Re: [PATCH v19 038/130] KVM: TDX: create/destroy VM structure

[Yuan Yao]

On Fri, Mar 22, 2024 at 11:46:41AM +0800, Yuan Yao wrote:
> On Thu, Mar 21, 2024 at 07:17:09AM -0700, Isaku Yamahata wrote:
> > On Wed, Mar 20, 2024 at 01:12:01PM +0800,
> > Chao Gao <chao.gao@xxxxxxxxx> wrote:
..
> > > >+static int __tdx_td_init(struct kvm *kvm)
> > > >+{
> > > >+	struct kvm_tdx *kvm_tdx = to_kvm_tdx(kvm);
> > > >+	cpumask_var_t packages;
> > > >+	unsigned long *tdcs_pa = NULL;
> > > >+	unsigned long tdr_pa = 0;
> > > >+	unsigned long va;
> > > >+	int ret, i;
> > > >+	u64 err;
> > > >+
> > > >+	ret = tdx_guest_keyid_alloc();
> > > >+	if (ret < 0)
> > > >+		return ret;
> > > >+	kvm_tdx->hkid = ret;
> > > >+
> > > >+	va = __get_free_page(GFP_KERNEL_ACCOUNT);
> > > >+	if (!va)
> > > >+		goto free_hkid;
> > > >+	tdr_pa = __pa(va);
> > > >+
> > > >+	tdcs_pa = kcalloc(tdx_info->nr_tdcs_pages, sizeof(*kvm_tdx->tdcs_pa),
> > > >+			  GFP_KERNEL_ACCOUNT | __GFP_ZERO);
> > > >+	if (!tdcs_pa)
> > > >+		goto free_tdr;
> > > >+	for (i = 0; i < tdx_info->nr_tdcs_pages; i++) {
> > > >+		va = __get_free_page(GFP_KERNEL_ACCOUNT);
> > > >+		if (!va)
> > > >+			goto free_tdcs;
> > > >+		tdcs_pa[i] = __pa(va);
> > > >+	}
> > > >+
> > > >+	if (!zalloc_cpumask_var(&packages, GFP_KERNEL)) {
> > > >+		ret = -ENOMEM;
> > > >+		goto free_tdcs;
> > > >+	}
> > > >+	cpus_read_lock();
> > > >+	/*
> > > >+	 * Need at least one CPU of the package to be online in order to
> > > >+	 * program all packages for host key id.  Check it.
> > > >+	 */
> > > >+	for_each_present_cpu(i)
> > > >+		cpumask_set_cpu(topology_physical_package_id(i), packages);
> > > >+	for_each_online_cpu(i)
> > > >+		cpumask_clear_cpu(topology_physical_package_id(i), packages);
> > > >+	if (!cpumask_empty(packages)) {
> > > >+		ret = -EIO;
> > > >+		/*
> > > >+		 * Because it's hard for human operator to figure out the
> > > >+		 * reason, warn it.
> > > >+		 */
> > > >+#define MSG_ALLPKG	"All packages need to have online CPU to create TD. Online CPU and retry.\n"
> > > >+		pr_warn_ratelimited(MSG_ALLPKG);
> > > >+		goto free_packages;
> > > >+	}
> > > >+
> > > >+	/*
> > > >+	 * Acquire global lock to avoid TDX_OPERAND_BUSY:
> > > >+	 * TDH.MNG.CREATE and other APIs try to lock the global Key Owner
> > > >+	 * Table (KOT) to track the assigned TDX private HKID.  It doesn't spin
> > > >+	 * to acquire the lock, returns TDX_OPERAND_BUSY instead, and let the
> > > >+	 * caller to handle the contention.  This is because of time limitation
> > > >+	 * usable inside the TDX module and OS/VMM knows better about process
> > > >+	 * scheduling.
> > > >+	 *
> > > >+	 * APIs to acquire the lock of KOT:
> > > >+	 * TDH.MNG.CREATE, TDH.MNG.KEY.FREEID, TDH.MNG.VPFLUSHDONE, and
> > > >+	 * TDH.PHYMEM.CACHE.WB.
> > > >+	 */
> > > >+	mutex_lock(&tdx_lock);
> > > >+	err = tdh_mng_create(tdr_pa, kvm_tdx->hkid);
> > > >+	mutex_unlock(&tdx_lock);
> > > >+	if (err == TDX_RND_NO_ENTROPY) {
> > > >+		ret = -EAGAIN;
> > > >+		goto free_packages;
> > > >+	}
> > > >+	if (WARN_ON_ONCE(err)) {
> > > >+		pr_tdx_error(TDH_MNG_CREATE, err, NULL);
> > > >+		ret = -EIO;
> > > >+		goto free_packages;
> > > >+	}
> > > >+	kvm_tdx->tdr_pa = tdr_pa;
> > > >+
> > > >+	for_each_online_cpu(i) {
> > > >+		int pkg = topology_physical_package_id(i);
> > > >+
> > > >+		if (cpumask_test_and_set_cpu(pkg, packages))
> > > >+			continue;
> > > >+
> > > >+		/*
> > > >+		 * Program the memory controller in the package with an
> > > >+		 * encryption key associated to a TDX private host key id
> > > >+		 * assigned to this TDR.  Concurrent operations on same memory
> > > >+		 * controller results in TDX_OPERAND_BUSY.  Avoid this race by
> > > >+		 * mutex.
> > > >+		 */
> > > >+		mutex_lock(&tdx_mng_key_config_lock[pkg]);
> > >
> > > the lock is superfluous to me. with cpu lock held, even if multiple CPUs try to
> > > create TDs, the same set of CPUs (the first online CPU of each package) will be
> > > selected to configure the key because of the cpumask_test_and_set_cpu() above.
> > > it means, we never have two CPUs in the same socket trying to program the key,
> > > i.e., no concurrent calls.
> >
> > Makes sense. Will drop the lock.
>
> Not get the point, the variable "packages" on stack, and it's
> possible that "i" is same for 2 threads which are trying to create td.
> Anything I missed ?

Got the point after synced with chao.
in case of using for_each_online_cpu() it's safe to remove the mutex_lock(&tdx_mng_key_config_lock[pkg]),
since every thread will select only 1 cpu for each sockets in same order, and requests submited
to same cpu by smp_call_on_cpu() are ordered on the target cpu. That means removing the lock works for
using for_each_online_cpu() but does NOT work for randomly pick up a cpu per socket.

Maybe it's just my issue that doesn't realize what's going on here, but
I think it still worth to give comment here for why it works/does not work.

>
> >
> >
> > > >+		ret = smp_call_on_cpu(i, tdx_do_tdh_mng_key_config,
> > > >+				      &kvm_tdx->tdr_pa, true);
> > > >+		mutex_unlock(&tdx_mng_key_config_lock[pkg]);
> > > >+		if (ret)
> > > >+			break;
> > > >+	}
> > > >+	cpus_read_unlock();
> > > >+	free_cpumask_var(packages);
> > > >+	if (ret) {
> > > >+		i = 0;
> > > >+		goto teardown;
> > > >+	}
> > > >+
> > > >+	kvm_tdx->tdcs_pa = tdcs_pa;
> > > >+	for (i = 0; i < tdx_info->nr_tdcs_pages; i++) {
> > > >+		err = tdh_mng_addcx(kvm_tdx->tdr_pa, tdcs_pa[i]);
> > > >+		if (err == TDX_RND_NO_ENTROPY) {
> > > >+			/* Here it's hard to allow userspace to retry. */
> > > >+			ret = -EBUSY;
> > > >+			goto teardown;
> > > >+		}
> > > >+		if (WARN_ON_ONCE(err)) {
> > > >+			pr_tdx_error(TDH_MNG_ADDCX, err, NULL);
> > > >+			ret = -EIO;
> > > >+			goto teardown;
> > > >+		}
> > > >+	}
> > > >+
> > > >+	/*
> > > >+	 * Note, TDH_MNG_INIT cannot be invoked here.  TDH_MNG_INIT requires a dedicated
> > > >+	 * ioctl() to define the configure CPUID values for the TD.
> > > >+	 */
> > > >+	return 0;
> > > >+
> > > >+	/*
> > > >+	 * The sequence for freeing resources from a partially initialized TD
> > > >+	 * varies based on where in the initialization flow failure occurred.
> > > >+	 * Simply use the full teardown and destroy, which naturally play nice
> > > >+	 * with partial initialization.
> > > >+	 */
> > > >+teardown:
> > > >+	for (; i < tdx_info->nr_tdcs_pages; i++) {
> > > >+		if (tdcs_pa[i]) {
> > > >+			free_page((unsigned long)__va(tdcs_pa[i]));
> > > >+			tdcs_pa[i] = 0;
> > > >+		}
> > > >+	}
> > > >+	if (!kvm_tdx->tdcs_pa)
> > > >+		kfree(tdcs_pa);
> > > >+	tdx_mmu_release_hkid(kvm);
> > > >+	tdx_vm_free(kvm);
> > > >+	return ret;
> > > >+
> > > >+free_packages:
> > > >+	cpus_read_unlock();
> > > >+	free_cpumask_var(packages);
> > > >+free_tdcs:
> > > >+	for (i = 0; i < tdx_info->nr_tdcs_pages; i++) {
> > > >+		if (tdcs_pa[i])
> > > >+			free_page((unsigned long)__va(tdcs_pa[i]));
> > > >+	}
> > > >+	kfree(tdcs_pa);
> > > >+	kvm_tdx->tdcs_pa = NULL;
> > > >+
> > > >+free_tdr:
> > > >+	if (tdr_pa)
> > > >+		free_page((unsigned long)__va(tdr_pa));
> > > >+	kvm_tdx->tdr_pa = 0;
> > > >+free_hkid:
> > > >+	if (is_hkid_assigned(kvm_tdx))
> > >
> > > IIUC, this is always true because you just return if keyid
> > > allocation fails.
> >
> > You're right. Will fix
> > --
> > Isaku Yamahata <isaku.yamahata@xxxxxxxxx>
> >
>