crypto: qat - Enforce valid numa configuration
authorTadeusz Struk <tadeusz.struk@intel.com>
Tue, 14 Oct 2014 01:24:32 +0000 (18:24 -0700)
committerHerbert Xu <herbert@gondor.apana.org.au>
Fri, 24 Oct 2014 14:37:36 +0000 (22:37 +0800)
In a system with NUMA configuration we want to enforce that the accelerator is
connected to a node with memory to avoid cross QPI memory transaction.
Otherwise there is no point in using the accelerator as the encryption in
software will be faster.

Cc: stable@vger.kernel.org
Signed-off-by: Tadeusz Struk <tadeusz.struk@intel.com>
Tested-by: Nikolay Aleksandrov <nikolay@redhat.com>
Reviewed-by: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
drivers/crypto/qat/qat_common/adf_accel_devices.h
drivers/crypto/qat/qat_common/adf_transport.c
drivers/crypto/qat/qat_common/qat_algs.c
drivers/crypto/qat/qat_common/qat_crypto.c
drivers/crypto/qat/qat_dh895xcc/adf_admin.c
drivers/crypto/qat/qat_dh895xcc/adf_drv.c
drivers/crypto/qat/qat_dh895xcc/adf_isr.c

index 9282381..fe7b3f0 100644 (file)
@@ -198,8 +198,7 @@ struct adf_accel_dev {
        struct dentry *debugfs_dir;
        struct list_head list;
        struct module *owner;
-       uint8_t accel_id;
-       uint8_t numa_node;
        struct adf_accel_pci accel_pci_dev;
+       uint8_t accel_id;
 } __packed;
 #endif
index 5f3fa45..9dd2cb7 100644 (file)
@@ -419,9 +419,10 @@ static int adf_init_bank(struct adf_accel_dev *accel_dev,
                WRITE_CSR_RING_BASE(csr_addr, bank_num, i, 0);
                ring = &bank->rings[i];
                if (hw_data->tx_rings_mask & (1 << i)) {
-                       ring->inflights = kzalloc_node(sizeof(atomic_t),
-                                                      GFP_KERNEL,
-                                                      accel_dev->numa_node);
+                       ring->inflights =
+                               kzalloc_node(sizeof(atomic_t),
+                                            GFP_KERNEL,
+                                            dev_to_node(&GET_DEV(accel_dev)));
                        if (!ring->inflights)
                                goto err;
                } else {
@@ -469,13 +470,14 @@ int adf_init_etr_data(struct adf_accel_dev *accel_dev)
        int i, ret;
 
        etr_data = kzalloc_node(sizeof(*etr_data), GFP_KERNEL,
-                               accel_dev->numa_node);
+                               dev_to_node(&GET_DEV(accel_dev)));
        if (!etr_data)
                return -ENOMEM;
 
        num_banks = GET_MAX_BANKS(accel_dev);
        size = num_banks * sizeof(struct adf_etr_bank_data);
-       etr_data->banks = kzalloc_node(size, GFP_KERNEL, accel_dev->numa_node);
+       etr_data->banks = kzalloc_node(size, GFP_KERNEL,
+                                      dev_to_node(&GET_DEV(accel_dev)));
        if (!etr_data->banks) {
                ret = -ENOMEM;
                goto err_bank;
index 699ccf4..9e9619c 100644 (file)
@@ -596,7 +596,8 @@ static int qat_alg_sgl_to_bufl(struct qat_crypto_instance *inst,
        if (unlikely(!n))
                return -EINVAL;
 
-       bufl = kmalloc_node(sz, GFP_ATOMIC, inst->accel_dev->numa_node);
+       bufl = kmalloc_node(sz, GFP_ATOMIC,
+                           dev_to_node(&GET_DEV(inst->accel_dev)));
        if (unlikely(!bufl))
                return -ENOMEM;
 
@@ -642,7 +643,7 @@ static int qat_alg_sgl_to_bufl(struct qat_crypto_instance *inst,
                struct qat_alg_buf *bufers;
 
                buflout = kmalloc_node(sz, GFP_ATOMIC,
-                                      inst->accel_dev->numa_node);
+                                      dev_to_node(&GET_DEV(inst->accel_dev)));
                if (unlikely(!buflout))
                        goto err;
                bloutp = dma_map_single(dev, buflout, sz, DMA_TO_DEVICE);
index 0d59bcb..828f2a6 100644 (file)
@@ -109,12 +109,14 @@ struct qat_crypto_instance *qat_crypto_get_instance_node(int node)
 
        list_for_each(itr, adf_devmgr_get_head()) {
                accel_dev = list_entry(itr, struct adf_accel_dev, list);
-               if (accel_dev->numa_node == node && adf_dev_started(accel_dev))
+               if ((node == dev_to_node(&GET_DEV(accel_dev)) ||
+                       dev_to_node(&GET_DEV(accel_dev)) < 0)
+                               && adf_dev_started(accel_dev))
                        break;
                accel_dev = NULL;
        }
        if (!accel_dev) {
-               pr_err("QAT: Could not find device on give node\n");
+               pr_err("QAT: Could not find device on node %d\n", node);
                accel_dev = adf_devmgr_get_first();
        }
        if (!accel_dev || !adf_dev_started(accel_dev))
@@ -164,7 +166,7 @@ static int qat_crypto_create_instances(struct adf_accel_dev *accel_dev)
 
        for (i = 0; i < num_inst; i++) {
                inst = kzalloc_node(sizeof(*inst), GFP_KERNEL,
-                                   accel_dev->numa_node);
+                                   dev_to_node(&GET_DEV(accel_dev)));
                if (!inst)
                        goto err;
 
index 978d6c5..53c491b 100644 (file)
@@ -108,7 +108,7 @@ int adf_init_admin_comms(struct adf_accel_dev *accel_dev)
        uint64_t reg_val;
 
        admin = kzalloc_node(sizeof(*accel_dev->admin), GFP_KERNEL,
-                            accel_dev->numa_node);
+                            dev_to_node(&GET_DEV(accel_dev)));
        if (!admin)
                return -ENOMEM;
        admin->virt_addr = dma_zalloc_coherent(&GET_DEV(accel_dev), PAGE_SIZE,
index 0d0435a..948f66b 100644 (file)
@@ -119,21 +119,6 @@ static void adf_cleanup_accel(struct adf_accel_dev *accel_dev)
        kfree(accel_dev);
 }
 
-static uint8_t adf_get_dev_node_id(struct pci_dev *pdev)
-{
-       unsigned int bus_per_cpu = 0;
-       struct cpuinfo_x86 *c = &cpu_data(num_online_cpus() - 1);
-
-       if (!c->phys_proc_id)
-               return 0;
-
-       bus_per_cpu = 256 / (c->phys_proc_id + 1);
-
-       if (bus_per_cpu != 0)
-               return pdev->bus->number / bus_per_cpu;
-       return 0;
-}
-
 static int qat_dev_start(struct adf_accel_dev *accel_dev)
 {
        int cpus = num_online_cpus();
@@ -235,7 +220,6 @@ static int adf_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
        void __iomem *pmisc_bar_addr = NULL;
        char name[ADF_DEVICE_NAME_LENGTH];
        unsigned int i, bar_nr;
-       uint8_t node;
        int ret;
 
        switch (ent->device) {
@@ -246,12 +230,19 @@ static int adf_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
                return -ENODEV;
        }
 
-       node = adf_get_dev_node_id(pdev);
-       accel_dev = kzalloc_node(sizeof(*accel_dev), GFP_KERNEL, node);
+       if (num_possible_nodes() > 1 && dev_to_node(&pdev->dev) < 0) {
+               /* If the accelerator is connected to a node with no memory
+                * there is no point in using the accelerator since the remote
+                * memory transaction will be very slow. */
+               dev_err(&pdev->dev, "Invalid NUMA configuration.\n");
+               return -EINVAL;
+       }
+
+       accel_dev = kzalloc_node(sizeof(*accel_dev), GFP_KERNEL,
+                                dev_to_node(&pdev->dev));
        if (!accel_dev)
                return -ENOMEM;
 
-       accel_dev->numa_node = node;
        INIT_LIST_HEAD(&accel_dev->crypto_list);
 
        /* Add accel device to accel table.
@@ -264,7 +255,8 @@ static int adf_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 
        accel_dev->owner = THIS_MODULE;
        /* Allocate and configure device configuration structure */
-       hw_data = kzalloc_node(sizeof(*hw_data), GFP_KERNEL, node);
+       hw_data = kzalloc_node(sizeof(*hw_data), GFP_KERNEL,
+                              dev_to_node(&pdev->dev));
        if (!hw_data) {
                ret = -ENOMEM;
                goto out_err;
index 67ec61e..d96ee21 100644 (file)
@@ -168,7 +168,7 @@ static int adf_isr_alloc_msix_entry_table(struct adf_accel_dev *accel_dev)
        uint32_t msix_num_entries = hw_data->num_banks + 1;
 
        entries = kzalloc_node(msix_num_entries * sizeof(*entries),
-                              GFP_KERNEL, accel_dev->numa_node);
+                              GFP_KERNEL, dev_to_node(&GET_DEV(accel_dev)));
        if (!entries)
                return -ENOMEM;