Skip to contentSkip to conversation

Enhance the hv_vmbus driver to support hibernation

PATCH Superseded Dexuan Cui 2 participants 14 messages Sep 5, 2019 – Sep 8, 2019

Subsystems (2)

v5 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Dexuan Cui

author maintainer
#0 lore raw

v5 cover Enhance the hv_vmbus driver to support hibernation

This patchset (consisting of 9 patches) was part of the v4 patchset (consisting of 12 patches):

    https://lkml.org/lkml/2019/9/2/894

The other 3 patches in v4 are posted in another patchset, which will go through the tip.git tree.

All the 9 patches here are now rebased to the hyperv tree's hyperv-next branch:
https://git.kernel.org/pub/scm/linux/kernel/git/hyperv/linux.git/log/?h=hyperv-next
, and all the 9 patches have Michael Kelley's Signed-off-by's.

Please review.

Thanks!
Dexuan

Dexuan Cui (9):
Drivers: hv: vmbus: Break out synic enable and disable operations
Drivers: hv: vmbus: Suspend/resume the synic for hibernation
Drivers: hv: vmbus: Add a helper function is_sub_channel()
Drivers: hv: vmbus: Implement suspend/resume for VSC drivers for
hibernation
Drivers: hv: vmbus: Ignore the offers when resuming from hibernation
Drivers: hv: vmbus: Suspend/resume the vmbus itself for hibernation
Drivers: hv: vmbus: Clean up hv_sock channels by force upon suspend
Drivers: hv: vmbus: Suspend after cleaning up hv_sock and sub channels
Drivers: hv: vmbus: Resume after fixing up old primary channels

 drivers/hv/channel_mgmt.c | 161 +++++++++++++++++++++++++---
 drivers/hv/connection.c   |   8 +-
 drivers/hv/hv.c           |  66 +++++++-----
 drivers/hv/hyperv_vmbus.h |  30 ++++++
 drivers/hv/vmbus_drv.c    | 265 ++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/hyperv.h    |  16 ++-
 6 files changed, 497 insertions(+), 49 deletions(-)
Signature · 1 line
1.8.3.1
author maintainer
#1 lore raw

PATCH v5 1/9 Drivers: hv: vmbus: Break out synic enable and disable operations

Break out synic enable and disable operations into separate hv_synic_disable_regs() and hv_synic_enable_regs() functions for use by a later patch to support hibernation.

There is no functional change except the unnecessary check
"if (sctrl.enable != 1) return -EFAULT;" which is removed, because when we're in hv_synic_cleanup(), we're absolutely sure sctrl.enable must be 1.

---
 drivers/hv/hv.c           | 66 ++++++++++++++++++++++++++---------------------
 drivers/hv/hyperv_vmbus.h |  2 ++
 2 files changed, 39 insertions(+), 29 deletions(-)
drivers/hv/hv.c +37 -29
index 6188fb7..fcc5279 100644
@@ -154,7 +154,7 @@ void hv_synic_free(void)
  * retrieve the initialized message and event pages.  Otherwise, we create and
  * initialize the message and event pages.
  */
-int hv_synic_init(unsigned int cpu)
+void hv_synic_enable_regs(unsigned int cpu)
 {
 	struct hv_per_cpu_context *hv_cpu
 		= per_cpu_ptr(hv_context.cpu_context, cpu);
@@ -196,6 +196,11 @@ int hv_synic_init(unsigned int cpu)
 	sctrl.enable = 1;
 
 	hv_set_synic_state(sctrl.as_uint64);
+}
+
+int hv_synic_init(unsigned int cpu)
+{
+	hv_synic_enable_regs(cpu);
 
 	hv_stimer_init(cpu);
 
@@ -205,20 +210,45 @@ int hv_synic_init(unsigned int cpu)
 /*
  * hv_synic_cleanup - Cleanup routine for hv_synic_init().
  */
-int hv_synic_cleanup(unsigned int cpu)
+void hv_synic_disable_regs(unsigned int cpu)
 {
 	union hv_synic_sint shared_sint;
 	union hv_synic_simp simp;
 	union hv_synic_siefp siefp;
 	union hv_synic_scontrol sctrl;
+
+	hv_get_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
+
+	shared_sint.masked = 1;
+
+	/* Need to correctly cleanup in the case of SMP!!! */
+	/* Disable the interrupt */
+	hv_set_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
+
+	hv_get_simp(simp.as_uint64);
+	simp.simp_enabled = 0;
+	simp.base_simp_gpa = 0;
+
+	hv_set_simp(simp.as_uint64);
+
+	hv_get_siefp(siefp.as_uint64);
+	siefp.siefp_enabled = 0;
+	siefp.base_siefp_gpa = 0;
+
+	hv_set_siefp(siefp.as_uint64);
+
+	/* Disable the global synic bit */
+	hv_get_synic_state(sctrl.as_uint64);
+	sctrl.enable = 0;
+	hv_set_synic_state(sctrl.as_uint64);
+}
+
+int hv_synic_cleanup(unsigned int cpu)
+{
 	struct vmbus_channel *channel, *sc;
 	bool channel_found = false;
 	unsigned long flags;
 
-	hv_get_synic_state(sctrl.as_uint64);
-	if (sctrl.enable != 1)
-		return -EFAULT;
-
 	/*
 	 * Search for channels which are bound to the CPU we're about to
 	 * cleanup. In case we find one and vmbus is still connected we need to
@@ -249,29 +279,7 @@ int hv_synic_cleanup(unsigned int cpu)
 
 	hv_stimer_cleanup(cpu);
 
-	hv_get_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
-
-	shared_sint.masked = 1;
-
-	/* Need to correctly cleanup in the case of SMP!!! */
-	/* Disable the interrupt */
-	hv_set_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
-
-	hv_get_simp(simp.as_uint64);
-	simp.simp_enabled = 0;
-	simp.base_simp_gpa = 0;
-
-	hv_set_simp(simp.as_uint64);
-
-	hv_get_siefp(siefp.as_uint64);
-	siefp.siefp_enabled = 0;
-	siefp.base_siefp_gpa = 0;
-
-	hv_set_siefp(siefp.as_uint64);
-
-	/* Disable the global synic bit */
-	sctrl.enable = 0;
-	hv_set_synic_state(sctrl.as_uint64);
+	hv_synic_disable_regs(cpu);
 
 	return 0;
 }
drivers/hv/hyperv_vmbus.h +2 -0
index 362e70e..26ea161 100644
@@ -171,8 +171,10 @@ extern int hv_post_message(union hv_connection_id connection_id,
 
 extern void hv_synic_free(void);
 
+extern void hv_synic_enable_regs(unsigned int cpu);
 extern int hv_synic_init(unsigned int cpu);
 
+extern void hv_synic_disable_regs(unsigned int cpu);
 extern int hv_synic_cleanup(unsigned int cpu);
 
 /* Interface */
Signature · 1 line
1.8.3.1
author maintainer
#2 lore raw

PATCH v5 2/9 Drivers: hv: vmbus: Suspend/resume the synic for hibernation

In reply to Dexuan Cui

This is needed when we resume the old kernel from the "current" kernel.

Note: when hv_synic_suspend() and hv_synic_resume() run, all the non-boot CPUs have been offlined, and interrupts are disabled on CPU0.

---
 drivers/hv/vmbus_drv.c | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 46 insertions(+)
drivers/hv/vmbus_drv.c +46 -0
index ebd35fc..2ef375c 100644
@@ -30,6 +30,7 @@
 #include <linux/kdebug.h>
 #include <linux/efi.h>
 #include <linux/random.h>
+#include <linux/syscore_ops.h>
 #include <clocksource/hyperv_timer.h>
 #include "hyperv_vmbus.h"
 
@@ -2086,6 +2087,47 @@ static void hv_crash_handler(struct pt_regs *regs)
 	hyperv_cleanup();
 };
 
+static int hv_synic_suspend(void)
+{
+	/*
+	 * When we reach here, all the non-boot CPUs have been offlined, and
+	 * the stimers on them have been unbound in hv_synic_cleanup() ->
+	 * hv_stimer_cleanup() -> clockevents_unbind_device().
+	 *
+	 * hv_synic_suspend() only runs on CPU0 with interrupts disabled. Here
+	 * we do not unbind the stimer on CPU0 because: 1) it's unnecessary
+	 * because the interrupts remain disabled between syscore_suspend()
+	 * and syscore_resume(): see create_image() and resume_target_kernel();
+	 * 2) the stimer on CPU0 is automatically disabled later by
+	 * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ...
+	 * -> clockevents_shutdown() -> ... -> hv_ce_shutdown(); 3) a warning
+	 * would be triggered if we call clockevents_unbind_device(), which
+	 * may sleep, in an interrupts-disabled context. So, we intentionally
+	 * don't call hv_stimer_cleanup(0) here.
+	 */
+
+	hv_synic_disable_regs(0);
+
+	return 0;
+}
+
+static void hv_synic_resume(void)
+{
+	hv_synic_enable_regs(0);
+
+	/*
+	 * Note: we don't need to call hv_stimer_init(0), because the timer
+	 * on CPU0 is not unbound in hv_synic_suspend(), and the timer is
+	 * automatically re-enabled in timekeeping_resume().
+	 */
+}
+
+/* The callbacks run only on CPU0, with irqs_disabled. */
+static struct syscore_ops hv_synic_syscore_ops = {
+	.suspend = hv_synic_suspend,
+	.resume = hv_synic_resume,
+};
+
 static int __init hv_acpi_init(void)
 {
 	int ret, t;
@@ -2116,6 +2158,8 @@ static int __init hv_acpi_init(void)
 	hv_setup_kexec_handler(hv_kexec_handler);
 	hv_setup_crash_handler(hv_crash_handler);
 
+	register_syscore_ops(&hv_synic_syscore_ops);
+
 	return 0;
 
 cleanup:
@@ -2128,6 +2172,8 @@ static void __exit vmbus_exit(void)
 {
 	int cpu;
 
+	unregister_syscore_ops(&hv_synic_syscore_ops);
+
 	hv_remove_kexec_handler();
 	hv_remove_crash_handler();
 	vmbus_connection.conn_state = DISCONNECTED;
Signature · 1 line
1.8.3.1
author maintainer
#3 lore raw

PATCH v5 3/9 Drivers: hv: vmbus: Add a helper function is_sub_channel()

In reply to Dexuan Cui

The existing method of telling if a channel is sub-channel in vmbus_process_offer() is cumbersome. This new simple helper function is preferred in future.

---
 include/linux/hyperv.h | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
include/linux/hyperv.h +9 -1
index 6256cc3..2d39248 100644
@@ -245,7 +245,10 @@ struct vmbus_channel_offer {
 		} pipe;
 	} u;
 	/*
-	 * The sub_channel_index is defined in win8.
+	 * The sub_channel_index is defined in Win8: a value of zero means a
+	 * primary channel and a value of non-zero means a sub-channel.
+	 *
+	 * Before Win8, the field is reserved, meaning it's always zero.
 	 */
 	u16 sub_channel_index;
 	u16 reserved3;
@@ -934,6 +937,11 @@ static inline bool is_hvsock_channel(const struct vmbus_channel *c)
 		  VMBUS_CHANNEL_TLNPI_PROVIDER_OFFER);
 }
 
+static inline bool is_sub_channel(const struct vmbus_channel *c)
+{
+	return c->offermsg.offer.sub_channel_index != 0;
+}
+
 static inline void set_channel_affinity_state(struct vmbus_channel *c,
 					      enum hv_numa_policy policy)
 {
Signature · 1 line
1.8.3.1
author maintainer
#4 lore raw

PATCH v5 4/9 Drivers: hv: vmbus: Implement suspend/resume for VSC drivers for hibernation

In reply to Dexuan Cui

The high-level VSC drivers will implement device-specific callbacks.

---
 drivers/hv/vmbus_drv.c | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/hyperv.h |  3 +++
 2 files changed, 49 insertions(+)
drivers/hv/vmbus_drv.c +46 -0
index 2ef375c..a30c70a 100644
@@ -911,6 +911,43 @@ static void vmbus_shutdown(struct device *child_device)
 		drv->shutdown(dev);
 }
 
+/*
+ * vmbus_suspend - Suspend a vmbus device
+ */
+static int vmbus_suspend(struct device *child_device)
+{
+	struct hv_driver *drv;
+	struct hv_device *dev = device_to_hv_device(child_device);
+
+	/* The device may not be attached yet */
+	if (!child_device->driver)
+		return 0;
+
+	drv = drv_to_hv_drv(child_device->driver);
+	if (!drv->suspend)
+		return -EOPNOTSUPP;
+
+	return drv->suspend(dev);
+}
+
+/*
+ * vmbus_resume - Resume a vmbus device
+ */
+static int vmbus_resume(struct device *child_device)
+{
+	struct hv_driver *drv;
+	struct hv_device *dev = device_to_hv_device(child_device);
+
+	/* The device may not be attached yet */
+	if (!child_device->driver)
+		return 0;
+
+	drv = drv_to_hv_drv(child_device->driver);
+	if (!drv->resume)
+		return -EOPNOTSUPP;
+
+	return drv->resume(dev);
+}
 
 /*
  * vmbus_device_release - Final callback release of the vmbus child device
@@ -926,6 +963,14 @@ static void vmbus_device_release(struct device *device)
 	kfree(hv_dev);
 }
 
+/*
+ * Note: we must use SET_NOIRQ_SYSTEM_SLEEP_PM_OPS rather than
+ * SET_SYSTEM_SLEEP_PM_OPS: see the comment before vmbus_bus_pm.
+ */
+static const struct dev_pm_ops vmbus_pm = {
+	SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(vmbus_suspend, vmbus_resume)
+};
+
 /* The one and only one */
 static struct bus_type  hv_bus = {
 	.name =		"vmbus",
@@ -936,6 +981,7 @@ static void vmbus_device_release(struct device *device)
 	.uevent =		vmbus_uevent,
 	.dev_groups =		vmbus_dev_groups,
 	.drv_groups =		vmbus_drv_groups,
+	.pm =			&vmbus_pm,
 };
 
 struct onmessage_work_context {
include/linux/hyperv.h +3 -0
index 2d39248..8a60e77 100644
@@ -1157,6 +1157,9 @@ struct hv_driver {
 	int (*remove)(struct hv_device *);
 	void (*shutdown)(struct hv_device *);
 
+	int (*suspend)(struct hv_device *);
+	int (*resume)(struct hv_device *);
+
 };
 
 /* Base device object */
Signature · 1 line
1.8.3.1
author maintainer
#5 lore raw

PATCH v5 5/9 Drivers: hv: vmbus: Ignore the offers when resuming from hibernation

In reply to Dexuan Cui

When the VM resumes, the host re-sends the offers. We should not add the offers to the global vmbus_connection.chn_list again.

This patch assumes the RELIDs of the channels don't change across hibernation. Actually this is not always true, especially in the case of NIC SR-IOV the VF vmbus device's RELID sometimes can change. A later patch will address this issue by mapping the new offers to the old channels and fixing up the old channels, if necessary.

---
 drivers/hv/channel_mgmt.c | 58 ++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 57 insertions(+), 1 deletion(-)
drivers/hv/channel_mgmt.c +57 -1
index addcef5..44b92fa 100644
@@ -848,18 +848,74 @@ void vmbus_initiate_unload(bool crash)
 }
 
 /*
+ * find_primary_channel_by_offer - Get the channel object given the new offer.
+ * This is only used in the resume path of hibernation.
+ */
+static struct vmbus_channel *
+find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
+{
+	struct vmbus_channel *channel = NULL, *iter;
+	const guid_t *inst1, *inst2;
+
+	/* Ignore sub-channel offers. */
+	if (offer->offer.sub_channel_index != 0)
+		return NULL;
+
+	mutex_lock(&vmbus_connection.channel_mutex);
+
+	list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
+		inst1 = &iter->offermsg.offer.if_instance;
+		inst2 = &offer->offer.if_instance;
+
+		if (guid_equal(inst1, inst2)) {
+			channel = iter;
+			break;
+		}
+	}
+
+	mutex_unlock(&vmbus_connection.channel_mutex);
+
+	return channel;
+}
+
+/*
  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
  *
  */
 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 {
 	struct vmbus_channel_offer_channel *offer;
-	struct vmbus_channel *newchannel;
+	struct vmbus_channel *oldchannel, *newchannel;
+	size_t offer_sz;
 
 	offer = (struct vmbus_channel_offer_channel *)hdr;
 
 	trace_vmbus_onoffer(offer);
 
+	oldchannel = find_primary_channel_by_offer(offer);
+
+	if (oldchannel != NULL) {
+		atomic_dec(&vmbus_connection.offer_in_progress);
+
+		/*
+		 * We're resuming from hibernation: we expect the host to send
+		 * exactly the same offers that we had before the hibernation.
+		 */
+		offer_sz = sizeof(*offer);
+		if (memcmp(offer, &oldchannel->offermsg, offer_sz) == 0)
+			return;
+
+		pr_debug("Mismatched offer from the host (relid=%d)\n",
+			 offer->child_relid);
+
+		print_hex_dump_debug("Old vmbus offer: ", DUMP_PREFIX_OFFSET,
+				     16, 4, &oldchannel->offermsg, offer_sz,
+				     false);
+		print_hex_dump_debug("New vmbus offer: ", DUMP_PREFIX_OFFSET,
+				     16, 4, offer, offer_sz, false);
+		return;
+	}
+
 	/* Allocate the channel object and save this offer. */
 	newchannel = alloc_channel();
 	if (!newchannel) {
Signature · 1 line
1.8.3.1
author maintainer
#6 lore raw

PATCH v5 6/9 Drivers: hv: vmbus: Suspend/resume the vmbus itself for hibernation

In reply to Dexuan Cui

Before Linux enters hibernation, it sends the CHANNELMSG_UNLOAD message to the host so all the offers are gone. After hibernation, Linux needs to re-negotiate with the host using the same vmbus protocol version (which was in use before hibernation), and ask the host to re-offer the vmbus devices.

---
 drivers/hv/connection.c   |  3 +--
 drivers/hv/hyperv_vmbus.h |  2 ++
 drivers/hv/vmbus_drv.c    | 59 +++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 62 insertions(+), 2 deletions(-)
drivers/hv/connection.c +1 -2
index 09829e1..806319c 100644
@@ -59,8 +59,7 @@ static __u32 vmbus_get_next_version(__u32 current_version)
 	}
 }
 
-static int vmbus_negotiate_version(struct vmbus_channel_msginfo *msginfo,
-					__u32 version)
+int vmbus_negotiate_version(struct vmbus_channel_msginfo *msginfo, u32 version)
 {
 	int ret = 0;
 	unsigned int cur_cpu;
drivers/hv/hyperv_vmbus.h +2 -0
index 26ea161..e657197 100644
@@ -274,6 +274,8 @@ struct vmbus_msginfo {
 
 extern struct vmbus_connection vmbus_connection;
 
+int vmbus_negotiate_version(struct vmbus_channel_msginfo *msginfo, u32 version);
+
 static inline void vmbus_send_interrupt(u32 relid)
 {
 	sync_set_bit(relid, vmbus_connection.send_int_page);
drivers/hv/vmbus_drv.c +59 -0

77 lines not on this page · raw message

Signature · 1 line
1.8.3.1
author maintainer
#7 lore raw

PATCH v5 7/9 Drivers: hv: vmbus: Clean up hv_sock channels by force upon suspend

In reply to Dexuan Cui

Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for hibernation. There is no better method to clean up the channels since some of the channels may still be referenced by the userspace apps when hibernation is triggered: in this case, with this patch, the "rescind" fields of the channels are set, and the apps will thoroughly destroy the channels after hibernation.

---
 drivers/hv/vmbus_drv.c | 55 ++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 55 insertions(+)
drivers/hv/vmbus_drv.c +55 -0
index ce9974b..45b976e 100644
@@ -24,6 +24,7 @@
 #include <linux/sched/task_stack.h>
 
 #include <asm/mshyperv.h>
+#include <linux/delay.h>
 #include <linux/notifier.h>
 #include <linux/ptrace.h>
 #include <linux/screen_info.h>
@@ -1069,6 +1070,41 @@ void vmbus_on_msg_dpc(unsigned long data)
 	vmbus_signal_eom(msg, message_type);
 }
 
+/*
+ * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for
+ * hibernation, because hv_sock connections can not persist across hibernation.
+ */
+static void vmbus_force_channel_rescinded(struct vmbus_channel *channel)
+{
+	struct onmessage_work_context *ctx;
+	struct vmbus_channel_rescind_offer *rescind;
+
+	WARN_ON(!is_hvsock_channel(channel));
+
+	/*
+	 * sizeof(*ctx) is small and the allocation should really not fail,
+	 * otherwise the state of the hv_sock connections ends up in limbo.
+	 */
+	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL | __GFP_NOFAIL);
+
+	/*
+	 * So far, these are not really used by Linux. Just set them to the
+	 * reasonable values conforming to the definitions of the fields.
+	 */
+	ctx->msg.header.message_type = 1;
+	ctx->msg.header.payload_size = sizeof(*rescind);
+
+	/* These values are actually used by Linux. */
+	rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.u.payload;
+	rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER;
+	rescind->child_relid = channel->offermsg.child_relid;
+
+	INIT_WORK(&ctx->work, vmbus_onmessage_work);
+
+	queue_work_on(vmbus_connection.connect_cpu,
+		      vmbus_connection.work_queue,
+		      &ctx->work);
+}
 
 /*
  * Direct callback for channels using other deferred processing
@@ -2091,6 +2127,25 @@ static int vmbus_acpi_add(struct acpi_device *device)
 
 static int vmbus_bus_suspend(struct device *dev)
 {
+	struct vmbus_channel *channel;
+
+	while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
+		/*
+		 * We wait here until the completion of any channel
+		 * offers that are currently in progress.
+		 */
+		msleep(1);
+	}
+
+	mutex_lock(&vmbus_connection.channel_mutex);
+	list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
+		if (!is_hvsock_channel(channel))
+			continue;
+
+		vmbus_force_channel_rescinded(channel);
+	}
+	mutex_unlock(&vmbus_connection.channel_mutex);
+
 	vmbus_initiate_unload(false);
 
 	vmbus_connection.conn_state = DISCONNECTED;
Signature · 1 line
1.8.3.1
author maintainer
#8 lore raw

PATCH v5 8/9 Drivers: hv: vmbus: Suspend after cleaning up hv_sock and sub channels

In reply to Dexuan Cui

Before suspend, Linux must make sure all the hv_sock channels have been properly cleaned up, because a hv_sock connection can not persist across hibernation, and the user-space app must be properly notified of the state change of the connection.

Before suspend, Linux also must make sure all the sub-channels have been destroyed, i.e. the related channel structs of the sub-channels must be properly removed, otherwise they would cause a conflict when the sub-channels are recreated upon resume.

Add a counter to track such channels, and vmbus_bus_suspend() should wait for the counter to drop to zero.

---
 drivers/hv/channel_mgmt.c | 26 ++++++++++++++++++++++++++
 drivers/hv/connection.c   |  3 +++
 drivers/hv/hyperv_vmbus.h | 12 ++++++++++++
 drivers/hv/vmbus_drv.c    | 44 +++++++++++++++++++++++++++++++++++++++++++-
 4 files changed, 84 insertions(+), 1 deletion(-)
drivers/hv/channel_mgmt.c +26 -0
index 44b92fa..5518d03 100644
@@ -545,6 +545,10 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
 
 	mutex_lock(&vmbus_connection.channel_mutex);
 
+	/* Remember the channels that should be cleaned up upon suspend. */
+	if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
+		atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
+
 	/*
 	 * Now that we have acquired the channel_mutex,
 	 * we can release the potentially racing rescind thread.
@@ -944,6 +948,16 @@ static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 	vmbus_process_offer(newchannel);
 }
 
+static void check_ready_for_suspend_event(void)
+{
+	/*
+	 * If all the sub-channels or hv_sock channels have been cleaned up,
+	 * then it's safe to suspend.
+	 */
+	if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
+		complete(&vmbus_connection.ready_for_suspend_event);
+}
+
 /*
  * vmbus_onoffer_rescind - Rescind offer handler.
  *
@@ -954,6 +968,7 @@ static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 	struct vmbus_channel_rescind_offer *rescind;
 	struct vmbus_channel *channel;
 	struct device *dev;
+	bool clean_up_chan_for_suspend;
 
 	rescind = (struct vmbus_channel_rescind_offer *)hdr;
 
@@ -993,6 +1008,8 @@ static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 		return;
 	}
 
+	clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
+				    is_sub_channel(channel);
 	/*
 	 * Before setting channel->rescind in vmbus_rescind_cleanup(), we
 	 * should make sure the channel callback is not running any more.
@@ -1018,6 +1035,10 @@ static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 	if (channel->device_obj) {
 		if (channel->chn_rescind_callback) {
 			channel->chn_rescind_callback(channel);
+
+			if (clean_up_chan_for_suspend)
+				check_ready_for_suspend_event();
+
 			return;
 		}
 		/*
@@ -1050,6 +1071,11 @@ static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 		}
 		mutex_unlock(&vmbus_connection.channel_mutex);
 	}
+
+	/* The "channel" may have been freed. Do not access it any longer. */
+
+	if (clean_up_chan_for_suspend)
+		check_ready_for_suspend_event();
 }
 
 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
drivers/hv/connection.c +3 -0
index 806319c..99851ea 100644
@@ -26,6 +26,9 @@
 struct vmbus_connection vmbus_connection = {
 	.conn_state		= DISCONNECTED,
 	.next_gpadl_handle	= ATOMIC_INIT(0xE1E10),
+
+	.ready_for_suspend_event= COMPLETION_INITIALIZER(
+				  vmbus_connection.ready_for_suspend_event),
 };
 EXPORT_SYMBOL_GPL(vmbus_connection);
 
drivers/hv/hyperv_vmbus.h +12 -0
index e657197..974b747 100644
@@ -260,6 +260,18 @@ struct vmbus_connection {
 	struct workqueue_struct *work_queue;
 	struct workqueue_struct *handle_primary_chan_wq;
 	struct workqueue_struct *handle_sub_chan_wq;
+
+	/*
+	 * The number of sub-channels and hv_sock channels that should be
+	 * cleaned up upon suspend: sub-channels will be re-created upon
+	 * resume, and hv_sock channels should not survive suspend.
+	 */
+	atomic_t nr_chan_close_on_suspend;
+	/*
+	 * vmbus_bus_suspend() waits for "nr_chan_close_on_suspend" to
+	 * drop to zero.
+	 */
+	struct completion ready_for_suspend_event;
 };
 
 
drivers/hv/vmbus_drv.c +43 -1

62 lines not on this page · raw message

Signature · 1 line
1.8.3.1
author maintainer
#9 lore raw

PATCH v5 9/9 Drivers: hv: vmbus: Resume after fixing up old primary channels

In reply to Dexuan Cui

When the host re-offers the primary channels upon resume, the host only guarantees the Instance GUID doesn't change, so vmbus_bus_suspend() should invalidate channel->offermsg.child_relid and figure out the number of primary channels that need to be fixed up upon resume.

Upon resume, vmbus_onoffer() finds the old channel structs, and maps the new offers to the old channels, and fixes up the old structs, and finally the resume callbacks of the VSC drivers will re-open the channels.

---
 drivers/hv/channel_mgmt.c | 85 ++++++++++++++++++++++++++++++++++++-----------
 drivers/hv/connection.c   |  2 ++
 drivers/hv/hyperv_vmbus.h | 14 ++++++++
 drivers/hv/vmbus_drv.c    | 17 ++++++++++
 include/linux/hyperv.h    |  3 ++
 5 files changed, 101 insertions(+), 20 deletions(-)
drivers/hv/channel_mgmt.c +65 -20
index 5518d03..8eb1675 100644
@@ -407,7 +407,15 @@ void hv_process_channel_removal(struct vmbus_channel *channel)
 		cpumask_clear_cpu(channel->target_cpu,
 				  &primary_channel->alloced_cpus_in_node);
 
-	vmbus_release_relid(channel->offermsg.child_relid);
+	/*
+	 * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
+	 * the relid is invalidated; after hibernation, when the user-space app
+	 * destroys the channel, the relid is INVALID_RELID, and in this case
+	 * it's unnecessary and unsafe to release the old relid, since the same
+	 * relid can refer to a completely different channel now.
+	 */
+	if (channel->offermsg.child_relid != INVALID_RELID)
+		vmbus_release_relid(channel->offermsg.child_relid);
 
 	free_channel(channel);
 }
@@ -851,6 +859,36 @@ void vmbus_initiate_unload(bool crash)
 		vmbus_wait_for_unload();
 }
 
+static void check_ready_for_resume_event(void)
+{
+	/*
+	 * If all the old primary channels have been fixed up, then it's safe
+	 * to resume.
+	 */
+	if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
+		complete(&vmbus_connection.ready_for_resume_event);
+}
+
+static void vmbus_setup_channel_state(struct vmbus_channel *channel,
+				      struct vmbus_channel_offer_channel *offer)
+{
+	/*
+	 * Setup state for signalling the host.
+	 */
+	channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
+
+	if (vmbus_proto_version != VERSION_WS2008) {
+		channel->is_dedicated_interrupt =
+				(offer->is_dedicated_interrupt != 0);
+		channel->sig_event = offer->connection_id;
+	}
+
+	memcpy(&channel->offermsg, offer,
+	       sizeof(struct vmbus_channel_offer_channel));
+	channel->monitor_grp = (u8)offer->monitorid / 32;
+	channel->monitor_bit = (u8)offer->monitorid % 32;
+}
+
 /*
  * find_primary_channel_by_offer - Get the channel object given the new offer.
  * This is only used in the resume path of hibernation.
@@ -902,14 +940,29 @@ static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 		atomic_dec(&vmbus_connection.offer_in_progress);
 
 		/*
-		 * We're resuming from hibernation: we expect the host to send
-		 * exactly the same offers that we had before the hibernation.
+		 * We're resuming from hibernation: all the sub-channel and
+		 * hv_sock channels we had before the hibernation should have
+		 * been cleaned up, and now we must be seeing a re-offered
+		 * primary channel that we had before the hibernation.
 		 */
+
+		WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
+		/* Fix up the relid. */
+		oldchannel->offermsg.child_relid = offer->child_relid;
+
 		offer_sz = sizeof(*offer);
-		if (memcmp(offer, &oldchannel->offermsg, offer_sz) == 0)
+		if (memcmp(offer, &oldchannel->offermsg, offer_sz) == 0) {
+			check_ready_for_resume_event();
 			return;
+		}
 
-		pr_debug("Mismatched offer from the host (relid=%d)\n",
+		/*
+		 * This is not an error, since the host can also change the
+		 * other field(s) of the offer, e.g. on WS RS5 (Build 17763),
+		 * the offer->connection_id of the Mellanox VF vmbus device
+		 * can change when the host reoffers the device upon resume.
+		 */
+		pr_debug("vmbus offer changed: relid=%d\n",
 			 offer->child_relid);
 
 		print_hex_dump_debug("Old vmbus offer: ", DUMP_PREFIX_OFFSET,
@@ -917,6 +970,12 @@ static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 				     false);
 		print_hex_dump_debug("New vmbus offer: ", DUMP_PREFIX_OFFSET,
 				     16, 4, offer, offer_sz, false);
+
+		/* Fix up the old channel. */
+		vmbus_setup_channel_state(oldchannel, offer);
+
+		check_ready_for_resume_event();
+
 		return;
 	}
 
@@ -929,21 +988,7 @@ static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 		return;
 	}
 
-	/*
-	 * Setup state for signalling the host.
-	 */
-	newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
-
-	if (vmbus_proto_version != VERSION_WS2008) {
-		newchannel->is_dedicated_interrupt =
-				(offer->is_dedicated_interrupt != 0);
-		newchannel->sig_event = offer->connection_id;
-	}
-
-	memcpy(&newchannel->offermsg, offer,
-	       sizeof(struct vmbus_channel_offer_channel));
-	newchannel->monitor_grp = (u8)offer->monitorid / 32;
-	newchannel->monitor_bit = (u8)offer->monitorid % 32;
+	vmbus_setup_channel_state(newchannel, offer);
 
 	vmbus_process_offer(newchannel);
 }
drivers/hv/connection.c +2 -0
index 99851ea..6e4c015 100644
@@ -29,6 +29,8 @@ struct vmbus_connection vmbus_connection = {
 
 	.ready_for_suspend_event= COMPLETION_INITIALIZER(
 				  vmbus_connection.ready_for_suspend_event),
+	.ready_for_resume_event	= COMPLETION_INITIALIZER(
+				  vmbus_connection.ready_for_resume_event),
 };
 EXPORT_SYMBOL_GPL(vmbus_connection);
 
drivers/hv/hyperv_vmbus.h +14 -0
index 974b747..f7a5f56 100644
@@ -272,6 +272,20 @@ struct vmbus_connection {
 	 * drop to zero.
 	 */
 	struct completion ready_for_suspend_event;
+
+	/*
+	 * The number of primary channels that should be "fixed up"
+	 * upon resume: these channels are re-offered upon resume, and some
+	 * fields of the channel offers (i.e. child_relid and connection_id)
+	 * can change, so the old offermsg must be fixed up, before the resume
+	 * callbacks of the VSC drivers start to further touch the channels.
+	 */
+	atomic_t nr_chan_fixup_on_resume;
+	/*
+	 * vmbus_bus_resume() waits for "nr_chan_fixup_on_resume" to
+	 * drop to zero.
+	 */
+	struct completion ready_for_resume_event;
 };
 
 
drivers/hv/vmbus_drv.c +17 -0
index 32ec951..391f0b2 100644
@@ -2164,9 +2164,17 @@ static int vmbus_bus_suspend(struct device *dev)
 	if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
 		wait_for_completion(&vmbus_connection.ready_for_suspend_event);
 
+	WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0);
+
 	mutex_lock(&vmbus_connection.channel_mutex);
 
 	list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
+		/*
+		 * Invalidate the field. Upon resume, vmbus_onoffer() will fix
+		 * up the field, and the other fields (if necessary).
+		 */
+		channel->offermsg.child_relid = INVALID_RELID;
+
 		if (is_hvsock_channel(channel)) {
 			if (!channel->rescind) {
 				pr_err("hv_sock channel not rescinded!\n");
@@ -2181,6 +2189,8 @@ static int vmbus_bus_suspend(struct device *dev)
 			WARN_ON_ONCE(1);
 		}
 		spin_unlock_irqrestore(&channel->lock, flags);
+
+		atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
 	}
 
 	mutex_unlock(&vmbus_connection.channel_mutex);
@@ -2189,6 +2199,9 @@ static int vmbus_bus_suspend(struct device *dev)
 
 	vmbus_connection.conn_state = DISCONNECTED;
 
+	/* Reset the event for the next resume. */
+	reinit_completion(&vmbus_connection.ready_for_resume_event);
+
 	return 0;
 }
 
@@ -2223,8 +2236,12 @@ static int vmbus_bus_resume(struct device *dev)
 	if (ret != 0)
 		return ret;
 
+	WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
+
 	vmbus_request_offers();
 
+	wait_for_completion(&vmbus_connection.ready_for_resume_event);
+
 	/* Reset the event for the next suspend. */
 	reinit_completion(&vmbus_connection.ready_for_suspend_event);
 
include/linux/hyperv.h +3 -0
index 8a60e77..a3aa9e9 100644
@@ -426,6 +426,9 @@ enum vmbus_channel_message_type {
 	CHANNELMSG_COUNT
 };
 
+/* Hyper-V supports about 2048 channels, and the RELIDs start with 1. */
+#define INVALID_RELID	U32_MAX
+
 struct vmbus_channel_message_header {
 	enum vmbus_channel_message_type msgtype;
 	u32 padding;
Signature · 1 line
1.8.3.1
#10 lore raw
In reply to Dexuan Cui
Quoted 4 lines · Dexuan Cui

This patchset (consisting of 9 patches) was part of the v4 patchset (consisting of 12 patches):

   https://lkml.org/lkml/2019/9/2/894
Quoted 8 lines

The other 3 patches in v4 are posted in another patchset, which will go through the tip.git tree.

All the 9 patches here are now rebased to the hyperv tree's hyperv-next branch:
https://git.kernel.org/pub/scm/linux/kernel/git/hyperv/linux.git/log/?h=hyperv-next
, and all the 9 patches have Michael Kelley's Signed-off-by's.

Please review.

Given that these two series depend on each other, I'd much prefer for them to go through one tree.

But, I may be wrong, and I'm going to see if a scenario such as this make sense. I've queued this one to the hyperv-next, but I'll wait for the x86 folks to send their pull request to Linus first before I do it for these patches.

Usually cases such as these are the exception, but for Hyper-V it seems to be the norm, so I'm curious to see how this will unfold.

--
Thanks,
Sasha

author maintainer
#11 lore raw
Quoted 9 lines

From: Sasha Levin <[email protected]>
Sent: Friday, September 6, 2019 1:03 PM

Quoted 6 lines · Dexuan Cui

This patchset (consisting of 9 patches) was part of the v4 patchset (consisting of 12 patches):

The other 3 patches in v4 are posted in another patchset, which will go through the tip.git tree.

Quoted 7 lines
Quoted 1 line

All the 9 patches here are now rebased to the hyperv tree's hyperv-next

branch, and all the 9 patches have Michael Kelley's Signed-off-by's.

Quoted 2 lines

Please review.

Given that these two series depend on each other, I'd much prefer for them to go through one tree.

Hi Sasha,
Yeah, that would be ideal. The problem here is: the other patchset conflicts with the existing patches in the tip.git tree's timers/core branch, so IMO the 3 patches have to go through the tip tree:

[PATCH v5 1/3] x86/hyper-v: Suspend/resume the hypercall page for hibernation [PATCH v5 2/3] x86/hyper-v: Implement hv_is_hibernation_supported() [PATCH v5 3/3] clocksource/drivers: Suspend/resume Hyper-V clocksource for hibernation

Quoted 4 lines

But, I may be wrong, and I'm going to see if a scenario such as this make sense. I've queued this one to the hyperv-next, but I'll wait for the x86 folks to send their pull request to Linus first before I do it for these patches.

Actually IMHO you don't need to wait, because there is not a build dependency, so either patchset can go into the Linus's tree first.

The 2 patchsets are just the first step to make hibernation work for Linux VM running on Hyper-V. Next I'm going to post some high-level VSC patches for hv_balloon, hv_utils, hv_netvsc, hid_hyperv, hv_storvsc, hyperv_keyboard, hyperv_fb,etc. All of these should go through the hyperv tree, since they're pure hyper-v changes, and they depend on this 9-patch patchset. I'll make a note in every patch so the subsystem maintainers will be aware and ack it.

Among the VSC patches, the hv_balloon patch does depend on the 2nd patch:
[PATCH v5 2/3] x86/hyper-v: Implement hv_is_hibernation_supported().
I think I'll wait for the aforementioned 2 patchsets to be in first, before posting the hv_balloon patch.

Quoted 5 lines

Usually cases such as these are the exception, but for Hyper-V it seems to be the norm, so I'm curious to see how this will unfold.

Thanks,
Sasha

Thanks for taking care all the patches!

Thanks,
-- Dexuan

#12 lore raw
Quoted 26 lines · Dexuan Cui
Quoted 9 lines

From: Sasha Levin <[email protected]>
Sent: Friday, September 6, 2019 1:03 PM

Quoted 6 lines · Dexuan Cui

This patchset (consisting of 9 patches) was part of the v4 patchset (consisting of 12 patches):

The other 3 patches in v4 are posted in another patchset, which will go through the tip.git tree.

Quoted 7 lines
Quoted 1 line

All the 9 patches here are now rebased to the hyperv tree's hyperv-next

branch, and all the 9 patches have Michael Kelley's Signed-off-by's.

Quoted 2 lines

Please review.

Given that these two series depend on each other, I'd much prefer for them to go through one tree.

Hi Sasha,
Yeah, that would be ideal. The problem here is: the other patchset conflicts with the existing patches in the tip.git tree's timers/core branch, so IMO the 3 patches have to go through the tip tree:

[PATCH v5 1/3] x86/hyper-v: Suspend/resume the hypercall page for hibernation [PATCH v5 2/3] x86/hyper-v: Implement hv_is_hibernation_supported() [PATCH v5 3/3] clocksource/drivers: Suspend/resume Hyper-V clocksource for hibernation

Quoted 7 lines
Quoted 4 lines

But, I may be wrong, and I'm going to see if a scenario such as this make sense. I've queued this one to the hyperv-next, but I'll wait for the x86 folks to send their pull request to Linus first before I do it for these patches.

Actually IMHO you don't need to wait, because there is not a build dependency, so either patchset can go into the Linus's tree first.

It'll build, sure. But did anyone actually test one without the other? What happens if Thomas doesn't send his batch at all during the merge window?

--
Thanks,
Sasha

author maintainer
#13 lore raw
Quoted 37 lines

From: Sasha Levin <[email protected]>
Sent: Sunday, September 8, 2019 5:13 AM

Quoted 4 lines · Dexuan Cui
Quoted 4 lines

From: Sasha Levin <[email protected]>
Sent: Friday, September 6, 2019 1:03 PM

Quoted 1 line · Dexuan Cui

This patchset (consisting of 9 patches) was part of the v4 patchset

(consisting

Quoted 13 lines
Quoted 5 lines
Quoted 5 lines

of 12 patches):

The other 3 patches in v4 are posted in another patchset, which will go through the tip.git tree.

Quoted 7 lines
Quoted 1 line

All the 9 patches here are now rebased to the hyperv tree's hyperv-next

branch, and all the 9 patches have Michael Kelley's Signed-off-by's.

Quoted 2 lines

Please review.

Given that these two series depend on each other, I'd much prefer for them to go through one tree.

Quoted 6 lines

Hi Sasha,
Yeah, that would be ideal. The problem here is: the other patchset conflicts with the existing patches in the tip.git tree's timers/core branch, so IMO the 3 patches have to go through the tip tree:

[PATCH v5 1/3] x86/hyper-v: Suspend/resume the hypercall page for

hibernation

Quoted 2 lines

[PATCH v5 2/3] x86/hyper-v: Implement hv_is_hibernation_supported() [PATCH v5 3/3] clocksource/drivers: Suspend/resume Hyper-V clocksource for

hibernation

Quoted 6 lines
Quoted 5 lines

But, I may be wrong, and I'm going to see if a scenario such as this make sense. I've queued this one to the hyperv-next, but I'll wait for the x86 folks to send their pull request to Linus first before I do it for these patches.

Quoted 4 lines
Quoted 2 lines

Actually IMHO you don't need to wait, because there is not a build dependency, so either patchset can go into the Linus's tree first.

It'll build, sure. But did anyone actually test one without the other?

Nobody tested this.

The fact is: even if we have the 2 patchsets, hibernation still can not work for Linux VM on Hyper-V, because we also need the high level driver changes to hv_netvsc, hv_storvsc, etc. I'm going to send out these patches soon.

Quoted 3 lines

What happens if Thomas doesn't send his batch at all during the merge window?
Sasha

We need all the patches together to make hibernation work.
I just meant that the 2 patchsets don't have to go into Linus's tree in a special order, as there is no build issue.

Thanks,
-- Dexuan