powerpc/pseries: Add support for dynamic dma windows
Nishanth Aravamudan [Thu, 10 Feb 2011 09:10:47 +0000 (09:10 +0000)]
If firmware allows us to map all of a partition's memory for DMA on a
particular bridge, create a 1:1 mapping of that memory. Add hooks for
dealing with hotplug events. Dynamic DMA windows can use larger than the
default page size, and we use the largest one possible.

Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>

Documentation/kernel-parameters.txt
arch/powerpc/platforms/pseries/iommu.c

index f4a04c0..b245ce6 100644 (file)
@@ -626,6 +626,10 @@ bytes respectively. Such letter suffixes can also be entirely omitted.
        disable=        [IPV6]
                        See Documentation/networking/ipv6.txt.
 
+       disable_ddw     [PPC/PSERIES]
+                       Disable Dynamic DMA Window support. Use this if
+                       to workaround buggy firmware.
+
        disable_ipv6=   [IPV6]
                        See Documentation/networking/ipv6.txt.
 
index edea60b..154c464 100644 (file)
@@ -33,6 +33,7 @@
 #include <linux/pci.h>
 #include <linux/dma-mapping.h>
 #include <linux/crash_dump.h>
+#include <linux/memory.h>
 #include <asm/io.h>
 #include <asm/prom.h>
 #include <asm/rtas.h>
@@ -45,6 +46,7 @@
 #include <asm/tce.h>
 #include <asm/ppc-pci.h>
 #include <asm/udbg.h>
+#include <asm/mmzone.h>
 
 #include "plpar_wrappers.h"
 
@@ -270,6 +272,152 @@ static unsigned long tce_get_pSeriesLP(struct iommu_table *tbl, long tcenum)
        return tce_ret;
 }
 
+/* this is compatable with cells for the device tree property */
+struct dynamic_dma_window_prop {
+       __be32  liobn;          /* tce table number */
+       __be64  dma_base;       /* address hi,lo */
+       __be32  tce_shift;      /* ilog2(tce_page_size) */
+       __be32  window_shift;   /* ilog2(tce_window_size) */
+};
+
+struct direct_window {
+       struct device_node *device;
+       const struct dynamic_dma_window_prop *prop;
+       struct list_head list;
+};
+
+/* Dynamic DMA Window support */
+struct ddw_query_response {
+       u32 windows_available;
+       u32 largest_available_block;
+       u32 page_size;
+       u32 migration_capable;
+};
+
+struct ddw_create_response {
+       u32 liobn;
+       u32 addr_hi;
+       u32 addr_lo;
+};
+
+static LIST_HEAD(direct_window_list);
+/* prevents races between memory on/offline and window creation */
+static DEFINE_SPINLOCK(direct_window_list_lock);
+/* protects initializing window twice for same device */
+static DEFINE_MUTEX(direct_window_init_mutex);
+#define DIRECT64_PROPNAME "linux,direct64-ddr-window-info"
+
+static int tce_clearrange_multi_pSeriesLP(unsigned long start_pfn,
+                                       unsigned long num_pfn, const void *arg)
+{
+       const struct dynamic_dma_window_prop *maprange = arg;
+       int rc;
+       u64 tce_size, num_tce, dma_offset, next;
+       u32 tce_shift;
+       long limit;
+
+       tce_shift = be32_to_cpu(maprange->tce_shift);
+       tce_size = 1ULL << tce_shift;
+       next = start_pfn << PAGE_SHIFT;
+       num_tce = num_pfn << PAGE_SHIFT;
+
+       /* round back to the beginning of the tce page size */
+       num_tce += next & (tce_size - 1);
+       next &= ~(tce_size - 1);
+
+       /* covert to number of tces */
+       num_tce |= tce_size - 1;
+       num_tce >>= tce_shift;
+
+       do {
+               /*
+                * Set up the page with TCE data, looping through and setting
+                * the values.
+                */
+               limit = min_t(long, num_tce, 512);
+               dma_offset = next + be64_to_cpu(maprange->dma_base);
+
+               rc = plpar_tce_stuff((u64)be32_to_cpu(maprange->liobn),
+                                            dma_offset,
+                                            0, limit);
+               num_tce -= limit;
+       } while (num_tce > 0 && !rc);
+
+       return rc;
+}
+
+static int tce_setrange_multi_pSeriesLP(unsigned long start_pfn,
+                                       unsigned long num_pfn, const void *arg)
+{
+       const struct dynamic_dma_window_prop *maprange = arg;
+       u64 *tcep, tce_size, num_tce, dma_offset, next, proto_tce, liobn;
+       u32 tce_shift;
+       u64 rc = 0;
+       long l, limit;
+
+       local_irq_disable();    /* to protect tcep and the page behind it */
+       tcep = __get_cpu_var(tce_page);
+
+       if (!tcep) {
+               tcep = (u64 *)__get_free_page(GFP_ATOMIC);
+               if (!tcep) {
+                       local_irq_enable();
+                       return -ENOMEM;
+               }
+               __get_cpu_var(tce_page) = tcep;
+       }
+
+       proto_tce = TCE_PCI_READ | TCE_PCI_WRITE;
+
+       liobn = (u64)be32_to_cpu(maprange->liobn);
+       tce_shift = be32_to_cpu(maprange->tce_shift);
+       tce_size = 1ULL << tce_shift;
+       next = start_pfn << PAGE_SHIFT;
+       num_tce = num_pfn << PAGE_SHIFT;
+
+       /* round back to the beginning of the tce page size */
+       num_tce += next & (tce_size - 1);
+       next &= ~(tce_size - 1);
+
+       /* covert to number of tces */
+       num_tce |= tce_size - 1;
+       num_tce >>= tce_shift;
+
+       /* We can map max one pageful of TCEs at a time */
+       do {
+               /*
+                * Set up the page with TCE data, looping through and setting
+                * the values.
+                */
+               limit = min_t(long, num_tce, 4096/TCE_ENTRY_SIZE);
+               dma_offset = next + be64_to_cpu(maprange->dma_base);
+
+               for (l = 0; l < limit; l++) {
+                       tcep[l] = proto_tce | next;
+                       next += tce_size;
+               }
+
+               rc = plpar_tce_put_indirect(liobn,
+                                           dma_offset,
+                                           (u64)virt_to_abs(tcep),
+                                           limit);
+
+               num_tce -= limit;
+       } while (num_tce > 0 && !rc);
+
+       /* error cleanup: caller will clear whole range */
+
+       local_irq_enable();
+       return rc;
+}
+
+static int tce_setrange_multi_pSeriesLP_walk(unsigned long start_pfn,
+               unsigned long num_pfn, void *arg)
+{
+       return tce_setrange_multi_pSeriesLP(start_pfn, num_pfn, arg);
+}
+
+
 #ifdef CONFIG_PCI
 static void iommu_table_setparms(struct pci_controller *phb,
                                 struct device_node *dn,
@@ -495,6 +643,329 @@ static void pci_dma_dev_setup_pSeries(struct pci_dev *dev)
                       pci_name(dev));
 }
 
+static int __read_mostly disable_ddw;
+
+static int __init disable_ddw_setup(char *str)
+{
+       disable_ddw = 1;
+       printk(KERN_INFO "ppc iommu: disabling ddw.\n");
+
+       return 0;
+}
+
+early_param("disable_ddw", disable_ddw_setup);
+
+static void remove_ddw(struct device_node *np)
+{
+       struct dynamic_dma_window_prop *dwp;
+       struct property *win64;
+       const u32 *ddr_avail;
+       u64 liobn;
+       int len, ret;
+
+       ddr_avail = of_get_property(np, "ibm,ddw-applicable", &len);
+       win64 = of_find_property(np, DIRECT64_PROPNAME, NULL);
+       if (!win64 || !ddr_avail || len < 3 * sizeof(u32))
+               return;
+
+       dwp = win64->value;
+       liobn = (u64)be32_to_cpu(dwp->liobn);
+
+       /* clear the whole window, note the arg is in kernel pages */
+       ret = tce_clearrange_multi_pSeriesLP(0,
+               1ULL << (be32_to_cpu(dwp->window_shift) - PAGE_SHIFT), dwp);
+       if (ret)
+               pr_warning("%s failed to clear tces in window.\n",
+                        np->full_name);
+       else
+               pr_debug("%s successfully cleared tces in window.\n",
+                        np->full_name);
+
+       ret = rtas_call(ddr_avail[2], 1, 1, NULL, liobn);
+       if (ret)
+               pr_warning("%s: failed to remove direct window: rtas returned "
+                       "%d to ibm,remove-pe-dma-window(%x) %llx\n",
+                       np->full_name, ret, ddr_avail[2], liobn);
+       else
+               pr_debug("%s: successfully removed direct window: rtas returned "
+                       "%d to ibm,remove-pe-dma-window(%x) %llx\n",
+                       np->full_name, ret, ddr_avail[2], liobn);
+}
+
+
+static int dupe_ddw_if_already_created(struct pci_dev *dev, struct device_node *pdn)
+{
+       struct device_node *dn;
+       struct pci_dn *pcidn;
+       struct direct_window *window;
+       const struct dynamic_dma_window_prop *direct64;
+       u64 dma_addr = 0;
+
+       dn = pci_device_to_OF_node(dev);
+       pcidn = PCI_DN(dn);
+       spin_lock(&direct_window_list_lock);
+       /* check if we already created a window and dupe that config if so */
+       list_for_each_entry(window, &direct_window_list, list) {
+               if (window->device == pdn) {
+                       direct64 = window->prop;
+                       dma_addr = direct64->dma_base;
+                       break;
+               }
+       }
+       spin_unlock(&direct_window_list_lock);
+
+       return dma_addr;
+}
+
+static u64 dupe_ddw_if_kexec(struct pci_dev *dev, struct device_node *pdn)
+{
+       struct device_node *dn;
+       struct pci_dn *pcidn;
+       int len;
+       struct direct_window *window;
+       const struct dynamic_dma_window_prop *direct64;
+       u64 dma_addr = 0;
+
+       dn = pci_device_to_OF_node(dev);
+       pcidn = PCI_DN(dn);
+       direct64 = of_get_property(pdn, DIRECT64_PROPNAME, &len);
+       if (direct64) {
+               window = kzalloc(sizeof(*window), GFP_KERNEL);
+               if (!window) {
+                       remove_ddw(pdn);
+               } else {
+                       window->device = pdn;
+                       window->prop = direct64;
+                       spin_lock(&direct_window_list_lock);
+                       list_add(&window->list, &direct_window_list);
+                       spin_unlock(&direct_window_list_lock);
+                       dma_addr = direct64->dma_base;
+               }
+       }
+
+       return dma_addr;
+}
+
+static int query_ddw(struct pci_dev *dev, const u32 *ddr_avail,
+                       struct ddw_query_response *query)
+{
+       struct device_node *dn;
+       struct pci_dn *pcidn;
+       u32 cfg_addr;
+       u64 buid;
+       int ret;
+
+       /*
+        * Get the config address and phb buid of the PE window.
+        * Rely on eeh to retrieve this for us.
+        * Retrieve them from the pci device, not the node with the
+        * dma-window property
+        */
+       dn = pci_device_to_OF_node(dev);
+       pcidn = PCI_DN(dn);
+       cfg_addr = pcidn->eeh_config_addr;
+       if (pcidn->eeh_pe_config_addr)
+               cfg_addr = pcidn->eeh_pe_config_addr;
+       buid = pcidn->phb->buid;
+       ret = rtas_call(ddr_avail[0], 3, 5, (u32 *)query,
+                 cfg_addr, BUID_HI(buid), BUID_LO(buid));
+       dev_info(&dev->dev, "ibm,query-pe-dma-windows(%x) %x %x %x"
+               " returned %d\n", ddr_avail[0], cfg_addr, BUID_HI(buid),
+               BUID_LO(buid), ret);
+       return ret;
+}
+
+static int create_ddw(struct pci_dev *dev, const u32 *ddr_avail,
+                       struct ddw_create_response *create, int page_shift,
+                       int window_shift)
+{
+       struct device_node *dn;
+       struct pci_dn *pcidn;
+       u32 cfg_addr;
+       u64 buid;
+       int ret;
+
+       /*
+        * Get the config address and phb buid of the PE window.
+        * Rely on eeh to retrieve this for us.
+        * Retrieve them from the pci device, not the node with the
+        * dma-window property
+        */
+       dn = pci_device_to_OF_node(dev);
+       pcidn = PCI_DN(dn);
+       cfg_addr = pcidn->eeh_config_addr;
+       if (pcidn->eeh_pe_config_addr)
+               cfg_addr = pcidn->eeh_pe_config_addr;
+       buid = pcidn->phb->buid;
+
+       do {
+               /* extra outputs are LIOBN and dma-addr (hi, lo) */
+               ret = rtas_call(ddr_avail[1], 5, 4, (u32 *)create, cfg_addr,
+                               BUID_HI(buid), BUID_LO(buid), page_shift, window_shift);
+       } while (rtas_busy_delay(ret));
+       dev_info(&dev->dev,
+               "ibm,create-pe-dma-window(%x) %x %x %x %x %x returned %d "
+               "(liobn = 0x%x starting addr = %x %x)\n", ddr_avail[1],
+                cfg_addr, BUID_HI(buid), BUID_LO(buid), page_shift,
+                window_shift, ret, create->liobn, create->addr_hi, create->addr_lo);
+
+       return ret;
+}
+
+/*
+ * If the PE supports dynamic dma windows, and there is space for a table
+ * that can map all pages in a linear offset, then setup such a table,
+ * and record the dma-offset in the struct device.
+ *
+ * dev: the pci device we are checking
+ * pdn: the parent pe node with the ibm,dma_window property
+ * Future: also check if we can remap the base window for our base page size
+ *
+ * returns the dma offset for use by dma_set_mask
+ */
+static u64 enable_ddw(struct pci_dev *dev, struct device_node *pdn)
+{
+       int len, ret;
+       struct ddw_query_response query;
+       struct ddw_create_response create;
+       int page_shift;
+       u64 dma_addr, max_addr;
+       struct device_node *dn;
+       const u32 *uninitialized_var(ddr_avail);
+       struct direct_window *window;
+       struct property *uninitialized_var(win64);
+       struct dynamic_dma_window_prop *ddwprop;
+
+       mutex_lock(&direct_window_init_mutex);
+
+       dma_addr = dupe_ddw_if_already_created(dev, pdn);
+       if (dma_addr != 0)
+               goto out_unlock;
+
+       dma_addr = dupe_ddw_if_kexec(dev, pdn);
+       if (dma_addr != 0)
+               goto out_unlock;
+
+       /*
+        * the ibm,ddw-applicable property holds the tokens for:
+        * ibm,query-pe-dma-window
+        * ibm,create-pe-dma-window
+        * ibm,remove-pe-dma-window
+        * for the given node in that order.
+        * the property is actually in the parent, not the PE
+        */
+       ddr_avail = of_get_property(pdn, "ibm,ddw-applicable", &len);
+       if (!ddr_avail || len < 3 * sizeof(u32))
+               goto out_unlock;
+
+       /*
+        * Query if there is a second window of size to map the
+        * whole partition.  Query returns number of windows, largest
+        * block assigned to PE (partition endpoint), and two bitmasks
+        * of page sizes: supported and supported for migrate-dma.
+        */
+       dn = pci_device_to_OF_node(dev);
+       ret = query_ddw(dev, ddr_avail, &query);
+       if (ret != 0)
+               goto out_unlock;
+
+       if (query.windows_available == 0) {
+               /*
+                * no additional windows are available for this device.
+                * We might be able to reallocate the existing window,
+                * trading in for a larger page size.
+                */
+               dev_dbg(&dev->dev, "no free dynamic windows");
+               goto out_unlock;
+       }
+       if (query.page_size & 4) {
+               page_shift = 24; /* 16MB */
+       } else if (query.page_size & 2) {
+               page_shift = 16; /* 64kB */
+       } else if (query.page_size & 1) {
+               page_shift = 12; /* 4kB */
+       } else {
+               dev_dbg(&dev->dev, "no supported direct page size in mask %x",
+                         query.page_size);
+               goto out_unlock;
+       }
+       /* verify the window * number of ptes will map the partition */
+       /* check largest block * page size > max memory hotplug addr */
+       max_addr = memory_hotplug_max();
+       if (query.largest_available_block < (max_addr >> page_shift)) {
+               dev_dbg(&dev->dev, "can't map partiton max 0x%llx with %u "
+                         "%llu-sized pages\n", max_addr,  query.largest_available_block,
+                         1ULL << page_shift);
+               goto out_unlock;
+       }
+       len = order_base_2(max_addr);
+       win64 = kzalloc(sizeof(struct property), GFP_KERNEL);
+       if (!win64) {
+               dev_info(&dev->dev,
+                       "couldn't allocate property for 64bit dma window\n");
+               goto out_unlock;
+       }
+       win64->name = kstrdup(DIRECT64_PROPNAME, GFP_KERNEL);
+       win64->value = ddwprop = kmalloc(sizeof(*ddwprop), GFP_KERNEL);
+       if (!win64->name || !win64->value) {
+               dev_info(&dev->dev,
+                       "couldn't allocate property name and value\n");
+               goto out_free_prop;
+       }
+
+       ret = create_ddw(dev, ddr_avail, &create, page_shift, len);
+       if (ret != 0)
+               goto out_free_prop;
+
+       ddwprop->liobn = cpu_to_be32(create.liobn);
+       ddwprop->dma_base = cpu_to_be64(of_read_number(&create.addr_hi, 2));
+       ddwprop->tce_shift = cpu_to_be32(page_shift);
+       ddwprop->window_shift = cpu_to_be32(len);
+
+       dev_dbg(&dev->dev, "created tce table LIOBN 0x%x for %s\n",
+                 create.liobn, dn->full_name);
+
+       window = kzalloc(sizeof(*window), GFP_KERNEL);
+       if (!window)
+               goto out_clear_window;
+
+       ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
+                       win64->value, tce_setrange_multi_pSeriesLP_walk);
+       if (ret) {
+               dev_info(&dev->dev, "failed to map direct window for %s: %d\n",
+                        dn->full_name, ret);
+               goto out_clear_window;
+       }
+
+       ret = prom_add_property(pdn, win64);
+       if (ret) {
+               dev_err(&dev->dev, "unable to add dma window property for %s: %d",
+                        pdn->full_name, ret);
+               goto out_clear_window;
+       }
+
+       window->device = pdn;
+       window->prop = ddwprop;
+       spin_lock(&direct_window_list_lock);
+       list_add(&window->list, &direct_window_list);
+       spin_unlock(&direct_window_list_lock);
+
+       dma_addr = of_read_number(&create.addr_hi, 2);
+       goto out_unlock;
+
+out_clear_window:
+       remove_ddw(pdn);
+
+out_free_prop:
+       kfree(win64->name);
+       kfree(win64->value);
+       kfree(win64);
+
+out_unlock:
+       mutex_unlock(&direct_window_init_mutex);
+       return dma_addr;
+}
+
 static void pci_dma_dev_setup_pSeriesLP(struct pci_dev *dev)
 {
        struct device_node *pdn, *dn;
@@ -541,23 +1012,137 @@ static void pci_dma_dev_setup_pSeriesLP(struct pci_dev *dev)
 
        set_iommu_table_base(&dev->dev, pci->iommu_table);
 }
+
+static int dma_set_mask_pSeriesLP(struct device *dev, u64 dma_mask)
+{
+       bool ddw_enabled = false;
+       struct device_node *pdn, *dn;
+       struct pci_dev *pdev;
+       const void *dma_window = NULL;
+       u64 dma_offset;
+
+       if (!dev->dma_mask || !dma_supported(dev, dma_mask))
+               return -EIO;
+
+       /* only attempt to use a new window if 64-bit DMA is requested */
+       if (!disable_ddw && dma_mask == DMA_BIT_MASK(64)) {
+               pdev = to_pci_dev(dev);
+
+               dn = pci_device_to_OF_node(pdev);
+               dev_dbg(dev, "node is %s\n", dn->full_name);
+
+               /*
+                * the device tree might contain the dma-window properties
+                * per-device and not neccesarily for the bus. So we need to
+                * search upwards in the tree until we either hit a dma-window
+                * property, OR find a parent with a table already allocated.
+                */
+               for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->iommu_table;
+                               pdn = pdn->parent) {
+                       dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
+                       if (dma_window)
+                               break;
+               }
+               if (pdn && PCI_DN(pdn)) {
+                       dma_offset = enable_ddw(pdev, pdn);
+                       if (dma_offset != 0) {
+                               dev_info(dev, "Using 64-bit direct DMA at offset %llx\n", dma_offset);
+                               set_dma_offset(dev, dma_offset);
+                               set_dma_ops(dev, &dma_direct_ops);
+                               ddw_enabled = true;
+                       }
+               }
+       }
+
+       /* fall-through to iommu ops */
+       if (!ddw_enabled) {
+               dev_info(dev, "Using 32-bit DMA via iommu\n");
+               set_dma_ops(dev, &dma_iommu_ops);
+       }
+
+       *dev->dma_mask = dma_mask;
+       return 0;
+}
+
 #else  /* CONFIG_PCI */
 #define pci_dma_bus_setup_pSeries      NULL
 #define pci_dma_dev_setup_pSeries      NULL
 #define pci_dma_bus_setup_pSeriesLP    NULL
 #define pci_dma_dev_setup_pSeriesLP    NULL
+#define dma_set_mask_pSeriesLP         NULL
 #endif /* !CONFIG_PCI */
 
+static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
+               void *data)
+{
+       struct direct_window *window;
+       struct memory_notify *arg = data;
+       int ret = 0;
+
+       switch (action) {
+       case MEM_GOING_ONLINE:
+               spin_lock(&direct_window_list_lock);
+               list_for_each_entry(window, &direct_window_list, list) {
+                       ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
+                                       arg->nr_pages, window->prop);
+                       /* XXX log error */
+               }
+               spin_unlock(&direct_window_list_lock);
+               break;
+       case MEM_CANCEL_ONLINE:
+       case MEM_OFFLINE:
+               spin_lock(&direct_window_list_lock);
+               list_for_each_entry(window, &direct_window_list, list) {
+                       ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
+                                       arg->nr_pages, window->prop);
+                       /* XXX log error */
+               }
+               spin_unlock(&direct_window_list_lock);
+               break;
+       default:
+               break;
+       }
+       if (ret && action != MEM_CANCEL_ONLINE)
+               return NOTIFY_BAD;
+
+       return NOTIFY_OK;
+}
+
+static struct notifier_block iommu_mem_nb = {
+       .notifier_call = iommu_mem_notifier,
+};
+
 static int iommu_reconfig_notifier(struct notifier_block *nb, unsigned long action, void *node)
 {
        int err = NOTIFY_OK;
        struct device_node *np = node;
        struct pci_dn *pci = PCI_DN(np);
+       struct direct_window *window;
 
        switch (action) {
        case PSERIES_RECONFIG_REMOVE:
                if (pci && pci->iommu_table)
                        iommu_free_table(pci->iommu_table, np->full_name);
+
+               spin_lock(&direct_window_list_lock);
+               list_for_each_entry(window, &direct_window_list, list) {
+                       if (window->device == np) {
+                               list_del(&window->list);
+                               kfree(window);
+                               break;
+                       }
+               }
+               spin_unlock(&direct_window_list_lock);
+
+               /*
+                * Because the notifier runs after isolation of the
+                * slot, we are guaranteed any DMA window has already
+                * been revoked and the TCEs have been marked invalid,
+                * so we don't need a call to remove_ddw(np). However,
+                * if an additional notifier action is added before the
+                * isolate call, we should update this code for
+                * completeness with such a call.
+                */
                break;
        default:
                err = NOTIFY_DONE;
@@ -587,6 +1172,7 @@ void iommu_init_early_pSeries(void)
                ppc_md.tce_get   = tce_get_pSeriesLP;
                ppc_md.pci_dma_bus_setup = pci_dma_bus_setup_pSeriesLP;
                ppc_md.pci_dma_dev_setup = pci_dma_dev_setup_pSeriesLP;
+               ppc_md.dma_set_mask = dma_set_mask_pSeriesLP;
        } else {
                ppc_md.tce_build = tce_build_pSeries;
                ppc_md.tce_free  = tce_free_pSeries;
@@ -597,6 +1183,7 @@ void iommu_init_early_pSeries(void)
 
 
        pSeries_reconfig_notifier_register(&iommu_reconfig_nb);
+       register_memory_notifier(&iommu_mem_nb);
 
        set_pci_dma_ops(&dma_iommu_ops);
 }