_qsfp_mod_present(ppd)) {
+ ppd->cpspec->qsfp_data.modpresent = 0;
+ /* Set the physical link to disabled */
+ qib_set_ib_7322_lstate(ppd, 0,
+ QLOGIC_IB_IBCC_LINKINITCMD_DISABLE);
+ spin_lock_irqsave(&ppd->lflags_lock, flags);
+ ppd->lflags &= ~QIBL_LINKV;
+ spin_unlock_irqrestore(&ppd->lflags_lock, flags);
+ } else {
+ /*
+ * Some QSFP's not only do not respond until the full power-up
+ * time, but may behave badly if we try. So hold off responding
+ * to insertion.
+ */
+ while (1) {
+ u64 now = get_jiffies_64();
+ if (time_after64(now, pwrup))
+ break;
+ msleep(20);
+ }
+
+ ret = qib_refresh_qsfp_cache(ppd, &qd->cache);
+
+ /*
+ * Need to change LE2 back to defaults if we couldn't
+ * read the cable type (to handle cable swaps), so do this
+ * even on failure to read cable information. We don't
+ * get here for QME, so IS_QME check not needed here.
+ */
+ if (!ret && !ppd->dd->cspec->r1) {
+ if (QSFP_IS_ACTIVE_FAR(qd->cache.tech))
+ le2 = LE2_QME;
+ else if (qd->cache.atten[1] >= qib_long_atten &&
+ QSFP_IS_CU(qd->cache.tech))
+ le2 = LE2_5m;
+ else
+ le2 = LE2_DEFAULT;
+ } else
le2 = LE2_DEFAULT;
- } else
- le2 = LE2_DEFAULT;
- ibsd_wr_allchans(ppd, 13, (le2 << 7), BMASK(9, 7));
- init_txdds_table(ppd, 0);
+ ibsd_wr_allchans(ppd, 13, (le2 << 7), BMASK(9, 7));
+ /*
+ * We always change parameteters, since we can choose
+ * values for cables without eeproms, and the cable may have
+ * changed from a cable with full or partial eeprom content
+ * to one with partial or no content.
+ */
+ init_txdds_table(ppd, 0);
+ /* The physical link is being re-enabled only when the
+ * previous state was DISABLED and the VALID bit is not
+ * set. This should only happen when the cable has been
+ * physically pulled. */
+ if (!ppd->cpspec->qsfp_data.modpresent &&
+ (ppd->lflags & (QIBL_LINKV | QIBL_IB_LINK_DISABLED))) {
+ ppd->cpspec->qsfp_data.modpresent = 1;
+ qib_set_ib_7322_lstate(ppd, 0,
+ QLOGIC_IB_IBCC_LINKINITCMD_SLEEP);
+ spin_lock_irqsave(&ppd->lflags_lock, flags);
+ ppd->lflags |= QIBL_LINKV;
+ spin_unlock_irqrestore(&ppd->lflags_lock, flags);
+ }
+ }
}
/*
@@ -5727,7 +5768,8 @@ static void set_no_qsfp_atten(struct qib_devdata *dd, int change)
/* now change the IBC and serdes, overriding generic */
init_txdds_table(ppd, 1);
/* Re-enable the physical state machine on mezz boards
- * now that the correct settings have been set. */
+ * now that the correct settings have been set.
+ * QSFP boards are handles by the QSFP event handler */
if (IS_QMH(dd) || IS_QME(dd))
qib_set_ib_7322_lstate(ppd, 0,
QLOGIC_IB_IBCC_LINKINITCMD_SLEEP);
@@ -6205,6 +6247,8 @@ static int qib_init_7322_variables(struct qib_devdata *dd)
/* we always allocate at least 2048 bytes for eager buffers */
dd->rcvegrbufsize = max(mtu, 2048);
+ BUG_ON(!is_power_of_2(dd->rcvegrbufsize));
+ dd->rcvegrbufsize_shift = ilog2(dd->rcvegrbufsize);
qib_7322_tidtemplate(dd);
@@ -7147,7 +7191,8 @@ static void find_best_ent(struct qib_pportdata *ppd,
}
}
- /* Lookup serdes setting by cable type and attenuation */
+ /* Active cables don't have attenuation so we only set SERDES
+ * settings to account for the attenuation of the board traces. */
if (!override && QSFP_IS_ACTIVE(qd->tech)) {
*sdr_dds = txdds_sdr + ppd->dd->board_atten;
*ddr_dds = txdds_ddr + ppd->dd->board_atten;
@@ -7464,12 +7509,6 @@ static int serdes_7322_init_new(struct qib_pportdata *ppd)
u32 le_val, rxcaldone;
int chan, chan_done = (1 << SERDES_CHANS) - 1;
- /*
- * Initialize the Tx DDS tables. Also done every QSFP event,
- * for adapters with QSFP
- */
- init_txdds_table(ppd, 0);
-
/* Clear cmode-override, may be set from older driver */
ahb_mod(ppd->dd, IBSD(ppd->hw_pidx), 5, 10, 0 << 14, 1 << 14);
@@ -7655,6 +7694,12 @@ static int serdes_7322_init_new(struct qib_pportdata *ppd)
/* VGA output common mode */
ibsd_wr_allchans(ppd, 12, (3 << 2), BMASK(3, 2));
+ /*
+ * Initialize the Tx DDS tables. Also done every QSFP event,
+ * for adapters with QSFP
+ */
+ init_txdds_table(ppd, 0);
+
return 0;
}
diff --git a/drivers/infiniband/hw/qib/qib_init.c b/drivers/infiniband/hw/qib/qib_init.c
index a01f3fce8eb..b093a0b53b2 100644
--- a/drivers/infiniband/hw/qib/qib_init.c
+++ b/drivers/infiniband/hw/qib/qib_init.c
@@ -183,6 +183,9 @@ struct qib_ctxtdata *qib_create_ctxtdata(struct qib_pportdata *ppd, u32 ctxt)
rcd->rcvegrbuf_chunks = (rcd->rcvegrcnt +
rcd->rcvegrbufs_perchunk - 1) /
rcd->rcvegrbufs_perchunk;
+ BUG_ON(!is_power_of_2(rcd->rcvegrbufs_perchunk));
+ rcd->rcvegrbufs_perchunk_shift =
+ ilog2(rcd->rcvegrbufs_perchunk);
}
return rcd;
}
@@ -398,6 +401,7 @@ static void enable_chip(struct qib_devdata *dd)
if (rcd)
dd->f_rcvctrl(rcd->ppd, rcvmask, i);
}
+ dd->freectxts = dd->cfgctxts - dd->first_user_ctxt;
}
static void verify_interrupt(unsigned long opaque)
@@ -581,10 +585,6 @@ int qib_init(struct qib_devdata *dd, int reinit)
continue;
}
- /* let link come up, and enable IBC */
- spin_lock_irqsave(&ppd->lflags_lock, flags);
- ppd->lflags &= ~QIBL_IB_LINK_DISABLED;
- spin_unlock_irqrestore(&ppd->lflags_lock, flags);
portok++;
}
diff --git a/drivers/infiniband/hw/qib/qib_qp.c b/drivers/infiniband/hw/qib/qib_qp.c
index e16751f8639..7e7e16fbee9 100644
--- a/drivers/infiniband/hw/qib/qib_qp.c
+++ b/drivers/infiniband/hw/qib/qib_qp.c
@@ -34,6 +34,7 @@
#include <linux/err.h>
#include <linux/vmalloc.h>
+#include <linux/jhash.h>
#include "qib.h"
@@ -204,6 +205,13 @@ static void free_qpn(struct qib_qpn_table *qpt, u32 qpn)
clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
}
+static inline unsigned qpn_hash(struct qib_ibdev *dev, u32 qpn)
+{
+ return jhash_1word(qpn, dev->qp_rnd) &
+ (dev->qp_table_size - 1);
+}
+
+
/*
* Put the QP into the hash table.
* The hash table holds a reference to the QP.
@@ -211,22 +219,23 @@ static void free_qpn(struct qib_qpn_table *qpt, u32 qpn)
static void insert_qp(struct qib_ibdev *dev, struct qib_qp *qp)
{
struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
- unsigned n = qp->ibqp.qp_num % dev->qp_table_size;
unsigned long flags;
+ unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
spin_lock_irqsave(&dev->qpt_lock, flags);
+ atomic_inc(&qp->refcount);
if (qp->ibqp.qp_num == 0)
- ibp->qp0 = qp;
+ rcu_assign_pointer(ibp->qp0, qp);
else if (qp->ibqp.qp_num == 1)
- ibp->qp1 = qp;
+ rcu_assign_pointer(ibp->qp1, qp);
else {
qp->next = dev->qp_table[n];
- dev->qp_table[n] = qp;
+ rcu_assign_pointer(dev->qp_table[n], qp);
}
- atomic_inc(&qp->refcount);
spin_unlock_irqrestore(&dev->qpt_lock, flags);
+ synchronize_rcu();
}
/*
@@ -236,29 +245,32 @@ static void insert_qp(struct qib_ibdev *dev, struct qib_qp *qp)
static void remove_qp(struct qib_ibdev *dev, struct qib_qp *qp)
{
struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
- struct qib_qp *q, **qpp;
+ unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
unsigned long flags;
- qpp = &dev->qp_table[qp->ibqp.qp_num % dev->qp_table_size];
-
spin_lock_irqsave(&dev->qpt_lock, flags);
if (ibp->qp0 == qp) {
- ibp->qp0 = NULL;
atomic_dec(&qp->refcount);
+ rcu_assign_pointer(ibp->qp0, NULL);
} else if (ibp->qp1 == qp) {
- ibp->qp1 = NULL;
atomic_dec(&qp->refcount);
- } else
+ rcu_assign_pointer(ibp->qp1, NULL);
+ } else {
+ struct qib_qp *q, **qpp;
+
+ qpp = &dev->qp_table[n];
for (; (q = *qpp) != NULL; qpp = &q->next)
if (q == qp) {
- *qpp = qp->next;
- qp->next = NULL;
atomic_dec(&qp->refcount);
+ rcu_assign_pointer(*qpp, qp->next);
+ qp->next = NULL;
break;
}
+ }
spin_unlock_irqrestore(&dev->qpt_lock, flags);
+ synchronize_rcu();
}
/**
@@ -280,21 +292,24 @@ unsigned qib_free_all_qps(struct qib_devdata *dd)
if (!qib_mcast_tree_empty(ibp))
qp_inuse++;
- if (ibp->qp0)
+ rcu_read_lock();
+ if (rcu_dereference(ibp->qp0))
qp_inuse++;
- if (ibp->qp1)
+ if (rcu_dereference(ibp->qp1))
qp_inuse++;
+ rcu_read_unlock();
}
spin_lock_irqsave(&dev->qpt_lock, flags);
for (n = 0; n < dev->qp_table_size; n++) {
qp = dev->qp_table[n];
- dev->qp_table[n] = NULL;
+ rcu_assign_pointer(dev->qp_table[n], NULL);
for (; qp; qp = qp->next)
qp_inuse++;
}
spin_unlock_irqrestore(&dev->qpt_lock, flags);
+ synchronize_rcu();
return qp_inuse;
}
@@ -309,25 +324,28 @@ unsigned qib_free_all_qps(struct qib_devdata *dd)
*/
struct qib_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn)
{
- struct qib_ibdev *dev = &ppd_from_ibp(ibp)->dd->verbs_dev;
- unsigned long flags;
- struct qib_qp *qp;
+ struct qib_qp *qp = NULL;
- spin_lock_irqsave(&dev->qpt_lock, flags);
+ if (unlikely(qpn <= 1)) {
+ rcu_read_lock();
+ if (qpn == 0)
+ qp = rcu_dereference(ibp->qp0);
+ else
+ qp = rcu_dereference(ibp->qp1);
+ } else {
+ struct qib_ibdev *dev = &ppd_from_ibp(ibp)->dd->verbs_dev;
+ unsigned n = qpn_hash(dev, qpn);
- if (qpn == 0)
- qp = ibp->qp0;
- else if (qpn == 1)
- qp = ibp->qp1;
- else
- for (qp = dev->qp_table[qpn % dev->qp_table_size]; qp;
- qp = qp->next)
+ rcu_read_lock();
+ for (qp = dev->qp_table[n]; rcu_dereference(qp); qp = qp->next)
if (qp->ibqp.qp_num == qpn)
break;
+ }
if (qp)
- atomic_inc(&qp->refcount);
+ if (unlikely(!atomic_inc_not_zero(&qp->refcount)))
+ qp = NULL;
- spin_unlock_irqrestore(&dev->qpt_lock, flags);
+ rcu_read_unlock();
return qp;
}
@@ -765,8 +783,10 @@ int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
}
}
- if (attr_mask & IB_QP_PATH_MTU)
+ if (attr_mask & IB_QP_PATH_MTU) {
qp->path_mtu = pmtu;
+ qp->pmtu = ib_mtu_enum_to_int(pmtu);
+ }
if (attr_mask & IB_QP_RETRY_CNT) {
qp->s_retry_cnt = attr->retry_cnt;
@@ -781,8 +801,12 @@ int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
if (attr_mask & IB_QP_MIN_RNR_TIMER)
qp->r_min_rnr_timer = attr->min_rnr_timer;
- if (attr_mask & IB_QP_TIMEOUT)
+ if (attr_mask & IB_QP_TIMEOUT) {
qp->timeout = attr->timeout;
+ qp->timeout_jiffies =
+ usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
+ 1000UL);
+ }
if (attr_mask & IB_QP_QKEY)
qp->qkey = attr->qkey;
@@ -1013,6 +1037,10 @@ struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
ret = ERR_PTR(-ENOMEM);
goto bail_swq;
}
+ RCU_INIT_POINTER(qp->next, NULL);
+ qp->timeout_jiffies =
+ usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
+ 1000UL);
if (init_attr->srq)
sz = 0;
else {
diff --git a/drivers/infiniband/hw/qib/qib_qsfp.c b/drivers/infiniband/hw/qib/qib_qsfp.c
index 3374a52232c..e06c4ed383f 100644
--- a/drivers/infiniband/hw/qib/qib_qsfp.c
+++ b/drivers/infiniband/hw/qib/qib_qsfp.c
@@ -273,18 +273,12 @@ int qib_refresh_qsfp_cache(struct qib_pportdata *ppd, struct qib_qsfp_cache *cp)
int ret;
int idx;
u16 cks;
- u32 mask;
u8 peek[4];
/* ensure sane contents on invalid reads, for cable swaps */
memset(cp, 0, sizeof(*cp));
- mask = QSFP_GPIO_MOD_PRS_N;
- if (ppd->hw_pidx)
- mask <<= QSFP_GPIO_PORT2_SHIFT;
-
- ret = ppd->dd->f_gpio_mod(ppd->dd, 0, 0, 0);
- if (ret & mask) {
+ if (!qib_qsfp_mod_present(ppd)) {
ret = -ENODEV;
goto bail;
}
@@ -444,6 +438,19 @@ const char * const qib_qsfp_devtech[16] = {
static const char *pwr_codes = "1.5W2.0W2.5W3.5W";
+int qib_qsfp_mod_present(struct qib_pportdata *ppd)
+{
+ u32 mask;
+ int ret;
+
+ mask = QSFP_GPIO_MOD_PRS_N <<
+ (ppd->hw_pidx * QSFP_GPIO_PORT2_SHIFT);
+ ret = ppd->dd->f_gpio_mod(ppd->dd, 0, 0, 0);
+
+ return !((ret & mask) >>
+ ((ppd->hw_pidx * QSFP_GPIO_PORT2_SHIFT) + 3));
+}
+
/*
* Initialize structures that control access to QSFP. Called once per port
* on cards that support QSFP.
@@ -452,7 +459,6 @@ void qib_qsfp_init(struct qib_qsfp_data *qd,
void (*fevent)(struct work_struct *))
{
u32 mask, highs;
- int pins;
struct qib_devdata *dd = qd->ppd->dd;
@@ -480,8 +486,7 @@ void qib_qsfp_init(struct qib_qsfp_data *qd,
mask <<= QSFP_GPIO_PORT2_SHIFT;
/* Do not try to wait here. Better to let event handle it */
- pins = dd->f_gpio_mod(dd, 0, 0, 0);
- if (pins & mask)
+ if (!qib_qsfp_mod_present(qd->ppd))
goto bail;
/* We see a module, but it may be unwise to look yet. Just schedule */
qd->t_insert = get_jiffies_64();
diff --git a/drivers/infiniband/hw/qib/qib_qsfp.h b/drivers/infiniband/hw/qib/qib_qsfp.h
index c109bbdc90a..46002a9417c 100644
--- a/drivers/infiniband/hw/qib/qib_qsfp.h
+++ b/drivers/infiniband/hw/qib/qib_qsfp.h
@@ -34,6 +34,7 @@
#define QSFP_DEV 0xA0
#define QSFP_PWR_LAG_MSEC 2000
+#define QSFP_MODPRS_LAG_MSEC 20
/*
* Below are masks for various QSFP signals, for Port 1.
@@ -177,10 +178,12 @@ struct qib_qsfp_data {
struct work_struct work;
struct qib_qsfp_cache cache;
u64 t_insert;
+ u8 modpresent;
};
extern int qib_refresh_qsfp_cache(struct qib_pportdata *ppd,
struct qib_qsfp_cache *cp);
+extern int qib_qsfp_mod_present(struct qib_pportdata *ppd);
extern void qib_qsfp_init(struct qib_qsfp_data *qd,
void (*fevent)(struct work_struct *));
extern void qib_qsfp_deinit(struct qib_qsfp_data *qd);
diff --git a/drivers/infiniband/hw/qib/qib_rc.c b/drivers/infiniband/hw/qib/qib_rc.c
index eca0c41f122..afaf4ac79f4 100644
--- a/drivers/infiniband/hw/qib/qib_rc.c
+++ b/drivers/infiniband/hw/qib/qib_rc.c
@@ -59,8 +59,7 @@ static void start_timer(struct qib_qp *qp)
qp->s_flags |= QIB_S_TIMER;
qp->s_timer.function = rc_timeout;
/* 4.096 usec. * (1 << qp->timeout) */
- qp->s_timer.expires = jiffies +
- usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / 1000UL);
+ qp->s_timer.expires = jiffies + qp->timeout_jiffies;
add_timer(&qp->s_timer);
}
@@ -239,7 +238,7 @@ int qib_make_rc_req(struct qib_qp *qp)
u32 len;
u32 bth0;
u32 bth2;
- u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
+ u32 pmtu = qp->pmtu;
char newreq;
unsigned long flags;
int ret = 0;
@@ -1519,9 +1518,7 @@ read_middle:
* 4.096 usec. * (1 << qp->timeout)
*/
qp->s_flags |= QIB_S_TIMER;
- mod_timer(&qp->s_timer, jiffies +
- usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
- 1000UL));
+ mod_timer(&qp->s_timer, jiffies + qp->timeout_jiffies);
if (qp->s_flags & QIB_S_WAIT_ACK) {
qp->s_flags &= ~QIB_S_WAIT_ACK;
qib_schedule_send(qp);
@@ -1732,7 +1729,7 @@ static int qib_rc_rcv_error(struct qib_other_headers *ohdr,
* same request.
*/
offset = ((psn - e->psn) & QIB_PSN_MASK) *
- ib_mtu_enum_to_int(qp->path_mtu);
+ qp->pmtu;
len = be32_to_cpu(reth->length);
if (unlikely(offset + len != e->rdma_sge.sge_length))
goto unlock_done;
@@ -1876,7 +1873,7 @@ void qib_rc_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr,
u32 psn;
u32 pad;
struct ib_wc wc;
- u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
+ u32 pmtu = qp->pmtu;
int diff;
struct ib_reth *reth;
unsigned long flags;
@@ -1892,10 +1889,8 @@ void qib_rc_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr,
}
opcode = be32_to_cpu(ohdr->bth[0]);
- spin_lock_irqsave(&qp->s_lock, flags);
if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode))
- goto sunlock;
- spin_unlock_irqrestore(&qp->s_lock, flags);
+ return;
psn = be32_to_cpu(ohdr->bth[2]);
opcode >>= 24;
@@ -1955,8 +1950,6 @@ void qib_rc_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr,
break;
}
- memset(&wc, 0, sizeof wc);
-
if (qp->state == IB_QPS_RTR && !(qp->r_flags & QIB_R_COMM_EST)) {
qp->r_flags |= QIB_R_COMM_EST;
if (qp->ibqp.event_handler) {
@@ -2009,16 +2002,19 @@ send_middle:
goto rnr_nak;
qp->r_rcv_len = 0;
if (opcode == OP(SEND_ONLY))
- goto send_last;
- /* FALLTHROUGH */
+ goto no_immediate_data;
+ /* FALLTHROUGH for SEND_ONLY_WITH_IMMEDIATE */
case OP(SEND_LAST_WITH_IMMEDIATE):
send_last_imm:
wc.ex.imm_data = ohdr->u.imm_data;
hdrsize += 4;
wc.wc_flags = IB_WC_WITH_IMM;
- /* FALLTHROUGH */
+ goto send_last;
case OP(SEND_LAST):
case OP(RDMA_WRITE_LAST):
+no_immediate_data:
+ wc.wc_flags = 0;
+ wc.ex.imm_data = 0;
send_last:
/* Get the number of bytes the message was padded by. */
pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
@@ -2051,6 +2047,12 @@ send_last:
wc.src_qp = qp->remote_qpn;
wc.slid = qp->remote_ah_attr.dlid;
wc.sl = qp->remote_ah_attr.sl;
+ /* zero fields that are N/A */
+ wc.vendor_err = 0;
+ wc.pkey_index = 0;
+ wc.dlid_path_bits = 0;
+ wc.port_num = 0;
+ wc.csum_ok = 0;
/* Signal completion event if the solicited bit is set. */
qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
(ohdr->bth[0] &
@@ -2089,7 +2091,7 @@ send_last:
if (opcode == OP(RDMA_WRITE_FIRST))
goto send_middle;
else if (opcode == OP(RDMA_WRITE_ONLY))
- goto send_last;
+ goto no_immediate_data;
ret = qib_get_rwqe(qp, 1);
if (ret < 0)
goto nack_op_err;
diff --git a/drivers/infiniband/hw/qib/qib_ruc.c b/drivers/infiniband/hw/qib/qib_ruc.c
index eb78d9367f0..b4b37e47321 100644
--- a/drivers/infiniband/hw/qib/qib_ruc.c
+++ b/drivers/infiniband/hw/qib/qib_ruc.c
@@ -260,12 +260,15 @@ static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
/*
*
- * This should be called with the QP s_lock held.
+ * This should be called with the QP r_lock held.
+ *
+ * The s_lock will be acquired around the qib_migrate_qp() call.
*/
int qib_ruc_check_hdr(struct qib_ibport *ibp, struct qib_ib_header *hdr,
int has_grh, struct qib_qp *qp, u32 bth0)
{
__be64 guid;
+ unsigned long flags;
if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
if (!has_grh) {
@@ -295,7 +298,9 @@ int qib_ruc_check_hdr(struct qib_ibport *ibp, struct qib_ib_header *hdr,
if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid ||
ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num)
goto err;
+ spin_lock_irqsave(&qp->s_lock, flags);
qib_migrate_qp(qp);
+ spin_unlock_irqrestore(&qp->s_lock, flags);
} else {
if (!has_grh) {
if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
diff --git a/drivers/infiniband/hw/qib/qib_srq.c b/drivers/infiniband/hw/qib/qib_srq.c
index c3ec8efc2ed..d6235931a1b 100644
--- a/drivers/infiniband/hw/qib/qib_srq.c
+++ b/drivers/infiniband/hw/qib/qib_srq.c
@@ -107,6 +107,11 @@ struct ib_srq *qib_create_srq(struct ib_pd *ibpd,
u32 sz;
struct ib_srq *ret;
+ if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
+ ret = ERR_PTR(-ENOSYS);
+ goto done;
+ }
+
if (srq_init_attr->attr.max_sge == 0 ||
srq_init_attr->attr.max_sge > ib_qib_max_srq_sges ||
srq_init_attr->attr.max_wr == 0 ||
diff --git a/drivers/infiniband/hw/qib/qib_sysfs.c b/drivers/infiniband/hw/qib/qib_sysfs.c
index 14d129de432..78fbd56879d 100644
--- a/drivers/infiniband/hw/qib/qib_sysfs.c
+++ b/drivers/infiniband/hw/qib/qib_sysfs.c
@@ -515,8 +515,7 @@ static ssize_t show_nfreectxts(struct device *device,
struct qib_devdata *dd = dd_from_dev(dev);
/* Return the number of free user ports (contexts) available. */
- return scnprintf(buf, PAGE_SIZE, "%u\n", dd->cfgctxts -
- dd->first_user_ctxt - (u32)qib_stats.sps_ctxts);
+ return scnprintf(buf, PAGE_SIZE, "%u\n", dd->freectxts);
}
static ssize_t show_serial(struct device *device,
diff --git a/drivers/infiniband/hw/qib/qib_uc.c b/drivers/infiniband/hw/qib/qib_uc.c
index 32ccf3c824c..847e7afdfd9 100644
--- a/drivers/infiniband/hw/qib/qib_uc.c
+++ b/drivers/infiniband/hw/qib/qib_uc.c
@@ -51,7 +51,7 @@ int qib_make_uc_req(struct qib_qp *qp)
u32 hwords;
u32 bth0;
u32 len;
- u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
+ u32 pmtu = qp->pmtu;
int ret = 0;
spin_lock_irqsave(&qp->s_lock, flags);
@@ -243,13 +243,12 @@ void qib_uc_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
int has_grh, void *data, u32 tlen, struct qib_qp *qp)
{
struct qib_other_headers *ohdr;
- unsigned long flags;
u32 opcode;
u32 hdrsize;
u32 psn;
u32 pad;
struct ib_wc wc;
- u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
+ u32 pmtu = qp->pmtu;
struct ib_reth *reth;
int ret;
@@ -263,14 +262,11 @@ void qib_uc_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
}
opcode = be32_to_cpu(ohdr->bth[0]);
- spin_lock_irqsave(&qp->s_lock, flags);
if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode))
- goto sunlock;
- spin_unlock_irqrestore(&qp->s_lock, flags);
+ return;
psn = be32_to_cpu(ohdr->bth[2]);
opcode >>= 24;
- memset(&wc, 0, sizeof wc);
/* Compare the PSN verses the expected PSN. */
if (unlikely(qib_cmp24(psn, qp->r_psn) != 0)) {
@@ -370,7 +366,7 @@ send_first:
}
qp->r_rcv_len = 0;
if (opcode == OP(SEND_ONLY))
- goto send_last;
+ goto no_immediate_data;
else if (opcode == OP(SEND_ONLY_WITH_IMMEDIATE))
goto send_last_imm;
/* FALLTHROUGH */
@@ -389,8 +385,11 @@ send_last_imm:
wc.ex.imm_data = ohdr->u.imm_data;
hdrsize += 4;
wc.wc_flags = IB_WC_WITH_IMM;
- /* FALLTHROUGH */
+ goto send_last;
case OP(SEND_LAST):
+no_immediate_data:
+ wc.ex.imm_data = 0;
+ wc.wc_flags = 0;
send_last:
/* Get the number of bytes the message was padded by. */
pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
@@ -418,6 +417,12 @@ last_imm:
wc.src_qp = qp->remote_qpn;
wc.slid = qp->remote_ah_attr.dlid;
wc.sl = qp->remote_ah_attr.sl;
+ /* zero fields that are N/A */
+ wc.vendor_err = 0;
+ wc.pkey_index = 0;
+ wc.dlid_path_bits = 0;
+ wc.port_num = 0;
+ wc.csum_ok = 0;
/* Signal completion event if the solicited bit is set. */
qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
(ohdr->bth[0] &
@@ -546,6 +551,4 @@ op_err:
qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
return;
-sunlock:
- spin_unlock_irqrestore(&qp->s_lock, flags);
}
diff --git a/drivers/infiniband/hw/qib/qib_verbs.c b/drivers/infiniband/hw/qib/qib_verbs.c
index 9fab4048885..9627cb73712 100644
--- a/drivers/infiniband/hw/qib/qib_verbs.c
+++ b/drivers/infiniband/hw/qib/qib_verbs.c
@@ -38,11 +38,12 @@
#include <linux/utsname.h>
#include <linux/rculist.h>
#include <linux/mm.h>
+#include <linux/random.h>
#include "qib.h"
#include "qib_common.h"
-static unsigned int ib_qib_qp_table_size = 251;
+static unsigned int ib_qib_qp_table_size = 256;
module_param_named(qp_table_size, ib_qib_qp_table_size, uint, S_IRUGO);
MODULE_PARM_DESC(qp_table_size, "QP table size");
@@ -659,17 +660,25 @@ void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen)
if (atomic_dec_return(&mcast->refcount) <= 1)
wake_up(&mcast->wait);
} else {
- qp = qib_lookup_qpn(ibp, qp_num);
- if (!qp)
- goto drop;
+ if (rcd->lookaside_qp) {
+ if (rcd->lookaside_qpn != qp_num) {
+ if (atomic_dec_and_test(
+ &rcd->lookaside_qp->refcount))
+ wake_up(
+ &rcd->lookaside_qp->wait);
+ rcd->lookaside_qp = NULL;
+ }
+ }
+ if (!rcd->lookaside_qp) {
+ qp = qib_lookup_qpn(ibp, qp_num);
+ if (!qp)
+ goto drop;
+ rcd->lookaside_qp = qp;
+ rcd->lookaside_qpn = qp_num;
+ } else
+ qp = rcd->lookaside_qp;
ibp->n_unicast_rcv++;
qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
- /*
- * Notify qib_destroy_qp() if it is waiting
- * for us to finish.
- */
- if (atomic_dec_and_test(&qp->refcount))
- wake_up(&qp->wait);
}
return;
@@ -1974,6 +1983,8 @@ static void init_ibport(struct qib_pportdata *ppd)
ibp->z_excessive_buffer_overrun_errors =
cntrs.excessive_buffer_overrun_errors;
ibp->z_vl15_dropped = cntrs.vl15_dropped;
+ RCU_INIT_POINTER(ibp->qp0, NULL);
+ RCU_INIT_POINTER(ibp->qp1, NULL);
}
/**
@@ -1990,12 +2001,15 @@ int qib_register_ib_device(struct qib_devdata *dd)
int ret;
dev->qp_table_size = ib_qib_qp_table_size;
- dev->qp_table = kzalloc(dev->qp_table_size * sizeof *dev->qp_table,
+ get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
+ dev->qp_table = kmalloc(dev->qp_table_size * sizeof *dev->qp_table,
GFP_KERNEL);
if (!dev->qp_table) {
ret = -ENOMEM;
goto err_qpt;
}
+ for (i = 0; i < dev->qp_table_size; i++)
+ RCU_INIT_POINTER(dev->qp_table[i], NULL);
for (i = 0; i < dd->num_pports; i++)
init_ibport(ppd + i);
diff --git a/drivers/infiniband/hw/qib/qib_verbs.h b/drivers/infiniband/hw/qib/qib_verbs.h
index 95e5b47223b..0c19ef0c412 100644
--- a/drivers/infiniband/hw/qib/qib_verbs.h
+++ b/drivers/infiniband/hw/qib/qib_verbs.h
@@ -485,6 +485,7 @@ struct qib_qp {
u8 alt_timeout; /* Alternate path timeout for this QP */
u8 port_num;
enum ib_mtu path_mtu;
+ u32 pmtu; /* decoded from path_mtu */
u32 remote_qpn;
u32 qkey; /* QKEY for this QP (for UD or RD) */
u32 s_size; /* send work queue size */
@@ -495,6 +496,7 @@ struct qib_qp {
u32 s_last; /* last completed entry */
u32 s_ssn; /* SSN of tail entry */
u32 s_lsn; /* limit sequence number (credit) */
+ unsigned long timeout_jiffies; /* computed from timeout */
struct qib_swqe *s_wq; /* send work queue */
struct qib_swqe *s_wqe;
struct qib_rq r_rq; /* receive work queue */
@@ -723,7 +725,8 @@ struct qib_ibdev {
dma_addr_t pio_hdrs_phys;
/* list of QPs waiting for RNR timer */
spinlock_t pending_lock; /* protect wait lists, PMA counters, etc. */
- unsigned qp_table_size; /* size of the hash table */
+ u32 qp_table_size; /* size of the hash table */
+ u32 qp_rnd; /* random bytes for hash */
spinlock_t qpt_lock;
u32 n_piowait;