Re: [PATCH v6 09/18] nitro_enclaves: Add logic for setting an enclave vCPU

From: Paraschiv, Andra-Irina
Date: Tue Aug 11 2020 - 04:53:17 EST




On 10/08/2020 10:33, Alexander Graf wrote:


On 05.08.20 11:10, Andra Paraschiv wrote:
An enclave, before being started, has its resources set. One of its
resources is CPU.

A NE CPU pool is set and enclave CPUs are chosen from it. Offline the
CPUs from the NE CPU pool during the pool setup and online them back
during the NE CPU pool teardown. The CPU offline is necessary so that
there would not be more vCPUs than physical CPUs available to the
primary / parent VM. In that case the CPUs would be overcommitted and
would change the initial configuration of the primary / parent VM of
having dedicated vCPUs to physical CPUs.

The enclave CPUs need to be full cores and from the same NUMA node. CPU
0 and its siblings have to remain available to the primary / parent VM.

Add ioctl command logic for setting an enclave vCPU.

Signed-off-by: Alexandru Vasile <lexnv@xxxxxxxxxx>
Signed-off-by: Andra Paraschiv <andraprs@xxxxxxxxxx>
---
Changelog

v5 -> v6

* Check CPUs are from the same NUMA node before going through CPU
   siblings during the NE CPU pool setup.
* Update documentation to kernel-doc format.

v4 -> v5

* Set empty string in case of invalid NE CPU pool.
* Clear NE CPU pool mask on pool setup failure.
* Setup NE CPU cores out of the NE CPU pool.
* Early exit on NE CPU pool setup if enclave(s) already running.
* Remove sanity checks for situations that shouldn't happen, only if
   buggy system or broken logic at all.
* Add check for maximum vCPU id possible before looking into the CPU
   pool.
* Remove log on copy_from_user() / copy_to_user() failure and on admin
   capability check for setting the NE CPU pool.
* Update the ioctl call to not create a file descriptor for the vCPU.
* Split the CPU pool usage logic in 2 separate functions - one to get a
   CPU from the pool and the other to check the given CPU is available in
   the pool.

v3 -> v4

* Setup the NE CPU pool at runtime via a sysfs file for the kernel
   parameter.
* Check enclave CPUs to be from the same NUMA node.
* Use dev_err instead of custom NE log pattern.
* Update the NE ioctl call to match the decoupling from the KVM API.

v2 -> v3

* Remove the WARN_ON calls.
* Update static calls sanity checks.
* Update kzfree() calls to kfree().
* Remove file ops that do nothing for now - open, ioctl and release.

v1 -> v2

* Add log pattern for NE.
* Update goto labels to match their purpose.
* Remove the BUG_ON calls.
* Check if enclave state is init when setting enclave vCPU.
---
  drivers/virt/nitro_enclaves/ne_misc_dev.c | 575 ++++++++++++++++++++++
  1 file changed, 575 insertions(+)

diff --git a/drivers/virt/nitro_enclaves/ne_misc_dev.c b/drivers/virt/nitro_enclaves/ne_misc_dev.c
index 6c8c12f65666..4787bc59d39d 100644
--- a/drivers/virt/nitro_enclaves/ne_misc_dev.c
+++ b/drivers/virt/nitro_enclaves/ne_misc_dev.c
@@ -57,8 +57,11 @@
   * TODO: Update logic to create new sysfs entries instead of using
   * a kernel parameter e.g. if multiple sysfs files needed.
   */
+static int ne_set_kernel_param(const char *val, const struct kernel_param *kp);
+
  static const struct kernel_param_ops ne_cpu_pool_ops = {
      .get    = param_get_string,
+    .set    = ne_set_kernel_param,
  };
    static char ne_cpus[NE_CPUS_SIZE];
@@ -87,6 +90,575 @@ struct ne_cpu_pool {
    static struct ne_cpu_pool ne_cpu_pool;
  +/**
+ * ne_check_enclaves_created() - Verify if at least one enclave has been created.
+ * @void:    No parameters provided.
+ *
+ * Context: Process context.
+ * Return:
+ * * True if at least one enclave is created.
+ * * False otherwise.
+ */
+static bool ne_check_enclaves_created(void)
+{
+    struct ne_pci_dev *ne_pci_dev = NULL;
+    /* TODO: Find another way to get the NE PCI device reference. */
+    struct pci_dev *pdev = pci_get_device(PCI_VENDOR_ID_AMAZON, PCI_DEVICE_ID_NE, NULL);

Can we just make this a global ref count instead?

I think we can use here as well the misc dev parent approach mentioned in the previous patches in this series. If no parent dev set, early exit.


+    bool ret = false;
+
+    if (!pdev)
+        return ret;
+
+    ne_pci_dev = pci_get_drvdata(pdev);
+    if (!ne_pci_dev) {
+        pci_dev_put(pdev);
+
+        return ret;
+    }
+
+    mutex_lock(&ne_pci_dev->enclaves_list_mutex);
+
+    if (!list_empty(&ne_pci_dev->enclaves_list))
+        ret = true;
+
+    mutex_unlock(&ne_pci_dev->enclaves_list_mutex);
+
+    pci_dev_put(pdev);
+
+    return ret;
+}
+
+/**
+ * ne_setup_cpu_pool() - Set the NE CPU pool after handling sanity checks such
+ *             as not sharing CPU cores with the primary / parent VM
+ *             or not using CPU 0, which should remain available for
+ *             the primary / parent VM. Offline the CPUs from the
+ *             pool after the checks passed.
+ * @ne_cpu_list:    The CPU list used for setting NE CPU pool.
+ *
+ * Context: Process context.
+ * Return:
+ * * 0 on success.
+ * * Negative return value on failure.
+ */
+static int ne_setup_cpu_pool(const char *ne_cpu_list)
+{
+    int core_id = -1;
+    unsigned int cpu = 0;
+    cpumask_var_t cpu_pool = NULL;
+    unsigned int cpu_sibling = 0;
+    unsigned int i = 0;
+    int numa_node = -1;
+    int rc = -EINVAL;
+
+    if (!ne_cpu_list)
+        return 0;
+
+    if (!zalloc_cpumask_var(&cpu_pool, GFP_KERNEL))
+        return -ENOMEM;
+
+    mutex_lock(&ne_cpu_pool.mutex);
+
+    rc = cpulist_parse(ne_cpu_list, cpu_pool);
+    if (rc < 0) {
+        pr_err("%s: Error in cpulist parse [rc=%d]\n", ne_misc_dev.name, rc);
+
+        goto free_pool_cpumask;
+    }
+
+    cpu = cpumask_any(cpu_pool);
+    if (cpu >= nr_cpu_ids) {
+        pr_err("%s: No CPUs available in CPU pool\n", ne_misc_dev.name);
+
+        rc = -EINVAL;
+
+        goto free_pool_cpumask;
+    }
+
+    /*
+     * Check if the CPUs from the NE CPU pool are from the same NUMA node.
+     */
+    for_each_cpu(cpu, cpu_pool) {
+        if (numa_node < 0) {
+            numa_node = cpu_to_node(cpu);
+            if (numa_node < 0) {
+                pr_err("%s: Invalid NUMA node %d\n",
+                       ne_misc_dev.name, numa_node);
+
+                rc = -EINVAL;
+
+                goto free_pool_cpumask;
+            }
+        } else {
+            if (numa_node != cpu_to_node(cpu)) {
+                pr_err("%s: CPUs with different NUMA nodes\n",
+                       ne_misc_dev.name);
+
+                rc = -EINVAL;
+
+                goto free_pool_cpumask;
+            }
+        }
+    }
+
+    /*
+     * Check if CPU 0 and its siblings are included in the provided CPU pool
+     * They should remain available for the primary / parent VM.
+     */
+    if (cpumask_test_cpu(0, cpu_pool)) {
+        pr_err("%s: CPU 0 has to remain available\n", ne_misc_dev.name);
+
+        rc = -EINVAL;
+
+        goto free_pool_cpumask;
+    }
+
+    for_each_cpu(cpu_sibling, topology_sibling_cpumask(0)) {
+        if (cpumask_test_cpu(cpu_sibling, cpu_pool)) {
+            pr_err("%s: CPU sibling %d for CPU 0 is in CPU pool\n",
+                   ne_misc_dev.name, cpu_sibling);
+
+            rc = -EINVAL;
+
+            goto free_pool_cpumask;
+        }
+    }
+
+    /*
+     * Check if CPU siblings are included in the provided CPU pool. The
+     * expectation is that CPU cores are made available in the CPU pool for
+     * enclaves.
+     */
+    for_each_cpu(cpu, cpu_pool) {
+        for_each_cpu(cpu_sibling, topology_sibling_cpumask(cpu)) {
+            if (!cpumask_test_cpu(cpu_sibling, cpu_pool)) {
+                pr_err("%s: CPU %d is not in CPU pool\n",
+                       ne_misc_dev.name, cpu_sibling);
+
+                rc = -EINVAL;
+
+                goto free_pool_cpumask;
+            }
+        }
+    }
+
+    ne_cpu_pool.avail_cores_size = nr_cpu_ids / smp_num_siblings;
+
+    ne_cpu_pool.avail_cores = kcalloc(ne_cpu_pool.avail_cores_size,
+                      sizeof(*ne_cpu_pool.avail_cores),
+                      GFP_KERNEL);
+    if (!ne_cpu_pool.avail_cores) {
+        rc = -ENOMEM;
+
+        goto free_pool_cpumask;
+    }
+
+    for (i = 0; i < ne_cpu_pool.avail_cores_size; i++)
+        if (!zalloc_cpumask_var(&ne_cpu_pool.avail_cores[i], GFP_KERNEL)) {
+            rc = -ENOMEM;
+
+            goto free_cores_cpumask;
+        }
+
+    /* Split the NE CPU pool in CPU cores. */
+    for_each_cpu(cpu, cpu_pool) {
+        core_id = topology_core_id(cpu);
+        if (core_id < 0 || core_id >= ne_cpu_pool.avail_cores_size) {
+            pr_err("%s: Invalid core id  %d\n",
+                   ne_misc_dev.name, core_id);
+
+            rc = -EINVAL;
+
+            goto clear_cpumask;
+        }
+
+        cpumask_set_cpu(cpu, ne_cpu_pool.avail_cores[core_id]);
+    }
+
+    /*
+     * CPUs that are given to enclave(s) should not be considered online
+     * by Linux anymore, as the hypervisor will degrade them to floating.
+     * The physical CPUs (full cores) are carved out of the primary / parent
+     * VM and given to the enclave VM. The same number of vCPUs would run
+     * on less pCPUs for the primary / parent VM.
+     *
+     * We offline them here, to not degrade performance and expose correct
+     * topology to Linux and user space.
+     */
+    for_each_cpu(cpu, cpu_pool) {
+        rc = remove_cpu(cpu);
+        if (rc != 0) {
+            pr_err("%s: CPU %d is not offlined [rc=%d]\n",
+                   ne_misc_dev.name, cpu, rc);
+
+            goto online_cpus;
+        }
+    }
+
+    free_cpumask_var(cpu_pool);
+
+    ne_cpu_pool.numa_node = numa_node;
+
+    mutex_unlock(&ne_cpu_pool.mutex);
+
+    return 0;
+
+online_cpus:
+    for_each_cpu(cpu, cpu_pool)
+        add_cpu(cpu);
+clear_cpumask:
+    for (i = 0; i < ne_cpu_pool.avail_cores_size; i++)
+        cpumask_clear(ne_cpu_pool.avail_cores[i]);
+free_cores_cpumask:
+    for (i = 0; i < ne_cpu_pool.avail_cores_size; i++)
+        free_cpumask_var(ne_cpu_pool.avail_cores[i]);
+    kfree(ne_cpu_pool.avail_cores);
+    ne_cpu_pool.avail_cores_size = 0;
+free_pool_cpumask:
+    free_cpumask_var(cpu_pool);
+    mutex_unlock(&ne_cpu_pool.mutex);
+
+    return rc;
+}
+
+/**
+ * ne_teardown_cpu_pool() - Online the CPUs from the NE CPU pool and cleanup the
+ *                CPU pool.
+ * @void:    No parameters provided.
+ *
+ * Context: Process context.
+ */
+static void ne_teardown_cpu_pool(void)
+{
+    unsigned int cpu = 0;
+    unsigned int i = 0;
+    int rc = -EINVAL;
+
+    mutex_lock(&ne_cpu_pool.mutex);
+
+    if (!ne_cpu_pool.avail_cores_size) {
+        mutex_unlock(&ne_cpu_pool.mutex);
+
+        return;
+    }
+
+    for (i = 0; i < ne_cpu_pool.avail_cores_size; i++) {
+        for_each_cpu(cpu, ne_cpu_pool.avail_cores[i]) {
+            rc = add_cpu(cpu);
+            if (rc != 0)
+                pr_err("%s: CPU %d is not onlined [rc=%d]\n",
+                       ne_misc_dev.name, cpu, rc);
+        }
+
+        cpumask_clear(ne_cpu_pool.avail_cores[i]);
+
+        free_cpumask_var(ne_cpu_pool.avail_cores[i]);
+    }
+
+    kfree(ne_cpu_pool.avail_cores);
+    ne_cpu_pool.avail_cores_size = 0;
+
+    mutex_unlock(&ne_cpu_pool.mutex);
+}
+
+/**
+ * ne_set_kernel_param() - Set the NE CPU pool value via the NE kernel parameter.
+ * @val:    NE CPU pool string value.
+ * @kp :    NE kernel parameter associated with the NE CPU pool.
+ *
+ * Context: Process context.
+ * Return:
+ * * 0 on success.
+ * * Negative return value on failure.
+ */
+static int ne_set_kernel_param(const char *val, const struct kernel_param *kp)
+{
+    char error_val[] = "";
+    int rc = -EINVAL;
+
+    if (!capable(CAP_SYS_ADMIN))
+        return -EPERM;
+
+    if (ne_check_enclaves_created()) {
+        pr_err("%s: The CPU pool is used by enclave(s)\n", ne_misc_dev.name);
+
+        return -EPERM;
+    }
+
+    ne_teardown_cpu_pool();
+
+    rc = ne_setup_cpu_pool(val);
+    if (rc < 0) {
+        pr_err("%s: Error in setup CPU pool [rc=%d]\n", ne_misc_dev.name, rc);
+
+        param_set_copystring(error_val, kp);
+
+        return rc;
+    }
+
+    return param_set_copystring(val, kp);
+}
+
+/**
+ * ne_get_cpu_from_cpu_pool() - Get a CPU from the NE CPU pool, either from the
+ *                remaining sibling(s) of a CPU core or the first
+ *                sibling of a new CPU core.
+ * @ne_enclave :    Private data associated with the current enclave.
+ *
+ * Context: Process context. This function is called with the ne_enclave mutex held.
+ * Return:
+ * * vCPU id.
+ * * 0, if no CPU available in the pool.
+ */
+static unsigned int ne_get_cpu_from_cpu_pool(struct ne_enclave *ne_enclave)
+{
+    int core_id = -1;
+    unsigned int cpu = 0;
+    unsigned int i = 0;
+    unsigned int vcpu_id = 0;
+
+    /* There are CPU siblings available to choose from. */
+    for (i = 0; i < ne_enclave->avail_cpu_cores_size; i++)
+        for_each_cpu(cpu, ne_enclave->avail_cpu_cores[i])

This is really hard to read. I think what you want to say here is something along these lines:

/*
 * If previously allocated a thread of a core to this enclave, we first
 * return its sibling(s) for new allocations, so that we can donate a
 * full core.
 */
for (i = 0; i < ne_enclave->nr_parent_cores; i++) {
    for_each_cpu(cpu, ne->enclave->parent_cores[i].reserved_siblings) {
        if (!ne_cpu_donated(cpu)) {
            [...]
        }
    }
}

Similarly for other places in this file where you do cpu mask logic. It always needs to be very obvious what the respective mask is for, because you have ~3 semantically different ones.

That's true, the motivation for all this logic is to split the NE CPU pool in threads / siblings per core and give to an enclave either the first sibling of a core or the remaining ones, till full core(s) is (are) given.

I'll update the naming, codebase logic and comments to make it more clear, this deserves it, indeed.

Thanks,
Andra



Amazon Development Center (Romania) S.R.L. registered office: 27A Sf. Lazar Street, UBC5, floor 2, Iasi, Iasi County, 700045, Romania. Registered in Romania. Registration number J22/2621/2005.