From: "Michael J. Ruhl" <michael.j.ruhl@intel.com>
Date: Fri, 4 Aug 2017 13:52:44 -0700
Subject: IB/hf1: User context locking is inconsistent
Patch-mainline: v4.14-rc1
Git-commit: d295dbeb2a0c93364444e76b3bb30f587a823e0e
References: bsc#1060463 FATE#323043
There is a mixture of mutex and spinlocks to protect receive context
(rcd/uctxt) information. This is not used consistently.
Use the mutex to protect device receive context information only.
Use the spinlock to protect sub context information only.
Protect access to items in the rcd array with a spinlock and
reference count.
Remove spinlock around dd->rcd array cleanup. Since interrupts are
disabled and cleaned up before this point, this lock is not useful.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Acked-by: Thomas Bogendoerfer <tbogendoerfer@suse.de>
---
drivers/infiniband/hw/hfi1/aspm.h | 35 +++---
drivers/infiniband/hw/hfi1/chip.c | 30 +++--
drivers/infiniband/hw/hfi1/debugfs.c | 32 +++--
drivers/infiniband/hw/hfi1/driver.c | 71 ++++++++----
drivers/infiniband/hw/hfi1/file_ops.c | 191 +++++++++++++++++++++------------
drivers/infiniband/hw/hfi1/hfi.h | 6 -
drivers/infiniband/hw/hfi1/init.c | 134 ++++++++++++++---------
drivers/infiniband/hw/hfi1/trace_rx.h | 12 +-
drivers/infiniband/hw/hfi1/vnic_main.c | 12 --
9 files changed, 329 insertions(+), 194 deletions(-)
--- a/drivers/infiniband/hw/hfi1/aspm.h
+++ b/drivers/infiniband/hw/hfi1/aspm.h
@@ -240,11 +240,14 @@ static inline void aspm_disable_all(stru
u16 i;
for (i = 0; i < dd->first_dyn_alloc_ctxt; i++) {
- rcd = dd->rcd[i];
- del_timer_sync(&rcd->aspm_timer);
- spin_lock_irqsave(&rcd->aspm_lock, flags);
- rcd->aspm_intr_enable = false;
- spin_unlock_irqrestore(&rcd->aspm_lock, flags);
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (rcd) {
+ del_timer_sync(&rcd->aspm_timer);
+ spin_lock_irqsave(&rcd->aspm_lock, flags);
+ rcd->aspm_intr_enable = false;
+ spin_unlock_irqrestore(&rcd->aspm_lock, flags);
+ hfi1_rcd_put(rcd);
+ }
}
aspm_disable(dd);
@@ -264,11 +267,14 @@ static inline void aspm_enable_all(struc
return;
for (i = 0; i < dd->first_dyn_alloc_ctxt; i++) {
- rcd = dd->rcd[i];
- spin_lock_irqsave(&rcd->aspm_lock, flags);
- rcd->aspm_intr_enable = true;
- rcd->aspm_enabled = true;
- spin_unlock_irqrestore(&rcd->aspm_lock, flags);
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (rcd) {
+ spin_lock_irqsave(&rcd->aspm_lock, flags);
+ rcd->aspm_intr_enable = true;
+ rcd->aspm_enabled = true;
+ spin_unlock_irqrestore(&rcd->aspm_lock, flags);
+ hfi1_rcd_put(rcd);
+ }
}
}
@@ -284,13 +290,18 @@ static inline void aspm_ctx_init(struct
static inline void aspm_init(struct hfi1_devdata *dd)
{
+ struct hfi1_ctxtdata *rcd;
u16 i;
spin_lock_init(&dd->aspm_lock);
dd->aspm_supported = aspm_hw_l1_supported(dd);
- for (i = 0; i < dd->first_dyn_alloc_ctxt; i++)
- aspm_ctx_init(dd->rcd[i]);
+ for (i = 0; i < dd->first_dyn_alloc_ctxt; i++) {
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (rcd)
+ aspm_ctx_init(rcd);
+ hfi1_rcd_put(rcd);
+ }
/* Start with ASPM disabled */
aspm_hw_set_l1_ent_latency(dd);
--- a/drivers/infiniband/hw/hfi1/chip.c
+++ b/drivers/infiniband/hw/hfi1/chip.c
@@ -6785,13 +6785,17 @@ static void wait_for_freeze_status(struc
static void rxe_freeze(struct hfi1_devdata *dd)
{
int i;
+ struct hfi1_ctxtdata *rcd;
/* disable port */
clear_rcvctrl(dd, RCV_CTRL_RCV_PORT_ENABLE_SMASK);
/* disable all receive contexts */
- for (i = 0; i < dd->num_rcv_contexts; i++)
- hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS, dd->rcd[i]);
+ for (i = 0; i < dd->num_rcv_contexts; i++) {
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS, rcd);
+ hfi1_rcd_put(rcd);
+ }
}
/*
@@ -6804,20 +6808,23 @@ static void rxe_kernel_unfreeze(struct h
{
u32 rcvmask;
u16 i;
+ struct hfi1_ctxtdata *rcd;
/* enable all kernel contexts */
for (i = 0; i < dd->num_rcv_contexts; i++) {
- struct hfi1_ctxtdata *rcd = dd->rcd[i];
+ rcd = hfi1_rcd_get_by_index(dd, i);
/* Ensure all non-user contexts(including vnic) are enabled */
- if (!rcd || !rcd->sc || (rcd->sc->type == SC_USER))
+ if (!rcd || !rcd->sc || (rcd->sc->type == SC_USER)) {
+ hfi1_rcd_put(rcd);
continue;
-
+ }
rcvmask = HFI1_RCVCTRL_CTXT_ENB;
/* HFI1_RCVCTRL_TAILUPD_[ENB|DIS] needs to be set explicitly */
rcvmask |= HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ?
HFI1_RCVCTRL_TAILUPD_ENB : HFI1_RCVCTRL_TAILUPD_DIS;
hfi1_rcvctrl(dd, rcvmask, rcd);
+ hfi1_rcd_put(rcd);
}
/* enable port */
@@ -8104,7 +8111,7 @@ static void is_rcv_avail_int(struct hfi1
char *err_detail;
if (likely(source < dd->num_rcv_contexts)) {
- rcd = dd->rcd[source];
+ rcd = hfi1_rcd_get_by_index(dd, source);
if (rcd) {
/* Check for non-user contexts, including vnic */
if ((source < dd->first_dyn_alloc_ctxt) ||
@@ -8112,6 +8119,8 @@ static void is_rcv_avail_int(struct hfi1
rcd->do_interrupt(rcd, 0);
else
handle_user_interrupt(rcd);
+
+ hfi1_rcd_put(rcd);
return; /* OK */
}
/* received an interrupt, but no rcd */
@@ -8133,12 +8142,14 @@ static void is_rcv_urgent_int(struct hfi
char *err_detail;
if (likely(source < dd->num_rcv_contexts)) {
- rcd = dd->rcd[source];
+ rcd = hfi1_rcd_get_by_index(dd, source);
if (rcd) {
/* only pay attention to user urgent interrupts */
if ((source >= dd->first_dyn_alloc_ctxt) &&
(!rcd->sc || (rcd->sc->type == SC_USER)))
handle_user_interrupt(rcd);
+
+ hfi1_rcd_put(rcd);
return; /* OK */
}
/* received an interrupt, but no rcd */
@@ -8343,7 +8354,7 @@ static irqreturn_t receive_context_inter
int disposition;
int present;
- trace_hfi1_receive_interrupt(dd, rcd->ctxt);
+ trace_hfi1_receive_interrupt(dd, rcd);
this_cpu_inc(*dd->int_counter);
aspm_ctx_disable(rcd);
@@ -13030,7 +13041,7 @@ static int request_msix_irqs(struct hfi1
me->type = IRQ_SDMA;
} else if (first_rx <= i && i < last_rx) {
idx = i - first_rx;
- rcd = dd->rcd[idx];
+ rcd = hfi1_rcd_get_by_index(dd, idx);
if (rcd) {
/*
* Set the interrupt register and mask for this
@@ -13049,6 +13060,7 @@ static int request_msix_irqs(struct hfi1
remap_intr(dd, IS_RCVAVAIL_START + idx, i);
me->type = IRQ_RCVCTXT;
rcd->msix_intr = i;
+ hfi1_rcd_put(rcd);
}
} else {
/* not in our expected range - complain, then
--- a/drivers/infiniband/hw/hfi1/debugfs.c
+++ b/drivers/infiniband/hw/hfi1/debugfs.c
@@ -173,12 +173,15 @@ static int _opcode_stats_seq_show(struct
u64 n_packets = 0, n_bytes = 0;
struct hfi1_ibdev *ibd = (struct hfi1_ibdev *)s->private;
struct hfi1_devdata *dd = dd_from_dev(ibd);
+ struct hfi1_ctxtdata *rcd;
for (j = 0; j < dd->first_dyn_alloc_ctxt; j++) {
- if (!dd->rcd[j])
- continue;
- n_packets += dd->rcd[j]->opstats->stats[i].n_packets;
- n_bytes += dd->rcd[j]->opstats->stats[i].n_bytes;
+ rcd = hfi1_rcd_get_by_index(dd, j);
+ if (rcd) {
+ n_packets += rcd->opstats->stats[i].n_packets;
+ n_bytes += rcd->opstats->stats[i].n_bytes;
+ }
+ hfi1_rcd_put(rcd);
}
if (!n_packets && !n_bytes)
return SEQ_SKIP;
@@ -231,6 +234,7 @@ static int _ctx_stats_seq_show(struct se
u64 n_packets = 0;
struct hfi1_ibdev *ibd = (struct hfi1_ibdev *)s->private;
struct hfi1_devdata *dd = dd_from_dev(ibd);
+ struct hfi1_ctxtdata *rcd;
if (v == SEQ_START_TOKEN) {
seq_puts(s, "Ctx:npkts\n");
@@ -240,11 +244,14 @@ static int _ctx_stats_seq_show(struct se
spos = v;
i = *spos;
- if (!dd->rcd[i])
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (!rcd)
return SEQ_SKIP;
- for (j = 0; j < ARRAY_SIZE(dd->rcd[i]->opstats->stats); j++)
- n_packets += dd->rcd[i]->opstats->stats[j].n_packets;
+ for (j = 0; j < ARRAY_SIZE(rcd->opstats->stats); j++)
+ n_packets += rcd->opstats->stats[j].n_packets;
+
+ hfi1_rcd_put(rcd);
if (!n_packets)
return SEQ_SKIP;
@@ -1098,12 +1105,15 @@ static int _fault_stats_seq_show(struct
u64 n_packets = 0, n_bytes = 0;
struct hfi1_ibdev *ibd = (struct hfi1_ibdev *)s->private;
struct hfi1_devdata *dd = dd_from_dev(ibd);
+ struct hfi1_ctxtdata *rcd;
for (j = 0; j < dd->first_dyn_alloc_ctxt; j++) {
- if (!dd->rcd[j])
- continue;
- n_packets += dd->rcd[j]->opstats->stats[i].n_packets;
- n_bytes += dd->rcd[j]->opstats->stats[i].n_bytes;
+ rcd = hfi1_rcd_get_by_index(dd, j);
+ if (rcd) {
+ n_packets += rcd->opstats->stats[i].n_packets;
+ n_bytes += rcd->opstats->stats[i].n_bytes;
+ }
+ hfi1_rcd_put(rcd);
}
if (!n_packets && !n_bytes)
return SEQ_SKIP;
--- a/drivers/infiniband/hw/hfi1/driver.c
+++ b/drivers/infiniband/hw/hfi1/driver.c
@@ -839,6 +839,7 @@ bail:
static inline void set_nodma_rtail(struct hfi1_devdata *dd, u16 ctxt)
{
+ struct hfi1_ctxtdata *rcd;
u16 i;
/*
@@ -847,18 +848,27 @@ static inline void set_nodma_rtail(struc
* interrupt handler for all statically allocated kernel contexts.
*/
if (ctxt >= dd->first_dyn_alloc_ctxt) {
- dd->rcd[ctxt]->do_interrupt =
- &handle_receive_interrupt_nodma_rtail;
+ rcd = hfi1_rcd_get_by_index(dd, ctxt);
+ if (rcd) {
+ rcd->do_interrupt =
+ &handle_receive_interrupt_nodma_rtail;
+ hfi1_rcd_put(rcd);
+ }
return;
}
- for (i = HFI1_CTRL_CTXT + 1; i < dd->first_dyn_alloc_ctxt; i++)
- dd->rcd[i]->do_interrupt =
- &handle_receive_interrupt_nodma_rtail;
+ for (i = HFI1_CTRL_CTXT + 1; i < dd->first_dyn_alloc_ctxt; i++) {
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (rcd)
+ rcd->do_interrupt =
+ &handle_receive_interrupt_nodma_rtail;
+ hfi1_rcd_put(rcd);
+ }
}
static inline void set_dma_rtail(struct hfi1_devdata *dd, u16 ctxt)
{
+ struct hfi1_ctxtdata *rcd;
u16 i;
/*
@@ -867,27 +877,39 @@ static inline void set_dma_rtail(struct
* interrupt handler for all statically allocated kernel contexts.
*/
if (ctxt >= dd->first_dyn_alloc_ctxt) {
- dd->rcd[ctxt]->do_interrupt =
- &handle_receive_interrupt_dma_rtail;
+ rcd = hfi1_rcd_get_by_index(dd, ctxt);
+ if (rcd) {
+ rcd->do_interrupt =
+ &handle_receive_interrupt_dma_rtail;
+ hfi1_rcd_put(rcd);
+ }
return;
}
- for (i = HFI1_CTRL_CTXT + 1; i < dd->first_dyn_alloc_ctxt; i++)
- dd->rcd[i]->do_interrupt =
- &handle_receive_interrupt_dma_rtail;
+ for (i = HFI1_CTRL_CTXT + 1; i < dd->first_dyn_alloc_ctxt; i++) {
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (rcd)
+ rcd->do_interrupt =
+ &handle_receive_interrupt_dma_rtail;
+ hfi1_rcd_put(rcd);
+ }
}
void set_all_slowpath(struct hfi1_devdata *dd)
{
+ struct hfi1_ctxtdata *rcd;
u16 i;
/* HFI1_CTRL_CTXT must always use the slow path interrupt handler */
for (i = HFI1_CTRL_CTXT + 1; i < dd->num_rcv_contexts; i++) {
- struct hfi1_ctxtdata *rcd = dd->rcd[i];
-
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (!rcd)
+ continue;
if ((i < dd->first_dyn_alloc_ctxt) ||
- (rcd && rcd->sc && (rcd->sc->type == SC_KERNEL)))
+ (rcd->sc && (rcd->sc->type == SC_KERNEL))) {
rcd->do_interrupt = &handle_receive_interrupt;
+ }
+ hfi1_rcd_put(rcd);
}
}
@@ -1068,6 +1090,7 @@ void receive_interrupt_work(struct work_
struct hfi1_pportdata *ppd = container_of(work, struct hfi1_pportdata,
linkstate_active_work);
struct hfi1_devdata *dd = ppd->dd;
+ struct hfi1_ctxtdata *rcd;
u16 i;
/* Received non-SC15 packet implies neighbor_normal */
@@ -1078,8 +1101,12 @@ void receive_interrupt_work(struct work_
* Interrupt all statically allocated kernel contexts that could
* have had an interrupt during auto activation.
*/
- for (i = HFI1_CTRL_CTXT; i < dd->first_dyn_alloc_ctxt; i++)
- force_recv_intr(dd->rcd[i]);
+ for (i = HFI1_CTRL_CTXT; i < dd->first_dyn_alloc_ctxt; i++) {
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (rcd)
+ force_recv_intr(rcd);
+ hfi1_rcd_put(rcd);
+ }
}
/*
@@ -1270,10 +1297,8 @@ void hfi1_start_led_override(struct hfi1
int hfi1_reset_device(int unit)
{
int ret;
- u16 i;
struct hfi1_devdata *dd = hfi1_lookup(unit);
struct hfi1_pportdata *ppd;
- unsigned long flags;
int pidx;
if (!dd) {
@@ -1291,17 +1316,15 @@ int hfi1_reset_device(int unit)
goto bail;
}
- spin_lock_irqsave(&dd->uctxt_lock, flags);
+ /* If there are any user/vnic contexts, we cannot reset */
+ mutex_lock(&hfi1_mutex);
if (dd->rcd)
- for (i = dd->first_dyn_alloc_ctxt;
- i < dd->num_rcv_contexts; i++) {
- if (!dd->rcd[i])
- continue;
- spin_unlock_irqrestore(&dd->uctxt_lock, flags);
+ if (hfi1_stats.sps_ctxts) {
+ mutex_unlock(&hfi1_mutex);
ret = -EBUSY;
goto bail;
}
- spin_unlock_irqrestore(&dd->uctxt_lock, flags);
+ mutex_unlock(&hfi1_mutex);
for (pidx = 0; pidx < dd->num_pports; ++pidx) {
ppd = dd->pport + pidx;
--- a/drivers/infiniband/hw/hfi1/file_ops.c
+++ b/drivers/infiniband/hw/hfi1/file_ops.c
@@ -757,7 +757,7 @@ static int hfi1_file_close(struct inode
if (!uctxt)
goto done;
- hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
+ hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
flush_wc();
/* drain user sdma queue */
@@ -770,6 +770,13 @@ static int hfi1_file_close(struct inode
hfi1_user_exp_rcv_free(fdata);
/*
+ * fdata->uctxt is used in the above cleanup. It is not ready to be
+ * removed until here.
+ */
+ fdata->uctxt = NULL;
+ hfi1_rcd_put(uctxt);
+
+ /*
* Clear any left over, unhandled events so the next process that
* gets this context doesn't get confused.
*/
@@ -777,16 +784,14 @@ static int hfi1_file_close(struct inode
HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
*ev = 0;
- mutex_lock(&hfi1_mutex);
+ spin_lock_irqsave(&dd->uctxt_lock, flags);
__clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
- fdata->uctxt = NULL;
- hfi1_rcd_put(uctxt); /* fdata reference */
if (!bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
- mutex_unlock(&hfi1_mutex);
+ spin_unlock_irqrestore(&dd->uctxt_lock, flags);
goto done;
}
+ spin_unlock_irqrestore(&dd->uctxt_lock, flags);
- spin_lock_irqsave(&dd->uctxt_lock, flags);
/*
* Disable receive context and interrupt available, reset all
* RcvCtxtCtrl bits to default values.
@@ -808,13 +813,11 @@ static int hfi1_file_close(struct inode
set_pio_integrity(uctxt->sc);
sc_disable(uctxt->sc);
}
- spin_unlock_irqrestore(&dd->uctxt_lock, flags);
hfi1_free_ctxt_rcv_groups(uctxt);
hfi1_clear_ctxt_pkey(dd, uctxt);
uctxt->event_flags = 0;
- mutex_unlock(&hfi1_mutex);
deallocate_ctxt(uctxt);
done:
@@ -844,9 +847,22 @@ static u64 kvirt_to_phys(void *addr)
return paddr;
}
+/**
+ * complete_subctxt
+ * @fd: valid filedata pointer
+ *
+ * Sub-context info can only be set up after the base context
+ * has been completed. This is indicated by the clearing of the
+ * HFI1_CTXT_BASE_UINIT bit.
+ *
+ * Wait for the bit to be cleared, and then complete the subcontext
+ * initialization.
+ *
+ */
static int complete_subctxt(struct hfi1_filedata *fd)
{
int ret;
+ unsigned long flags;
/*
* sub-context info can only be set up after the base context
@@ -859,7 +875,7 @@ static int complete_subctxt(struct hfi1_
if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
ret = -ENOMEM;
- /* The only thing a sub context needs is the user_xxx stuff */
+ /* Finish the sub-context init */
if (!ret) {
fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
ret = init_user_ctxt(fd, fd->uctxt);
@@ -868,9 +884,9 @@ static int complete_subctxt(struct hfi1_
if (ret) {
hfi1_rcd_put(fd->uctxt);
fd->uctxt = NULL;
- mutex_lock(&hfi1_mutex);
+ spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
__clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
- mutex_unlock(&hfi1_mutex);
+ spin_unlock_irqrestore(&fd->dd->uctxt_lock, flags);
}
return ret;
@@ -911,14 +927,15 @@ static int assign_ctxt(struct hfi1_filed
mutex_unlock(&hfi1_mutex);
- /* Depending on the context type, do the appropriate init */
+ /* Depending on the context type, finish the appropriate init */
switch (ret) {
case 0:
ret = setup_base_ctxt(fd, uctxt);
if (uctxt->subctxt_cnt) {
/*
- * Base context is done, notify anybody using a
- * sub-context that is waiting for this completion
+ * Base context is done (successfully or not), notify
+ * anybody using a sub-context that is waiting for
+ * this completion.
*/
clear_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
wake_up(&uctxt->wait);
@@ -934,58 +951,97 @@ static int assign_ctxt(struct hfi1_filed
return ret;
}
-/*
+/**
+ * match_ctxt
+ * @fd: valid filedata pointer
+ * @uinfo: user info to compare base context with
+ * @uctxt: context to compare uinfo to.
+ *
+ * Compare the given context with the given information to see if it
+ * can be used for a sub context.
+ */
+static int match_ctxt(struct hfi1_filedata *fd,
+ const struct hfi1_user_info *uinfo,
+ struct hfi1_ctxtdata *uctxt)
+{
+ struct hfi1_devdata *dd = fd->dd;
+ unsigned long flags;
+ u16 subctxt;
+
+ /* Skip dynamically allocated kernel contexts */
+ if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
+ return 0;
+
+ /* Skip ctxt if it doesn't match the requested one */
+ if (memcmp(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid)) ||
+ uctxt->jkey != generate_jkey(current_uid()) ||
+ uctxt->subctxt_id != uinfo->subctxt_id ||
+ uctxt->subctxt_cnt != uinfo->subctxt_cnt)
+ return 0;
+
+ /* Verify the sharing process matches the base */
+ if (uctxt->userversion != uinfo->userversion)
+ return -EINVAL;
+
+ /* Find an unused sub context */
+ spin_lock_irqsave(&dd->uctxt_lock, flags);
+ if (bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
+ /* context is being closed, do not use */
+ spin_unlock_irqrestore(&dd->uctxt_lock, flags);
+ return 0;
+ }
+
+ subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
+ HFI1_MAX_SHARED_CTXTS);
+ if (subctxt >= uctxt->subctxt_cnt) {
+ spin_unlock_irqrestore(&dd->uctxt_lock, flags);
+ return -EBUSY;
+ }
+
+ fd->subctxt = subctxt;
+ __set_bit(fd->subctxt, uctxt->in_use_ctxts);
+ spin_unlock_irqrestore(&dd->uctxt_lock, flags);
+
+ fd->uctxt = uctxt;
+ hfi1_rcd_get(uctxt);
+
+ return 1;
+}
+
+/**
+ * find_sub_ctxt
+ * @fd: valid filedata pointer
+ * @uinfo: matching info to use to find a possible context to share.
+ *
* The hfi1_mutex must be held when this function is called. It is
* necessary to ensure serialized creation of shared contexts.
+ *
+ * Return:
+ * 0 No sub-context found
+ * 1 Subcontext found and allocated
+ * errno EINVAL (incorrect parameters)
+ * EBUSY (all sub contexts in use)
*/
static int find_sub_ctxt(struct hfi1_filedata *fd,
const struct hfi1_user_info *uinfo)
{
- u16 i;
+ struct hfi1_ctxtdata *uctxt;
struct hfi1_devdata *dd = fd->dd;
- u16 subctxt;
+ u16 i;
+ int ret;
if (!uinfo->subctxt_cnt)
return 0;
for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
- struct hfi1_ctxtdata *uctxt = dd->rcd[i];
-
- /* Skip ctxts which are not yet open */
- if (!uctxt ||
- bitmap_empty(uctxt->in_use_ctxts,
- HFI1_MAX_SHARED_CTXTS))
- continue;
-
- /* Skip dynamically allocted kernel contexts */
- if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
- continue;
-
- /* Skip ctxt if it doesn't match the requested one */
- if (memcmp(uctxt->uuid, uinfo->uuid,
- sizeof(uctxt->uuid)) ||
- uctxt->jkey != generate_jkey(current_uid()) ||
- uctxt->subctxt_id != uinfo->subctxt_id ||
- uctxt->subctxt_cnt != uinfo->subctxt_cnt)
- continue;
-
- /* Verify the sharing process matches the master */
- if (uctxt->userversion != uinfo->userversion)
- return -EINVAL;
-
- /* Find an unused context */
- subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
- HFI1_MAX_SHARED_CTXTS);
- if (subctxt >= uctxt->subctxt_cnt)
- return -EBUSY;
-
- fd->uctxt = uctxt;
- fd->subctxt = subctxt;
-
- hfi1_rcd_get(uctxt);
- __set_bit(fd->subctxt, uctxt->in_use_ctxts);
-
- return 1;
+ uctxt = hfi1_rcd_get_by_index(dd, i);
+ if (uctxt) {
+ ret = match_ctxt(fd, uinfo, uctxt);
+ hfi1_rcd_put(uctxt);
+ /* value of != 0 will return */
+ if (ret)
+ return ret;
+ }
}
return 0;
@@ -993,7 +1049,7 @@ static int find_sub_ctxt(struct hfi1_fil
static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
struct hfi1_user_info *uinfo,
- struct hfi1_ctxtdata **cd)
+ struct hfi1_ctxtdata **rcd)
{
struct hfi1_ctxtdata *uctxt;
int ret, numa;
@@ -1066,12 +1122,12 @@ static int allocate_ctxt(struct hfi1_fil
if (dd->freectxts-- == dd->num_user_contexts)
aspm_disable_all(dd);
- *cd = uctxt;
+ *rcd = uctxt;
return 0;
ctxdata_free:
- hfi1_free_ctxt(dd, uctxt);
+ hfi1_free_ctxt(uctxt);
return ret;
}
@@ -1083,7 +1139,7 @@ static void deallocate_ctxt(struct hfi1_
aspm_enable_all(uctxt->dd);
mutex_unlock(&hfi1_mutex);
- hfi1_free_ctxt(uctxt->dd, uctxt);
+ hfi1_free_ctxt(uctxt);
}
static void init_subctxts(struct hfi1_ctxtdata *uctxt,
@@ -1279,8 +1335,10 @@ static int setup_base_ctxt(struct hfi1_f
return 0;
setup_failed:
+ /* Set the failed bit so sub-context init can do the right thing */
set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
deallocate_ctxt(uctxt);
+
return ret;
}
@@ -1417,18 +1475,13 @@ int hfi1_set_uevent_bits(struct hfi1_ppo
struct hfi1_ctxtdata *uctxt;
struct hfi1_devdata *dd = ppd->dd;
u16 ctxt;
- int ret = 0;
- unsigned long flags;
- if (!dd->events) {
- ret = -EINVAL;
- goto done;
- }
+ if (!dd->events)
+ return -EINVAL;
- spin_lock_irqsave(&dd->uctxt_lock, flags);
for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
ctxt++) {
- uctxt = dd->rcd[ctxt];
+ uctxt = hfi1_rcd_get_by_index(dd, ctxt);
if (uctxt) {
unsigned long *evs = dd->events +
(uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
@@ -1441,11 +1494,11 @@ int hfi1_set_uevent_bits(struct hfi1_ppo
set_bit(evtbit, evs);
for (i = 1; i < uctxt->subctxt_cnt; i++)
set_bit(evtbit, evs + i);
+ hfi1_rcd_put(uctxt);
}
}
- spin_unlock_irqrestore(&dd->uctxt_lock, flags);
-done:
- return ret;
+
+ return 0;
}
/**
--- a/drivers/infiniband/hw/hfi1/hfi.h
+++ b/drivers/infiniband/hw/hfi1/hfi.h
@@ -938,8 +938,7 @@ struct hfi1_devdata {
u64 __iomem *egrtidbase;
spinlock_t sendctrl_lock; /* protect changes to SendCtrl */
spinlock_t rcvctrl_lock; /* protect changes to RcvCtrl */
- /* around rcd and (user ctxts) ctxt_cnt use (intr vs free) */
- spinlock_t uctxt_lock; /* rcd and user context changes */
+ spinlock_t uctxt_lock; /* protect rcd changes */
struct mutex dc8051_lock; /* exclusive access to 8051 */
struct workqueue_struct *update_cntr_wq;
struct work_struct update_cntr_work;
@@ -1265,12 +1264,13 @@ int hfi1_setup_eagerbufs(struct hfi1_ctx
int hfi1_create_kctxts(struct hfi1_devdata *dd);
int hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, int numa,
struct hfi1_ctxtdata **rcd);
-void hfi1_free_ctxt(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd);
+void hfi1_free_ctxt(struct hfi1_ctxtdata *rcd);
void hfi1_init_pportdata(struct pci_dev *pdev, struct hfi1_pportdata *ppd,
struct hfi1_devdata *dd, u8 hw_pidx, u8 port);
void hfi1_free_ctxtdata(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd);
int hfi1_rcd_put(struct hfi1_ctxtdata *rcd);
void hfi1_rcd_get(struct hfi1_ctxtdata *rcd);
+struct hfi1_ctxtdata *hfi1_rcd_get_by_index(struct hfi1_devdata *dd, u16 ctxt);
int handle_receive_interrupt(struct hfi1_ctxtdata *rcd, int thread);
int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata *rcd, int thread);
int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata *rcd, int thread);
--- a/drivers/infiniband/hw/hfi1/init.c
+++ b/drivers/infiniband/hw/hfi1/init.c
@@ -188,7 +188,7 @@ int hfi1_create_kctxts(struct hfi1_devda
return 0;
bail:
for (i = 0; dd->rcd && i < dd->first_dyn_alloc_ctxt; ++i)
- hfi1_rcd_put(dd->rcd[i]);
+ hfi1_free_ctxt(dd->rcd[i]);
/* All the contexts should be freed, free the array */
kfree(dd->rcd);
@@ -197,7 +197,7 @@ bail:
}
/*
- * Helper routines for the receive context reference count (rcd and uctxt)
+ * Helper routines for the receive context reference count (rcd and uctxt).
*/
static void hfi1_rcd_init(struct hfi1_ctxtdata *rcd)
{
@@ -211,10 +211,16 @@ static void hfi1_rcd_init(struct hfi1_ct
*/
static void hfi1_rcd_free(struct kref *kref)
{
+ unsigned long flags;
struct hfi1_ctxtdata *rcd =
container_of(kref, struct hfi1_ctxtdata, kref);
hfi1_free_ctxtdata(rcd->dd, rcd);
+
+ spin_lock_irqsave(&rcd->dd->uctxt_lock, flags);
+ rcd->dd->rcd[rcd->ctxt] = NULL;
+ spin_unlock_irqrestore(&rcd->dd->uctxt_lock, flags);
+
kfree(rcd);
}
@@ -253,7 +259,7 @@ void hfi1_rcd_get(struct hfi1_ctxtdata *
* If the array is full, we are EBUSY.
*
*/
-static u16 allocate_rcd_index(struct hfi1_devdata *dd,
+static int allocate_rcd_index(struct hfi1_devdata *dd,
struct hfi1_ctxtdata *rcd, u16 *index)
{
unsigned long flags;
@@ -279,8 +285,36 @@ static u16 allocate_rcd_index(struct hfi
return 0;
}
+/**
+ * hfi1_rcd_get_by_index
+ * @dd: pointer to a valid devdata structure
+ * @ctxt: the index of an possilbe rcd
+ *
+ * We need to protect access to the rcd array. If access is needed to
+ * one or more index, get the protecting spinlock and then increment the
+ * kref.
+ *
+ * The caller is responsible for making the _put().
+ *
+ */
+struct hfi1_ctxtdata *hfi1_rcd_get_by_index(struct hfi1_devdata *dd, u16 ctxt)
+{
+ unsigned long flags;
+ struct hfi1_ctxtdata *rcd = NULL;
+
+ spin_lock_irqsave(&dd->uctxt_lock, flags);
+ if (dd->rcd[ctxt]) {
+ rcd = dd->rcd[ctxt];
+ hfi1_rcd_get(rcd);
+ }
+ spin_unlock_irqrestore(&dd->uctxt_lock, flags);
+
+ return rcd;
+}
+
/*
- * Common code for user and kernel context setup.
+ * Common code for user and kernel context create and setup.
+ * NOTE: the initial kref is done here (hf1_rcd_init()).
*/
int hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, int numa,
struct hfi1_ctxtdata **context)
@@ -300,8 +334,6 @@ int hfi1_create_ctxtdata(struct hfi1_ppo
u16 ctxt;
int ret;
- hfi1_cdbg(PROC, "setting up context %u\n", ctxt);
-
ret = allocate_rcd_index(dd, rcd, &ctxt);
if (ret) {
*context = NULL;
@@ -321,6 +353,8 @@ int hfi1_create_ctxtdata(struct hfi1_ppo
mutex_init(&rcd->exp_lock);
+ hfi1_cdbg(PROC, "setting up context %u\n", rcd->ctxt);
+
/*
* Calculate the context's RcvArray entry starting point.
* We do this here because we have to take into account all
@@ -425,28 +459,23 @@ int hfi1_create_ctxtdata(struct hfi1_ppo
bail:
*context = NULL;
- hfi1_free_ctxt(dd, rcd);
+ hfi1_free_ctxt(rcd);
return -ENOMEM;
}
/**
* hfi1_free_ctxt
- * @dd: Pointer to a valid device
* @rcd: pointer to an initialized rcd data structure
*
- * This is the "free" to match the _create_ctxtdata (alloc) function.
- * This is the final "put" for the kref.
+ * This wrapper is the free function that matches hfi1_create_ctxtdata().
+ * When a context is done being used (kernel or user), this function is called
+ * for the "final" put to match the kref init from hf1i_create_ctxtdata().
+ * Other users of the context do a get/put sequence to make sure that the
+ * structure isn't removed while in use.
*/
-void hfi1_free_ctxt(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
+void hfi1_free_ctxt(struct hfi1_ctxtdata *rcd)
{
- unsigned long flags;
-
- if (rcd) {
- spin_lock_irqsave(&dd->uctxt_lock, flags);
- dd->rcd[rcd->ctxt] = NULL;
- spin_unlock_irqrestore(&dd->uctxt_lock, flags);
- hfi1_rcd_put(rcd);
- }
+ hfi1_rcd_put(rcd);
}
/*
@@ -669,16 +698,19 @@ static int loadtime_init(struct hfi1_dev
static int init_after_reset(struct hfi1_devdata *dd)
{
int i;
-
+ struct hfi1_ctxtdata *rcd;
/*
* Ensure chip does no sends or receives, tail updates, or
* pioavail updates while we re-initialize. This is mostly
* for the driver data structures, not chip registers.
*/
- for (i = 0; i < dd->num_rcv_contexts; i++)
+ for (i = 0; i < dd->num_rcv_contexts; i++) {
+ rcd = hfi1_rcd_get_by_index(dd, i);
hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
HFI1_RCVCTRL_INTRAVAIL_DIS |
- HFI1_RCVCTRL_TAILUPD_DIS, dd->rcd[i]);
+ HFI1_RCVCTRL_TAILUPD_DIS, rcd);
+ hfi1_rcd_put(rcd);
+ }
pio_send_control(dd, PSC_GLOBAL_DISABLE);
for (i = 0; i < dd->num_send_contexts; i++)
sc_disable(dd->send_contexts[i].sc);
@@ -688,6 +720,7 @@ static int init_after_reset(struct hfi1_
static void enable_chip(struct hfi1_devdata *dd)
{
+ struct hfi1_ctxtdata *rcd;
u32 rcvmask;
u16 i;
@@ -699,17 +732,21 @@ static void enable_chip(struct hfi1_devd
* Other ctxts done as user opens and initializes them.
*/
for (i = 0; i < dd->first_dyn_alloc_ctxt; ++i) {
+ rcd = hfi1_rcd_get_by_index(dd, i);
+ if (!rcd)
+ continue;
rcvmask = HFI1_RCVCTRL_CTXT_ENB | HFI1_RCVCTRL_INTRAVAIL_ENB;
- rcvmask |= HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, DMA_RTAIL) ?
+ rcvmask |= HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ?
HFI1_RCVCTRL_TAILUPD_ENB : HFI1_RCVCTRL_TAILUPD_DIS;
- if (!HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, MULTI_PKT_EGR))
+ if (!HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR))
rcvmask |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
- if (HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, NODROP_RHQ_FULL))
+ if (HFI1_CAP_KGET_MASK(rcd->flags, NODROP_RHQ_FULL))
rcvmask |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
- if (HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, NODROP_EGR_FULL))
+ if (HFI1_CAP_KGET_MASK(rcd->flags, NODROP_EGR_FULL))
rcvmask |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
- hfi1_rcvctrl(dd, rcvmask, dd->rcd[i]);
- sc_enable(dd->rcd[i]->sc);
+ hfi1_rcvctrl(dd, rcvmask, rcd);
+ sc_enable(rcd->sc);
+ hfi1_rcd_put(rcd);
}
}
@@ -854,7 +891,7 @@ int hfi1_init(struct hfi1_devdata *dd, i
* existing, and re-allocate.
* Need to re-create rest of ctxt 0 ctxtdata as well.
*/
- rcd = dd->rcd[i];
+ rcd = hfi1_rcd_get_by_index(dd, i);
if (!rcd)
continue;
@@ -868,6 +905,7 @@ int hfi1_init(struct hfi1_devdata *dd, i
"failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
ret = lastfail;
}
+ hfi1_rcd_put(rcd);
}
/* Allocate enough memory for user event notification. */
@@ -987,6 +1025,7 @@ static void stop_timers(struct hfi1_devd
static void shutdown_device(struct hfi1_devdata *dd)
{
struct hfi1_pportdata *ppd;
+ struct hfi1_ctxtdata *rcd;
unsigned pidx;
int i;
@@ -1005,12 +1044,15 @@ static void shutdown_device(struct hfi1_
for (pidx = 0; pidx < dd->num_pports; ++pidx) {
ppd = dd->pport + pidx;
- for (i = 0; i < dd->num_rcv_contexts; i++)
+ for (i = 0; i < dd->num_rcv_contexts; i++) {
+ rcd = hfi1_rcd_get_by_index(dd, i);
hfi1_rcvctrl(dd, HFI1_RCVCTRL_TAILUPD_DIS |
HFI1_RCVCTRL_CTXT_DIS |
HFI1_RCVCTRL_INTRAVAIL_DIS |
HFI1_RCVCTRL_PKEY_DIS |
- HFI1_RCVCTRL_ONE_PKT_EGR_DIS, dd->rcd[i]);
+ HFI1_RCVCTRL_ONE_PKT_EGR_DIS, rcd);
+ hfi1_rcd_put(rcd);
+ }
/*
* Gracefully stop all sends allowing any in progress to
* trickle out first.
@@ -1450,8 +1492,6 @@ static void cleanup_device_data(struct h
{
int ctxt;
int pidx;
- struct hfi1_ctxtdata **tmp;
- unsigned long flags;
/* users can't do anything more with chip */
for (pidx = 0; pidx < dd->num_pports; ++pidx) {
@@ -1476,18 +1516,6 @@ static void cleanup_device_data(struct h
free_credit_return(dd);
- /*
- * Free any resources still in use (usually just kernel contexts)
- * at unload; we do for ctxtcnt, because that's what we allocate.
- * We acquire lock to be really paranoid that rcd isn't being
- * accessed from some interrupt-related code (that should not happen,
- * but best to be sure).
- */
- spin_lock_irqsave(&dd->uctxt_lock, flags);
- tmp = dd->rcd;
- dd->rcd = NULL;
- spin_unlock_irqrestore(&dd->uctxt_lock, flags);
-
if (dd->rcvhdrtail_dummy_kvaddr) {
dma_free_coherent(&dd->pcidev->dev, sizeof(u64),
(void *)dd->rcvhdrtail_dummy_kvaddr,
@@ -1495,16 +1523,22 @@ static void cleanup_device_data(struct h
dd->rcvhdrtail_dummy_kvaddr = NULL;
}
- for (ctxt = 0; tmp && ctxt < dd->num_rcv_contexts; ctxt++) {
- struct hfi1_ctxtdata *rcd = tmp[ctxt];
+ /*
+ * Free any resources still in use (usually just kernel contexts)
+ * at unload; we do for ctxtcnt, because that's what we allocate.
+ */
+ for (ctxt = 0; dd->rcd && ctxt < dd->num_rcv_contexts; ctxt++) {
+ struct hfi1_ctxtdata *rcd = dd->rcd[ctxt];
- tmp[ctxt] = NULL; /* debugging paranoia */
if (rcd) {
hfi1_clear_tids(rcd);
- hfi1_rcd_put(rcd);
+ hfi1_free_ctxt(rcd);
}
}
- kfree(tmp);
+
+ kfree(dd->rcd);
+ dd->rcd = NULL;
+
free_pio_map(dd);
/* must follow rcv context free - need to remove rcv's hooks */
for (ctxt = 0; ctxt < dd->num_send_contexts; ctxt++)
--- a/drivers/infiniband/hw/hfi1/trace_rx.h
+++ b/drivers/infiniband/hw/hfi1/trace_rx.h
@@ -114,24 +114,24 @@ TRACE_EVENT(hfi1_rcvhdr,
);
TRACE_EVENT(hfi1_receive_interrupt,
- TP_PROTO(struct hfi1_devdata *dd, u16 ctxt),
- TP_ARGS(dd, ctxt),
+ TP_PROTO(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd),
+ TP_ARGS(dd, rcd),
TP_STRUCT__entry(DD_DEV_ENTRY(dd)
__field(u32, ctxt)
__field(u8, slow_path)
__field(u8, dma_rtail)
),
TP_fast_assign(DD_DEV_ASSIGN(dd);
- __entry->ctxt = ctxt;
- if (dd->rcd[ctxt]->do_interrupt ==
+ __entry->ctxt = rcd->ctxt;
+ if (rcd->do_interrupt ==
&handle_receive_interrupt) {
__entry->slow_path = 1;
__entry->dma_rtail = 0xFF;
- } else if (dd->rcd[ctxt]->do_interrupt ==
+ } else if (rcd->do_interrupt ==
&handle_receive_interrupt_dma_rtail){
__entry->dma_rtail = 1;
__entry->slow_path = 0;
- } else if (dd->rcd[ctxt]->do_interrupt ==
+ } else if (rcd->do_interrupt ==
&handle_receive_interrupt_nodma_rtail) {
__entry->dma_rtail = 0;
__entry->slow_path = 0;
--- a/drivers/infiniband/hw/hfi1/vnic_main.c
+++ b/drivers/infiniband/hw/hfi1/vnic_main.c
@@ -146,11 +146,7 @@ static int allocate_vnic_ctxt(struct hfi
return ret;
bail:
- /*
- * hfi1_free_ctxt() will call hfi1_free_ctxtdata(), which will
- * release send_context structure if uctxt->sc is not null
- */
- hfi1_free_ctxt(dd, uctxt);
+ hfi1_free_ctxt(uctxt);
dd_dev_dbg(dd, "vnic allocation failed. rc %d\n", ret);
return ret;
}
@@ -158,15 +154,12 @@ bail:
static void deallocate_vnic_ctxt(struct hfi1_devdata *dd,
struct hfi1_ctxtdata *uctxt)
{
- unsigned long flags;
-
dd_dev_dbg(dd, "closing vnic context %d\n", uctxt->ctxt);
flush_wc();
if (dd->num_msix_entries)
hfi1_reset_vnic_msix_info(uctxt);
- spin_lock_irqsave(&dd->uctxt_lock, flags);
/*
* Disable receive context and interrupt available, reset all
* RcvCtxtCtrl bits to default values.
@@ -189,7 +182,6 @@ static void deallocate_vnic_ctxt(struct
sc_disable(uctxt->sc);
dd->send_contexts[uctxt->sc->sw_index].type = SC_USER;
- spin_unlock_irqrestore(&dd->uctxt_lock, flags);
uctxt->event_flags = 0;
@@ -198,7 +190,7 @@ static void deallocate_vnic_ctxt(struct
hfi1_stats.sps_ctxts--;
- hfi1_free_ctxt(dd, uctxt);
+ hfi1_free_ctxt(uctxt);
}
void hfi1_vnic_setup(struct hfi1_devdata *dd)