Re: [PATCH 4/4] PCI: pci-epf-test: Add support to defer core initialization

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 12/5/2019 4:52 PM, Kishon Vijay Abraham I wrote:
Hi,

On 01/12/19 7:59 pm, Vidya Sagar wrote:
On 11/27/2019 2:50 PM, Kishon Vijay Abraham I wrote:
Hi,

On 13/11/19 2:38 PM, Vidya Sagar wrote:
Add support to defer core initialization and to receive a notifier
when core is ready to accommodate platforms where core is not for
initialization untile reference clock from host is available.

Signed-off-by: Vidya Sagar <vidyas@xxxxxxxxxx>
---
  drivers/pci/endpoint/functions/pci-epf-test.c | 114 ++++++++++++------
  1 file changed, 77 insertions(+), 37 deletions(-)

diff --git a/drivers/pci/endpoint/functions/pci-epf-test.c b/drivers/pci/endpoint/functions/pci-epf-test.c
index bddff15052cc..068024fab544 100644
--- a/drivers/pci/endpoint/functions/pci-epf-test.c
+++ b/drivers/pci/endpoint/functions/pci-epf-test.c
@@ -360,18 +360,6 @@ static void pci_epf_test_cmd_handler(struct work_struct *work)
                 msecs_to_jiffies(1));
  }
-static int pci_epf_test_notifier(struct notifier_block *nb, unsigned long val,
-                 void *data)
-{
-    struct pci_epf *epf = container_of(nb, struct pci_epf, nb);
-    struct pci_epf_test *epf_test = epf_get_drvdata(epf);
-
-    queue_delayed_work(kpcitest_workqueue, &epf_test->cmd_handler,
-               msecs_to_jiffies(1));
-
-    return NOTIFY_OK;
-}
-
  static void pci_epf_test_unbind(struct pci_epf *epf)
  {
      struct pci_epf_test *epf_test = epf_get_drvdata(epf);
@@ -428,6 +416,78 @@ static int pci_epf_test_set_bar(struct pci_epf *epf)
      return 0;
  }
+static int pci_epf_test_core_init(struct pci_epf *epf)
+{
+    struct pci_epf_header *header = epf->header;
+    const struct pci_epc_features *epc_features;
+    struct pci_epc *epc = epf->epc;
+    struct device *dev = &epf->dev;
+    bool msix_capable = false;
+    bool msi_capable = true;
+    int ret;
+
+    epc_features = pci_epc_get_features(epc, epf->func_no);
+    if (epc_features) {
+        msix_capable = epc_features->msix_capable;
+        msi_capable = epc_features->msi_capable;
+    }
+
+    ret = pci_epc_write_header(epc, epf->func_no, header);
+    if (ret) {
+        dev_err(dev, "Configuration header write failed\n");
+        return ret;
+    }
+
+    ret = pci_epf_test_set_bar(epf);
+    if (ret)
+        return ret;
+
+    if (msi_capable) {
+        ret = pci_epc_set_msi(epc, epf->func_no, epf->msi_interrupts);
+        if (ret) {
+            dev_err(dev, "MSI configuration failed\n");
+            return ret;
+        }
+    }
+
+    if (msix_capable) {
+        ret = pci_epc_set_msix(epc, epf->func_no, epf->msix_interrupts);
+        if (ret) {
+            dev_err(dev, "MSI-X configuration failed\n");
+            return ret;
+        }
+    }
+
+    return 0;
+}
+
+static int pci_epf_test_notifier(struct notifier_block *nb, unsigned long val,
+                 void *data)
+{
+    struct pci_epf *epf = container_of(nb, struct pci_epf, nb);
+    struct pci_epf_test *epf_test = epf_get_drvdata(epf);
+    int ret;
+
+    switch (val) {
+    case CORE_INIT:
+        ret = pci_epf_test_core_init(epf);
+        if (ret)
+            return NOTIFY_BAD;
+        break;
+
+    case LINK_UP:
+        queue_delayed_work(kpcitest_workqueue, &epf_test->cmd_handler,
+                   msecs_to_jiffies(1));
+        break;
+
+    default:
+        dev_err(&epf->dev, "Invalid EPF test notifier event\n");
+        return NOTIFY_BAD;
+    }
+
+    return NOTIFY_OK;
+}
+
  static int pci_epf_test_alloc_space(struct pci_epf *epf)
  {
      struct pci_epf_test *epf_test = epf_get_drvdata(epf);
@@ -496,12 +556,11 @@ static int pci_epf_test_bind(struct pci_epf *epf)
  {
      int ret;
      struct pci_epf_test *epf_test = epf_get_drvdata(epf);
-    struct pci_epf_header *header = epf->header;
      const struct pci_epc_features *epc_features;
      enum pci_barno test_reg_bar = BAR_0;
      struct pci_epc *epc = epf->epc;
-    struct device *dev = &epf->dev;
      bool linkup_notifier = false;
+    bool skip_core_init = false;
      bool msix_capable = false;
      bool msi_capable = true;
@@ -511,6 +570,7 @@ static int pci_epf_test_bind(struct pci_epf *epf)
      epc_features = pci_epc_get_features(epc, epf->func_no);
      if (epc_features) {
          linkup_notifier = epc_features->linkup_notifier;
+        skip_core_init = epc_features->skip_core_init;
          msix_capable = epc_features->msix_capable;
          msi_capable = epc_features->msi_capable;

Are these used anywhere in this function?
Nope. I'll remove them.

          test_reg_bar = pci_epc_get_first_free_bar(epc_features);
@@ -520,34 +580,14 @@ static int pci_epf_test_bind(struct pci_epf *epf)
      epf_test->test_reg_bar = test_reg_bar;
      epf_test->epc_features = epc_features;
-    ret = pci_epc_write_header(epc, epf->func_no, header);
-    if (ret) {
-        dev_err(dev, "Configuration header write failed\n");
-        return ret;
-    }
-
      ret = pci_epf_test_alloc_space(epf);
      if (ret)
          return ret;
-    ret = pci_epf_test_set_bar(epf);
-    if (ret)
-        return ret;
-
-    if (msi_capable) {
-        ret = pci_epc_set_msi(epc, epf->func_no, epf->msi_interrupts);
-        if (ret) {
-            dev_err(dev, "MSI configuration failed\n");
-            return ret;
-        }
-    }
-
-    if (msix_capable) {
-        ret = pci_epc_set_msix(epc, epf->func_no, epf->msix_interrupts);
-        if (ret) {
-            dev_err(dev, "MSI-X configuration failed\n");
+    if (!skip_core_init) {
+        ret = pci_epf_test_core_init(epf);
+        if (ret)
              return ret;
-        }
      }
      if (linkup_notifier) {

This could as well be moved to pci_epf_test_core_init().
Yes, but I would like to keep only the code that touches hardware in pci_epf_test_core_init()
to minimize the time it takes to execute it. Is there any strong reason to move it? if not,
I would prefer to leave it here in this function itself.

There is no point in scheduling a work to check for commands from host when the EP itself is not initialized.
True. But, since this is more of preparatory work, I thought we should just have it done here itself.
Main reason being, once PERST is perceived, endpoint can't take too much initializing its core. So, I want to
keep that part as minimalistic as possible.

- Vidya Sagar


Thanks
Kishon





[Index of Archives]     [DMA Engine]     [Linux Coverity]     [Linux USB]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Greybus]

  Powered by Linux