Re: [PATCH v7 27/50] powerpc/powernv: Dynamically release PEs

[Gavin Shan <gwshan@linux.vnet.ibm.com>]

On Wed, Nov 18, 2015 at 01:23:05PM +1100, Alexey Kardashevskiy wrote:
>On 11/05/2015 12:12 AM, Gavin Shan wrote:
>>This adds a reference count of PE, representing the number of PCI
>>devices associated with the PE. The reference count is increased
>>or decreased when PCI devices join or leave the PE. Once it becomes
>>zero, the PE together with its used resources (IO, MMIO, DMA, PELTM,
>>PELTV) are released to support PCI hot unplug.
>
>
>The commit log suggest the patch only adds a counter, initializes it, and
>replaces unconditional release of an object (in this case - PE) with the
>conditional one. But it is more that that...
>

Yes, it's more than that as stated in the commit log.

>>Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
>>---
>>  arch/powerpc/platforms/powernv/pci-ioda.c | 245 ++++++++++++++++++++++++++----
>>  arch/powerpc/platforms/powernv/pci.h      |   1 +
>>  2 files changed, 218 insertions(+), 28 deletions(-)
>>
>>diff --git a/arch/powerpc/platforms/powernv/pci-ioda.c b/arch/powerpc/platforms/powernv/pci-ioda.c
>>index 0bb0056..dcffce5 100644
>>--- a/arch/powerpc/platforms/powernv/pci-ioda.c
>>+++ b/arch/powerpc/platforms/powernv/pci-ioda.c
>>@@ -129,6 +129,215 @@ static inline bool pnv_pci_is_mem_pref_64(unsigned long flags)
>>  		(IORESOURCE_MEM_64 | IORESOURCE_PREFETCH));
>>  }
>>
>>+static void pnv_pci_ioda1_release_dma_pe(struct pnv_ioda_pe *pe)
>>+{
>>+	struct pnv_phb *phb = pe->phb;
>>+	struct iommu_table *tbl;
>>+	int start, count, i;
>>+	int64_t rc;
>>+
>>+	/* Search for the used DMA32 segments */
>>+	start = -1;
>>+	count = 0;
>>+	for (i = 0; i < phb->ioda.dma32_count; i++) {
>>+		if (phb->ioda.dma32_segmap[i] != pe->pe_number)
>>+			continue;
>>+
>>+		count++;
>>+		if (start < 0)
>>+			start = i;
>>+	}
>>+
>>+	if (!count)
>>+		return;
>
>
>imho checking pe->table_group.tables[0] != NULL is shorter than the loop above.
>

Will use it in next revision.

>>+
>>+	/* Unlink IOMMU table from group */
>>+	tbl = pe->table_group.tables[0];
>>+	pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
>>+	if (pe->table_group.group) {
>>+		iommu_group_put(pe->table_group.group);
>>+		WARN_ON(pe->table_group.group);
>>+	}
>>+
>>+	/* Release IOMMU table */
>>+	pnv_pci_ioda2_table_free_pages(tbl);
>
>
>This is IODA2 helper with multilevel support, does IODA1 support multilevel
>TCE tables? If not, it should WARN_ON on levels!=1.
>
>Another thing is you should first unprogram TVEs (via
>opal_pci_map_pe_dma_window), then invalidate the cache (if required, not sure
>if this is needed on IODA1), only then free the actual table.
>
>
>>+	iommu_free_table(tbl, of_node_full_name(pci_bus_to_OF_node(pe->pbus)));
>>+
>>+	/* Disable TVE */
>>+	for (i = start; i < start + count; i++) {
>>+		rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
>>+						i, 0, 0ul, 0ul, 0ul);
>>+		if (rc)
>>+			pe_warn(pe, "Error %ld unmapping DMA32 seg#%d\n",
>>+				rc, i);
>>+
>>+		phb->ioda.dma32_segmap[i] = IODA_INVALID_PE;
>>+	}
>
>
>You could implement pnv_pci_ioda1_unset_window/pnv_ioda1_table_free as
>callbacks, change pnv_pci_ioda2_release_dma_pe() to use them (and rename it
>to reflect that it supports IODA1 and IODA2).
>
>
>>+}
>>+
>>+static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe);
>>+static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
>>+		int num);
>>+static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
>>+
>>+static void pnv_pci_ioda2_release_dma_pe(struct pnv_ioda_pe *pe)
>
>
>You moved this function and changed it, please do one thing at once (which is
>"change", not "move").
>
>>+{
>>+	struct iommu_table *tbl;
>>+	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
>>+	int64_t rc;
>>+
>>+	if (!weight)
>>+		return;
>
>
>Checking for pe->table_group.group is better because if we ever change the
>logic of what gets included to an IOMMU group, we will have to do the change
>where we add devices to a group but we won't have to touch releasing code.
>
>
>>+
>>+	tbl = pe->table_group.tables[0];
>>+	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
>>+	if (rc)
>>+		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
>>+
>>+	pnv_pci_ioda2_set_bypass(pe, false);
>>+	if (pe->table_group.group) {
>>+		iommu_group_put(pe->table_group.group);
>>+		WARN_ON(pe->table_group.group);
>>+	}
>>+
>>+	pnv_pci_ioda2_table_free_pages(tbl);
>>+	iommu_free_table(tbl, "pnv");
>>+}
>>+
>>+static void pnv_ioda_release_dma_pe(struct pnv_ioda_pe *pe)
>
>Merge this into pnv_ioda_release_pe() - it is small and called just once.
>
>
>>+{
>>+	struct pnv_phb *phb = pe->phb;
>>+
>>+	switch (phb->type) {
>>+	case PNV_PHB_IODA1:
>>+		pnv_pci_ioda1_release_dma_pe(pe);
>>+		break;
>>+	case PNV_PHB_IODA2:
>>+		pnv_pci_ioda2_release_dma_pe(pe);
>>+		break;
>>+	default:
>>+		WARN_ON(1);
>>+	}
>>+}
>>+
>>+static void pnv_ioda_release_window(struct pnv_ioda_pe *pe, int win)
>>+{
>>+	struct pnv_phb *phb = pe->phb;
>>+	int index, *segmap = NULL;
>>+	int64_t rc;
>>+
>>+	switch (win) {
>>+	case OPAL_IO_WINDOW_TYPE:
>>+		segmap = phb->ioda.io_segmap;
>>+		break;
>>+	case OPAL_M32_WINDOW_TYPE:
>>+		segmap = phb->ioda.m32_segmap;
>>+		break;
>>+	case OPAL_M64_WINDOW_TYPE:
>>+		if (phb->type != PNV_PHB_IODA1)
>>+			return;
>>+		segmap = phb->ioda.m64_segmap;
>>+		break;
>>+	default:
>>+		return;
>
>Unnecessary return.
>
>
>>+	}
>>+
>>+	for (index = 0; index < phb->ioda.total_pe_num; index++) {
>>+		if (segmap[index] != pe->pe_number)
>>+			continue;
>>+
>>+		if (win == OPAL_M64_WINDOW_TYPE)
>>+			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
>>+					phb->ioda.reserved_pe_idx, win,
>>+					index / PNV_IODA1_M64_SEGS,
>>+					index % PNV_IODA1_M64_SEGS);
>>+		else
>>+			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
>>+					phb->ioda.reserved_pe_idx, win,
>>+					0, index);
>>+
>>+		if (rc != OPAL_SUCCESS)
>>+			pe_warn(pe, "Error %ld unmapping (%d) segment#%d\n",
>>+				rc, win, index);
>>+
>>+		segmap[index] = IODA_INVALID_PE;
>>+	}
>>+}
>>+
>>+static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
>>+{
>>+	struct pnv_phb *phb = pe->phb;
>>+	int win;
>>+
>>+	for (win = OPAL_M32_WINDOW_TYPE; win <= OPAL_IO_WINDOW_TYPE; win++) {
>>+		if (phb->type == PNV_PHB_IODA2 && win == OPAL_IO_WINDOW_TYPE)
>>+			continue;
>
>Move this check to pnv_ioda_release_window() or move case(win ==
>OPAL_M64_WINDOW_TYPE):if(phb->type != PNV_PHB_IODA1) from that function here.
>
>
>>+
>>+		pnv_ioda_release_window(pe, win);
>>+	}
>>+}
>
>This is shorter and cleaner:
>
>
>static void pnv_ioda_release_window(struct pnv_ioda_pe *pe, int win, int
>*segmap
>{
>        struct pnv_phb *phb = pe->phb;
>        int index;
>        int64_t rc;
>
>        for (index = 0; index < phb->ioda.total_pe_num; index++) {
>                if (segmap[index] != pe->pe_number)
>                        continue;
>
>                if (win == OPAL_M64_WINDOW_TYPE)
>                        rc = opal_pci_map_pe_mmio_window(phb->opal_id,
>                                        phb->ioda.reserved_pe_idx, win,
>                                        index / PNV_IODA1_M64_SEGS,
>                                        index % PNV_IODA1_M64_SEGS);
>                else
>                        rc = opal_pci_map_pe_mmio_window(phb->opal_id,
>                                        phb->ioda.reserved_pe_idx, win,
>                                        0, index);
>
>                if (rc != OPAL_SUCCESS)
>                        pe_warn(pe, "Error %ld unmapping (%d) segment#%d\n",
>                                rc, win, index);
>
>                segmap[index] = IODA_INVALID_PE;
>        }
>}
>
>static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
>{
>        pnv_ioda_release_window(pe, OPAL_M32_WINDOW_TYPE,
>phb->ioda.m32_segmap);
>        if (phb->type != PNV_PHB_IODA2)
>                pnv_ioda_release_window(pe, OPAL_IO_WINDOW_TYPE,
>                                phb->ioda.io_segmap);
>	else
>                pnv_ioda_release_window(pe, OPAL_M64_WINDOW_TYPE,
>                                phb->ioda.m64_segmap);
>}
>
>
>I'd actually merge pnv_ioda_release_pe_seg() into pnv_ioda_release_pe() as
>well as it is also small and called once.
>
>
>>+
>>+static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb,
>>+				   struct pnv_ioda_pe *pe);
>>+static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe);
>>+static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
>>+{
>>+	struct pnv_ioda_pe *tmp, *slave;
>>+
>>+	/* Release slave PEs in compound PE */
>>+	if (pe->flags & PNV_IODA_PE_MASTER) {
>>+		list_for_each_entry_safe(slave, tmp, &pe->slaves, list)
>>+			pnv_ioda_release_pe(slave);
>>+	}
>>+
>>+	/* Remove the PE from the list */
>>+	list_del(&pe->list);
>>+
>>+	/* Release resources */
>>+	pnv_ioda_release_dma_pe(pe);
>>+	pnv_ioda_release_pe_seg(pe);
>>+	pnv_ioda_deconfigure_pe(pe->phb, pe);
>>+
>>+	pnv_ioda_free_pe(pe);
>>+}
>>+
>>+static inline struct pnv_ioda_pe *pnv_ioda_pe_get(struct pnv_ioda_pe *pe)
>>+{
>>+	if (!pe)
>>+		return NULL;
>>+
>>+	pe->device_count++;
>>+	return pe;
>>+}
>>+
>>+static inline void pnv_ioda_pe_put(struct pnv_ioda_pe *pe)
>
>
>Merge this into pnv_pci_release_device() as it is small and called only once.
>

I don't think so. The functions pnv_ioda_pe_{get,put}() are paired. I think it's
good enough to have separate function for the logic included in pnv_ioda_pe_put().

>>+{
>>+	if (!pe)
>>+		return;
>>+
>>+	pe->device_count--;
>>+	WARN_ON(pe->device_count < 0);
>>+	if (pe->device_count == 0)
>>+		pnv_ioda_release_pe(pe);
>>+}
>>+
>>+static void pnv_pci_release_device(struct pci_dev *pdev)
>>+{
>>+	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
>>+	struct pnv_phb *phb = hose->private_data;
>>+	struct pci_dn *pdn = pci_get_pdn(pdev);
>>+	struct pnv_ioda_pe *pe;
>>+
>>+	if (pdev->is_virtfn)
>>+		return;
>>+
>>+	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
>>+		return;
>>+
>>+	pe = &phb->ioda.pe_array[pdn->pe_number];
>>+	pnv_ioda_pe_put(pe);
>>+}
>>+
>>  static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
>>  {
>>  	phb->ioda.pe_array[pe_no].phb = phb;
>>@@ -724,7 +933,6 @@ static int pnv_ioda_set_peltv(struct pnv_phb *phb,
>>  	return 0;
>>  }
>>
>>-#ifdef CONFIG_PCI_IOV
>>  static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
>>  {
>>  	struct pci_dev *parent;
>>@@ -759,9 +967,11 @@ static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
>>  		}
>>  		rid_end = pe->rid + (count << 8);
>>  	} else {
>>+#ifdef CONFIG_PCI_IOV
>>  		if (pe->flags & PNV_IODA_PE_VF)
>>  			parent = pe->parent_dev;
>>  		else
>>+#endif
>>  			parent = pe->pdev->bus->self;
>>  		bcomp = OpalPciBusAll;
>>  		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
>>@@ -799,11 +1009,12 @@ static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
>>
>>  	pe->pbus = NULL;
>>  	pe->pdev = NULL;
>>+#ifdef CONFIG_PCI_IOV
>>  	pe->parent_dev = NULL;
>>+#endif
>
>
>These #ifdef movements seem very much unrelated.
>

It's related: pnv_ioda_deconfigure_pe() was used for VF PE only. Now it's used by all
types of PEs. pe->parent_dev is declared as below:

#ifdef CONFIG_PCI_IOV
        struct pci_dev          *parent_dev;
#endif

>
>>
>>  	return 0;
>>  }
>>-#endif /* CONFIG_PCI_IOV */
>>
>>  static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
>>  {
>>@@ -985,6 +1196,7 @@ static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
>>  			continue;
>>
>>  		pdn->pe_number = pe->pe_number;
>>+		pnv_ioda_pe_get(pe);
>>  		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
>>  			pnv_ioda_setup_same_PE(dev->subordinate, pe);
>>  	}
>>@@ -1047,9 +1259,8 @@ static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
>>  			bus->busn_res.start, pe->pe_number);
>>
>>  	if (pnv_ioda_configure_pe(phb, pe)) {
>>-		/* XXX What do we do here ? */
>>-		pnv_ioda_free_pe(pe);
>>  		pe->pbus = NULL;
>>+		pnv_ioda_release_pe(pe);
>
>
>This is unrelated unexplained change.
>

Will drop it in next revision.

>>  		return NULL;
>>  	}
>>
>>@@ -1199,29 +1410,6 @@ m64_failed:
>>  	return -EBUSY;
>>  }
>>
>>-static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
>>-		int num);
>>-static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
>>-
>>-static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
>>-{
>>-	struct iommu_table    *tbl;
>>-	int64_t               rc;
>>-
>>-	tbl = pe->table_group.tables[0];
>>-	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
>>-	if (rc)
>>-		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
>>-
>>-	pnv_pci_ioda2_set_bypass(pe, false);
>>-	if (pe->table_group.group) {
>>-		iommu_group_put(pe->table_group.group);
>>-		BUG_ON(pe->table_group.group);
>>-	}
>>-	pnv_pci_ioda2_table_free_pages(tbl);
>>-	iommu_free_table(tbl, of_node_full_name(dev->dev.of_node));
>>-}
>>-
>>  static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
>>  {
>>  	struct pci_bus        *bus;
>>@@ -1242,7 +1430,7 @@ static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
>>  		if (pe->parent_dev != pdev)
>>  			continue;
>>
>>-		pnv_pci_ioda2_release_dma_pe(pdev, pe);
>>+		pnv_pci_ioda2_release_dma_pe(pe);
>
>
>This is unrelated change.
>


>>
>>  		/* Remove from list */
>>  		mutex_lock(&phb->ioda.pe_list_mutex);
>>@@ -3124,6 +3312,7 @@ static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
>>  	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
>>  #endif
>>  	.enable_device_hook	= pnv_pci_enable_device_hook,
>>+	.release_device		= pnv_pci_release_device,
>>  	.window_alignment	= pnv_pci_window_alignment,
>>  	.setup_bridge		= pnv_pci_setup_bridge,
>>  	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
>>diff --git a/arch/powerpc/platforms/powernv/pci.h b/arch/powerpc/platforms/powernv/pci.h
>>index ef5271a..3bb10de 100644
>>--- a/arch/powerpc/platforms/powernv/pci.h
>>+++ b/arch/powerpc/platforms/powernv/pci.h
>>@@ -30,6 +30,7 @@ struct pnv_phb;
>>  struct pnv_ioda_pe {
>>  	unsigned long		flags;
>>  	struct pnv_phb		*phb;
>>+	int			device_count;
>
>Not atomic_t, no kref, no additional mutex, just "int"? Sure about it? If so,
>put a note to the commit log about what provides a guarantee that there is no
>race.
>
>

It was a kref. Something you suggested on v5 as below:

| You do not need kref here. You call kref_put() in a single location and can do
| stuff directly, without kref. Just have an "unsigned int" counter and that's
| it (it does not even have to be atomic if you do not have races but I am not
| sure you do not).
|

>>
>>  	/* A PE can be associated with a single device or an
>>  	 * entire bus (& children). In the former case, pdev
>>

Thanks,
Gavin