The rcd array can be accessed from user context or during interrupts.
Protecting this with a mutex isn't a good idea because the mutex should
not be used from an IRQ.
Protect the allocation and freeing of rcd array elements with a
spinlock.
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>
if (ret)
goto bail_cleanup;
- ret = hfi1_create_ctxts(dd);
+ ret = hfi1_create_kctxts(dd);
if (ret)
goto bail_cleanup;
+ /*
+ * Initialize aspm, to be done after gen3 transition and setting up
+ * contexts and before enabling interrupts
+ */
+ aspm_init(dd);
+
dd->rcvhdrsize = DEFAULT_RCVHDRSIZE;
/*
* rcd[0] is guaranteed to be valid by this point. Also, all
goto bail_cleanup;
}
- /* use contexts created by hfi1_create_ctxts */
+ /* use contexts created by hfi1_create_kctxts */
ret = set_up_interrupts(dd);
if (ret)
goto bail_cleanup;
static u64 kvirt_to_phys(void *addr);
static int assign_ctxt(struct hfi1_filedata *fd, struct hfi1_user_info *uinfo);
-static int init_subctxts(struct hfi1_ctxtdata *uctxt,
- const struct hfi1_user_info *uinfo);
+static void init_subctxts(struct hfi1_ctxtdata *uctxt,
+ const struct hfi1_user_info *uinfo);
static int init_user_ctxt(struct hfi1_filedata *fd,
struct hfi1_ctxtdata *uctxt);
static void user_init(struct hfi1_ctxtdata *uctxt);
goto done;
hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
- mutex_lock(&hfi1_mutex);
flush_wc();
/* drain user sdma queue */
HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
*ev = 0;
+ mutex_lock(&hfi1_mutex);
__clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
fdata->uctxt = NULL;
hfi1_rcd_put(uctxt); /* fdata reference */
return paddr;
}
+static int complete_subctxt(struct hfi1_filedata *fd)
+{
+ int ret;
+
+ /*
+ * sub-context info can only be set up after the base context
+ * has been completed.
+ */
+ ret = wait_event_interruptible(
+ fd->uctxt->wait,
+ !test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
+
+ 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 */
+ if (!ret) {
+ fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
+ ret = init_user_ctxt(fd, fd->uctxt);
+ }
+
+ if (ret) {
+ hfi1_rcd_put(fd->uctxt);
+ fd->uctxt = NULL;
+ mutex_lock(&hfi1_mutex);
+ __clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
+ mutex_unlock(&hfi1_mutex);
+ }
+
+ return ret;
+}
+
static int assign_ctxt(struct hfi1_filedata *fd, struct hfi1_user_info *uinfo)
{
int ret;
if (swmajor != HFI1_USER_SWMAJOR)
return -ENODEV;
+ if (uinfo->subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
+ return -EINVAL;
+
swminor = uinfo->userversion & 0xffff;
+ /*
+ * Acquire the mutex to protect against multiple creations of what
+ * could be a shared base context.
+ */
mutex_lock(&hfi1_mutex);
/*
- * Get a sub context if necessary.
+ * Get a sub context if available (fd->uctxt will be set).
* ret < 0 error, 0 no context, 1 sub-context found
*/
- ret = 0;
- if (uinfo->subctxt_cnt) {
- ret = find_sub_ctxt(fd, uinfo);
- if (ret > 0)
- fd->rec_cpu_num =
- hfi1_get_proc_affinity(fd->uctxt->numa_id);
- }
+ ret = find_sub_ctxt(fd, uinfo);
/*
- * Allocate a base context if context sharing is not required or we
- * couldn't find a sub context.
+ * Allocate a base context if context sharing is not required or a
+ * sub context wasn't found.
*/
if (!ret)
ret = allocate_ctxt(fd, fd->dd, uinfo, &uctxt);
mutex_unlock(&hfi1_mutex);
/* Depending on the context type, do the appropriate init */
- if (ret > 0) {
- /*
- * sub-context info can only be set up after the base
- * context has been completed.
- */
- ret = wait_event_interruptible(fd->uctxt->wait, !test_bit(
- HFI1_CTXT_BASE_UNINIT,
- &fd->uctxt->event_flags));
- 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 */
- if (!ret)
- ret = init_user_ctxt(fd, fd->uctxt);
-
- if (ret)
- clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
-
- } else if (!ret) {
+ switch (ret) {
+ case 0:
ret = setup_base_ctxt(fd, uctxt);
if (uctxt->subctxt_cnt) {
- /* If there is an error, set the failed bit. */
- if (ret)
- set_bit(HFI1_CTXT_BASE_FAILED,
- &uctxt->event_flags);
/*
* Base context is done, 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);
}
- if (ret)
- deallocate_ctxt(uctxt);
- }
-
- /* If an error occurred, clear the reference */
- if (ret && fd->uctxt) {
- hfi1_rcd_put(fd->uctxt);
- fd->uctxt = NULL;
+ break;
+ case 1:
+ ret = complete_subctxt(fd);
+ break;
+ default:
+ break;
}
return ret;
/*
* The hfi1_mutex must be held when this function is called. It is
- * necessary to ensure serialized access to the bitmask in_use_ctxts.
+ * necessary to ensure serialized creation of shared contexts.
*/
static int find_sub_ctxt(struct hfi1_filedata *fd,
const struct hfi1_user_info *uinfo)
struct hfi1_devdata *dd = fd->dd;
u16 subctxt;
+ 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];
struct hfi1_ctxtdata **cd)
{
struct hfi1_ctxtdata *uctxt;
- u16 ctxt;
int ret, numa;
if (dd->flags & HFI1_FROZEN) {
return -EIO;
}
- /*
- * This check is sort of redundant to the next EBUSY error. It would
- * also indicate an inconsistancy in the driver if this value was
- * zero, but there were still contexts available.
- */
if (!dd->freectxts)
return -EBUSY;
- for (ctxt = dd->first_dyn_alloc_ctxt;
- ctxt < dd->num_rcv_contexts; ctxt++)
- if (!dd->rcd[ctxt])
- break;
-
- if (ctxt == dd->num_rcv_contexts)
- return -EBUSY;
-
/*
* If we don't have a NUMA node requested, preference is towards
* device NUMA node.
numa = cpu_to_node(fd->rec_cpu_num);
else
numa = numa_node_id();
- uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
- if (!uctxt) {
- dd_dev_err(dd,
- "Unable to allocate ctxtdata memory, failing open\n");
- return -ENOMEM;
+ ret = hfi1_create_ctxtdata(dd->pport, numa, &uctxt);
+ if (ret < 0) {
+ dd_dev_err(dd, "user ctxtdata allocation failed\n");
+ return ret;
}
hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
/*
* Allocate and enable a PIO send context.
*/
- uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
- uctxt->dd->node);
+ uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize, dd->node);
if (!uctxt->sc) {
ret = -ENOMEM;
goto ctxdata_free;
goto ctxdata_free;
/*
- * Setup sub context resources if the user-level has requested
+ * Setup sub context information if the user-level has requested
* sub contexts.
* This has to be done here so the rest of the sub-contexts find the
- * proper master.
+ * proper base context.
*/
- if (uinfo->subctxt_cnt) {
- ret = init_subctxts(uctxt, uinfo);
- /*
- * On error, we don't need to disable and de-allocate the
- * send context because it will be done during file close
- */
- if (ret)
- goto ctxdata_free;
- }
+ if (uinfo->subctxt_cnt)
+ init_subctxts(uctxt, uinfo);
uctxt->userversion = uinfo->userversion;
uctxt->flags = hfi1_cap_mask; /* save current flag state */
init_waitqueue_head(&uctxt->wait);
return 0;
ctxdata_free:
- *cd = NULL;
- dd->rcd[ctxt] = NULL;
- hfi1_rcd_put(uctxt);
+ hfi1_free_ctxt(dd, uctxt);
return ret;
}
hfi1_stats.sps_ctxts--;
if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
aspm_enable_all(uctxt->dd);
-
- /* _rcd_put() should be done after releasing mutex */
- uctxt->dd->rcd[uctxt->ctxt] = NULL;
mutex_unlock(&hfi1_mutex);
- hfi1_rcd_put(uctxt); /* dd reference */
+
+ hfi1_free_ctxt(uctxt->dd, uctxt);
}
-static int init_subctxts(struct hfi1_ctxtdata *uctxt,
- const struct hfi1_user_info *uinfo)
+static void init_subctxts(struct hfi1_ctxtdata *uctxt,
+ const struct hfi1_user_info *uinfo)
{
- u16 num_subctxts;
-
- num_subctxts = uinfo->subctxt_cnt;
- if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
- return -EINVAL;
-
uctxt->subctxt_cnt = uinfo->subctxt_cnt;
uctxt->subctxt_id = uinfo->subctxt_id;
- uctxt->redirect_seq_cnt = 1;
set_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
-
- return 0;
}
static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
return 0;
setup_failed:
- /* Call _free_ctxtdata, not _rcd_put(). We still need the context. */
- hfi1_free_ctxtdata(dd, uctxt);
+ set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
+ deallocate_ctxt(uctxt);
return ret;
}
u16 poll_type;
/* receive packet sequence counter */
u8 seq_cnt;
- u8 redirect_seq_cnt;
/* ctxt rcvhdrq head offset */
u32 head;
/* QPs waiting for context processing */
int hfi1_create_rcvhdrq(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd);
int hfi1_setup_eagerbufs(struct hfi1_ctxtdata *rcd);
-int hfi1_create_ctxts(struct hfi1_devdata *dd);
-struct hfi1_ctxtdata *hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, u16 ctxt,
- int numa);
+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_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);
u32 hfi1_cpulist_count;
unsigned long *hfi1_cpulist;
-/*
- * Common code for creating the receive context array.
- */
-int hfi1_create_ctxts(struct hfi1_devdata *dd)
+static int hfi1_create_kctxt(struct hfi1_devdata *dd,
+ struct hfi1_pportdata *ppd)
{
- u16 i;
+ struct hfi1_ctxtdata *rcd;
int ret;
/* Control context has to be always 0 */
BUILD_BUG_ON(HFI1_CTRL_CTXT != 0);
+ ret = hfi1_create_ctxtdata(ppd, dd->node, &rcd);
+ if (ret < 0) {
+ dd_dev_err(dd, "Kernel receive context allocation failed\n");
+ return ret;
+ }
+
+ /*
+ * Set up the kernel context flags here and now because they use
+ * default values for all receive side memories. User contexts will
+ * be handled as they are created.
+ */
+ rcd->flags = HFI1_CAP_KGET(MULTI_PKT_EGR) |
+ HFI1_CAP_KGET(NODROP_RHQ_FULL) |
+ HFI1_CAP_KGET(NODROP_EGR_FULL) |
+ HFI1_CAP_KGET(DMA_RTAIL);
+
+ /* Control context must use DMA_RTAIL */
+ if (rcd->ctxt == HFI1_CTRL_CTXT)
+ rcd->flags |= HFI1_CAP_DMA_RTAIL;
+ rcd->seq_cnt = 1;
+
+ rcd->sc = sc_alloc(dd, SC_ACK, rcd->rcvhdrqentsize, dd->node);
+ if (!rcd->sc) {
+ dd_dev_err(dd, "Kernel send context allocation failed\n");
+ return -ENOMEM;
+ }
+ hfi1_init_ctxt(rcd->sc);
+
+ return 0;
+}
+
+/*
+ * Create the receive context array and one or more kernel contexts
+ */
+int hfi1_create_kctxts(struct hfi1_devdata *dd)
+{
+ u16 i;
+ int ret;
+
dd->rcd = kzalloc_node(dd->num_rcv_contexts * sizeof(*dd->rcd),
GFP_KERNEL, dd->node);
if (!dd->rcd)
- goto nomem;
+ return -ENOMEM;
- /* create one or more kernel contexts */
for (i = 0; i < dd->first_dyn_alloc_ctxt; ++i) {
- struct hfi1_pportdata *ppd;
- struct hfi1_ctxtdata *rcd;
-
- ppd = dd->pport + (i % dd->num_pports);
-
- /* dd->rcd[i] gets assigned inside the callee */
- rcd = hfi1_create_ctxtdata(ppd, i, dd->node);
- if (!rcd) {
- dd_dev_err(dd,
- "Unable to allocate kernel receive context, failing\n");
- goto nomem;
- }
- /*
- * Set up the kernel context flags here and now because they
- * use default values for all receive side memories. User
- * contexts will be handled as they are created.
- */
- rcd->flags = HFI1_CAP_KGET(MULTI_PKT_EGR) |
- HFI1_CAP_KGET(NODROP_RHQ_FULL) |
- HFI1_CAP_KGET(NODROP_EGR_FULL) |
- HFI1_CAP_KGET(DMA_RTAIL);
-
- /* Control context must use DMA_RTAIL */
- if (rcd->ctxt == HFI1_CTRL_CTXT)
- rcd->flags |= HFI1_CAP_DMA_RTAIL;
- rcd->seq_cnt = 1;
-
- rcd->sc = sc_alloc(dd, SC_ACK, rcd->rcvhdrqentsize, dd->node);
- if (!rcd->sc) {
- dd_dev_err(dd,
- "Unable to allocate kernel send context, failing\n");
- goto nomem;
- }
-
- hfi1_init_ctxt(rcd->sc);
+ ret = hfi1_create_kctxt(dd, dd->pport);
+ if (ret)
+ goto bail;
}
- /*
- * Initialize aspm, to be done after gen3 transition and setting up
- * contexts and before enabling interrupts
- */
- aspm_init(dd);
-
return 0;
-nomem:
- ret = -ENOMEM;
-
+bail:
for (i = 0; dd->rcd && i < dd->first_dyn_alloc_ctxt; ++i)
hfi1_rcd_put(dd->rcd[i]);
kref_init(&rcd->kref);
}
+/**
+ * hfi1_rcd_free - When reference is zero clean up.
+ * @kref: pointer to an initialized rcd data structure
+ *
+ */
static void hfi1_rcd_free(struct kref *kref)
{
struct hfi1_ctxtdata *rcd =
kfree(rcd);
}
+/**
+ * hfi1_rcd_put - decrement reference for rcd
+ * @rcd: pointer to an initialized rcd data structure
+ *
+ * Use this to put a reference after the init.
+ */
int hfi1_rcd_put(struct hfi1_ctxtdata *rcd)
{
if (rcd)
return 0;
}
+/**
+ * hfi1_rcd_get - increment reference for rcd
+ * @rcd: pointer to an initialized rcd data structure
+ *
+ * Use this to get a reference after the init.
+ */
void hfi1_rcd_get(struct hfi1_ctxtdata *rcd)
{
kref_get(&rcd->kref);
}
+/**
+ * allocate_rcd_index - allocate an rcd index from the rcd array
+ * @dd: pointer to a valid devdata structure
+ * @rcd: rcd data structure to assign
+ * @index: pointer to index that is allocated
+ *
+ * Find an empty index in the rcd array, and assign the given rcd to it.
+ * If the array is full, we are EBUSY.
+ *
+ */
+static u16 allocate_rcd_index(struct hfi1_devdata *dd,
+ struct hfi1_ctxtdata *rcd, u16 *index)
+{
+ unsigned long flags;
+ u16 ctxt;
+
+ spin_lock_irqsave(&dd->uctxt_lock, flags);
+ for (ctxt = 0; ctxt < dd->num_rcv_contexts; ctxt++)
+ if (!dd->rcd[ctxt])
+ break;
+
+ if (ctxt < dd->num_rcv_contexts) {
+ rcd->ctxt = ctxt;
+ dd->rcd[ctxt] = rcd;
+ hfi1_rcd_init(rcd);
+ }
+ spin_unlock_irqrestore(&dd->uctxt_lock, flags);
+
+ if (ctxt >= dd->num_rcv_contexts)
+ return -EBUSY;
+
+ *index = ctxt;
+
+ return 0;
+}
+
/*
* Common code for user and kernel context setup.
*/
-struct hfi1_ctxtdata *hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, u16 ctxt,
- int numa)
+int hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, int numa,
+ struct hfi1_ctxtdata **context)
{
struct hfi1_devdata *dd = ppd->dd;
struct hfi1_ctxtdata *rcd;
rcd = kzalloc_node(sizeof(*rcd), GFP_KERNEL, numa);
if (rcd) {
u32 rcvtids, max_entries;
+ u16 ctxt;
+ int ret;
hfi1_cdbg(PROC, "setting up context %u\n", ctxt);
+ ret = allocate_rcd_index(dd, rcd, &ctxt);
+ if (ret) {
+ *context = NULL;
+ kfree(rcd);
+ return ret;
+ }
+
INIT_LIST_HEAD(&rcd->qp_wait_list);
hfi1_exp_tid_group_init(&rcd->tid_group_list);
hfi1_exp_tid_group_init(&rcd->tid_used_list);
rcd->ppd = ppd;
rcd->dd = dd;
__set_bit(0, rcd->in_use_ctxts);
- rcd->ctxt = ctxt;
- dd->rcd[ctxt] = rcd;
rcd->numa_id = numa;
rcd->rcv_array_groups = dd->rcv_entries.ngroups;
goto bail;
}
- hfi1_rcd_init(rcd);
+ *context = rcd;
+ return 0;
}
- return rcd;
+
bail:
- dd->rcd[ctxt] = NULL;
- kfree(rcd->egrbufs.rcvtids);
- kfree(rcd->egrbufs.buffers);
- kfree(rcd);
- return NULL;
+ *context = NULL;
+ hfi1_free_ctxt(dd, 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.
+ */
+void hfi1_free_ctxt(struct hfi1_devdata *dd, 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);
+ }
}
/*
struct hfi1_ctxtdata **vnic_ctxt)
{
struct hfi1_ctxtdata *uctxt;
- u16 ctxt;
int ret;
if (dd->flags & HFI1_FROZEN)
return -EIO;
- for (ctxt = dd->first_dyn_alloc_ctxt;
- ctxt < dd->num_rcv_contexts; ctxt++)
- if (!dd->rcd[ctxt])
- break;
-
- if (ctxt == dd->num_rcv_contexts)
- return -EBUSY;
-
- uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, dd->node);
- if (!uctxt) {
+ ret = hfi1_create_ctxtdata(dd->pport, dd->node, &uctxt);
+ if (ret < 0) {
dd_dev_err(dd, "Unable to create ctxtdata, failing open\n");
return -ENOMEM;
}
return ret;
bail:
/*
- * hfi1_rcd_put() will call hfi1_free_ctxtdata(), which will
+ * hfi1_free_ctxt() will call hfi1_free_ctxtdata(), which will
* release send_context structure if uctxt->sc is not null
*/
- dd->rcd[uctxt->ctxt] = NULL;
- hfi1_rcd_put(uctxt);
+ hfi1_free_ctxt(dd, uctxt);
dd_dev_dbg(dd, "vnic allocation failed. rc %d\n", ret);
return ret;
}
dd->send_contexts[uctxt->sc->sw_index].type = SC_USER;
spin_unlock_irqrestore(&dd->uctxt_lock, flags);
- dd->rcd[uctxt->ctxt] = NULL;
uctxt->event_flags = 0;
hfi1_clear_tids(uctxt);
hfi1_clear_ctxt_pkey(dd, uctxt);
hfi1_stats.sps_ctxts--;
- hfi1_rcd_put(uctxt);
+
+ hfi1_free_ctxt(dd, uctxt);
}
void hfi1_vnic_setup(struct hfi1_devdata *dd)
projects / platform / kernel / linux-starfive.git / commitdiff