linux/drivers/hv/hv_balloon.c

1633 lines
39 KiB
C
Raw Normal View History

/*
* Copyright (c) 2012, Microsoft Corporation.
*
* Author:
* K. Y. Srinivasan <kys@microsoft.com>
*
* This program is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 as published
* by the Free Software Foundation.
*
* This program is distributed in the hope that it will be useful, but
* WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
* NON INFRINGEMENT. See the GNU General Public License for more
* details.
*
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/kernel.h>
#include <linux/jiffies.h>
#include <linux/mman.h>
#include <linux/delay.h>
#include <linux/init.h>
#include <linux/module.h>
#include <linux/slab.h>
#include <linux/kthread.h>
#include <linux/completion.h>
#include <linux/memory_hotplug.h>
#include <linux/memory.h>
#include <linux/notifier.h>
#include <linux/percpu_counter.h>
#include <linux/hyperv.h>
/*
* We begin with definitions supporting the Dynamic Memory protocol
* with the host.
*
* Begin protocol definitions.
*/
/*
* Protocol versions. The low word is the minor version, the high word the major
* version.
*
* History:
* Initial version 1.0
* Changed to 0.1 on 2009/03/25
* Changes to 0.2 on 2009/05/14
* Changes to 0.3 on 2009/12/03
* Changed to 1.0 on 2011/04/05
*/
#define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
#define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
#define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
enum {
DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
};
/*
* Message Types
*/
enum dm_message_type {
/*
* Version 0.3
*/
DM_ERROR = 0,
DM_VERSION_REQUEST = 1,
DM_VERSION_RESPONSE = 2,
DM_CAPABILITIES_REPORT = 3,
DM_CAPABILITIES_RESPONSE = 4,
DM_STATUS_REPORT = 5,
DM_BALLOON_REQUEST = 6,
DM_BALLOON_RESPONSE = 7,
DM_UNBALLOON_REQUEST = 8,
DM_UNBALLOON_RESPONSE = 9,
DM_MEM_HOT_ADD_REQUEST = 10,
DM_MEM_HOT_ADD_RESPONSE = 11,
DM_VERSION_03_MAX = 11,
/*
* Version 1.0.
*/
DM_INFO_MESSAGE = 12,
DM_VERSION_1_MAX = 12
};
/*
* Structures defining the dynamic memory management
* protocol.
*/
union dm_version {
struct {
__u16 minor_version;
__u16 major_version;
};
__u32 version;
} __packed;
union dm_caps {
struct {
__u64 balloon:1;
__u64 hot_add:1;
/*
* To support guests that may have alignment
* limitations on hot-add, the guest can specify
* its alignment requirements; a value of n
* represents an alignment of 2^n in mega bytes.
*/
__u64 hot_add_alignment:4;
__u64 reservedz:58;
} cap_bits;
__u64 caps;
} __packed;
union dm_mem_page_range {
struct {
/*
* The PFN number of the first page in the range.
* 40 bits is the architectural limit of a PFN
* number for AMD64.
*/
__u64 start_page:40;
/*
* The number of pages in the range.
*/
__u64 page_cnt:24;
} finfo;
__u64 page_range;
} __packed;
/*
* The header for all dynamic memory messages:
*
* type: Type of the message.
* size: Size of the message in bytes; including the header.
* trans_id: The guest is responsible for manufacturing this ID.
*/
struct dm_header {
__u16 type;
__u16 size;
__u32 trans_id;
} __packed;
/*
* A generic message format for dynamic memory.
* Specific message formats are defined later in the file.
*/
struct dm_message {
struct dm_header hdr;
__u8 data[]; /* enclosed message */
} __packed;
/*
* Specific message types supporting the dynamic memory protocol.
*/
/*
* Version negotiation message. Sent from the guest to the host.
* The guest is free to try different versions until the host
* accepts the version.
*
* dm_version: The protocol version requested.
* is_last_attempt: If TRUE, this is the last version guest will request.
* reservedz: Reserved field, set to zero.
*/
struct dm_version_request {
struct dm_header hdr;
union dm_version version;
__u32 is_last_attempt:1;
__u32 reservedz:31;
} __packed;
/*
* Version response message; Host to Guest and indicates
* if the host has accepted the version sent by the guest.
*
* is_accepted: If TRUE, host has accepted the version and the guest
* should proceed to the next stage of the protocol. FALSE indicates that
* guest should re-try with a different version.
*
* reservedz: Reserved field, set to zero.
*/
struct dm_version_response {
struct dm_header hdr;
__u64 is_accepted:1;
__u64 reservedz:63;
} __packed;
/*
* Message reporting capabilities. This is sent from the guest to the
* host.
*/
struct dm_capabilities {
struct dm_header hdr;
union dm_caps caps;
__u64 min_page_cnt;
__u64 max_page_number;
} __packed;
/*
* Response to the capabilities message. This is sent from the host to the
* guest. This message notifies if the host has accepted the guest's
* capabilities. If the host has not accepted, the guest must shutdown
* the service.
*
* is_accepted: Indicates if the host has accepted guest's capabilities.
* reservedz: Must be 0.
*/
struct dm_capabilities_resp_msg {
struct dm_header hdr;
__u64 is_accepted:1;
__u64 reservedz:63;
} __packed;
/*
* This message is used to report memory pressure from the guest.
* This message is not part of any transaction and there is no
* response to this message.
*
* num_avail: Available memory in pages.
* num_committed: Committed memory in pages.
* page_file_size: The accumulated size of all page files
* in the system in pages.
* zero_free: The nunber of zero and free pages.
* page_file_writes: The writes to the page file in pages.
* io_diff: An indicator of file cache efficiency or page file activity,
* calculated as File Cache Page Fault Count - Page Read Count.
* This value is in pages.
*
* Some of these metrics are Windows specific and fortunately
* the algorithm on the host side that computes the guest memory
* pressure only uses num_committed value.
*/
struct dm_status {
struct dm_header hdr;
__u64 num_avail;
__u64 num_committed;
__u64 page_file_size;
__u64 zero_free;
__u32 page_file_writes;
__u32 io_diff;
} __packed;
/*
* Message to ask the guest to allocate memory - balloon up message.
* This message is sent from the host to the guest. The guest may not be
* able to allocate as much memory as requested.
*
* num_pages: number of pages to allocate.
*/
struct dm_balloon {
struct dm_header hdr;
__u32 num_pages;
__u32 reservedz;
} __packed;
/*
* Balloon response message; this message is sent from the guest
* to the host in response to the balloon message.
*
* reservedz: Reserved; must be set to zero.
* more_pages: If FALSE, this is the last message of the transaction.
* if TRUE there will atleast one more message from the guest.
*
* range_count: The number of ranges in the range array.
*
* range_array: An array of page ranges returned to the host.
*
*/
struct dm_balloon_response {
struct dm_header hdr;
__u32 reservedz;
__u32 more_pages:1;
__u32 range_count:31;
union dm_mem_page_range range_array[];
} __packed;
/*
* Un-balloon message; this message is sent from the host
* to the guest to give guest more memory.
*
* more_pages: If FALSE, this is the last message of the transaction.
* if TRUE there will atleast one more message from the guest.
*
* reservedz: Reserved; must be set to zero.
*
* range_count: The number of ranges in the range array.
*
* range_array: An array of page ranges returned to the host.
*
*/
struct dm_unballoon_request {
struct dm_header hdr;
__u32 more_pages:1;
__u32 reservedz:31;
__u32 range_count;
union dm_mem_page_range range_array[];
} __packed;
/*
* Un-balloon response message; this message is sent from the guest
* to the host in response to an unballoon request.
*
*/
struct dm_unballoon_response {
struct dm_header hdr;
} __packed;
/*
* Hot add request message. Message sent from the host to the guest.
*
* mem_range: Memory range to hot add.
*
* On Linux we currently don't support this since we cannot hot add
* arbitrary granularity of memory.
*/
struct dm_hot_add {
struct dm_header hdr;
union dm_mem_page_range range;
} __packed;
/*
* Hot add response message.
* This message is sent by the guest to report the status of a hot add request.
* If page_count is less than the requested page count, then the host should
* assume all further hot add requests will fail, since this indicates that
* the guest has hit an upper physical memory barrier.
*
* Hot adds may also fail due to low resources; in this case, the guest must
* not complete this message until the hot add can succeed, and the host must
* not send a new hot add request until the response is sent.
* If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
* times it fails the request.
*
*
* page_count: number of pages that were successfully hot added.
*
* result: result of the operation 1: success, 0: failure.
*
*/
struct dm_hot_add_response {
struct dm_header hdr;
__u32 page_count;
__u32 result;
} __packed;
/*
* Types of information sent from host to the guest.
*/
enum dm_info_type {
INFO_TYPE_MAX_PAGE_CNT = 0,
MAX_INFO_TYPE
};
/*
* Header for the information message.
*/
struct dm_info_header {
enum dm_info_type type;
__u32 data_size;
} __packed;
/*
* This message is sent from the host to the guest to pass
* some relevant information (win8 addition).
*
* reserved: no used.
* info_size: size of the information blob.
* info: information blob.
*/
struct dm_info_msg {
struct dm_header hdr;
__u32 reserved;
__u32 info_size;
__u8 info[];
};
/*
* End protocol definitions.
*/
/*
* State to manage hot adding memory into the guest.
* The range start_pfn : end_pfn specifies the range
* that the host has asked us to hot add. The range
* start_pfn : ha_end_pfn specifies the range that we have
* currently hot added. We hot add in multiples of 128M
* chunks; it is possible that we may not be able to bring
* online all the pages in the region. The range
Drivers: hv: hv_balloon: don't lose memory when onlining order is not natural Memory blocks can be onlined in random order. When this order is not natural some memory pages are not onlined because of the redundant check in hv_online_page(). Here is a real world scenario: 1) Host tries to hot-add the following (process_hot_add): pg_start=rg_start=0x48000, pfn_cnt=111616, rg_size=262144 2) This results in adding 4 memory blocks: [ 109.057866] init_memory_mapping: [mem 0x48000000-0x4fffffff] [ 114.102698] init_memory_mapping: [mem 0x50000000-0x57ffffff] [ 119.168039] init_memory_mapping: [mem 0x58000000-0x5fffffff] [ 124.233053] init_memory_mapping: [mem 0x60000000-0x67ffffff] The last one is incomplete but we have special has->covered_end_pfn counter to avoid onlining non-backed frames and hv_bring_pgs_online() function to bring them online later on. 3) Now we have 4 offline memory blocks: /sys/devices/system/memory/memory9-12 $ for f in /sys/devices/system/memory/memory*/state; do echo $f `cat $f`; done | grep -v onlin /sys/devices/system/memory/memory10/state offline /sys/devices/system/memory/memory11/state offline /sys/devices/system/memory/memory12/state offline /sys/devices/system/memory/memory9/state offline 4) We bring them online in non-natural order: $grep MemTotal /proc/meminfo MemTotal: 966348 kB $echo online > /sys/devices/system/memory/memory12/state && grep MemTotal /proc/meminfo MemTotal: 1019596 kB $echo online > /sys/devices/system/memory/memory11/state && grep MemTotal /proc/meminfo MemTotal: 1150668 kB $echo online > /sys/devices/system/memory/memory9/state && grep MemTotal /proc/meminfo MemTotal: 1150668 kB As you can see memory9 block gives us zero additional memory. We can also observe a huge discrepancy between host- and guest-reported memory sizes. The root cause of the issue is the redundant pg >= covered_start_pfn check (and covered_start_pfn advancing) in hv_online_page(). When upper memory block in being onlined before the lower one (memory12 and memory11 in the above case) we advance the covered_start_pfn pointer and all memory9 pages do not pass the check. If the assumption that host always gives us requests in sequential order and pg_start always equals rg_start when the first request for the new HA region is received (that's the case in my testing) is correct than we can get rid of covered_start_pfn and pg >= start_pfn check in hv_online_page() is sufficient. The current char-next branch is broken and this patch fixes the bug. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-03-18 19:29:23 +00:00
* covered_end_pfn defines the pages that can
* be brough online.
*/
struct hv_hotadd_state {
struct list_head list;
unsigned long start_pfn;
unsigned long covered_end_pfn;
unsigned long ha_end_pfn;
unsigned long end_pfn;
};
struct balloon_state {
__u32 num_pages;
struct work_struct wrk;
};
struct hot_add_wrk {
union dm_mem_page_range ha_page_range;
union dm_mem_page_range ha_region_range;
struct work_struct wrk;
};
static bool hot_add = true;
static bool do_hot_add;
/*
* Delay reporting memory pressure by
* the specified number of seconds.
*/
static uint pressure_report_delay = 45;
/*
* The last time we posted a pressure report to host.
*/
static unsigned long last_post_time;
module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
static atomic_t trans_id = ATOMIC_INIT(0);
static int dm_ring_size = (5 * PAGE_SIZE);
/*
* Driver specific state.
*/
enum hv_dm_state {
DM_INITIALIZING = 0,
DM_INITIALIZED,
DM_BALLOON_UP,
DM_BALLOON_DOWN,
DM_HOT_ADD,
DM_INIT_ERROR
};
static __u8 recv_buffer[PAGE_SIZE];
static __u8 *send_buffer;
#define PAGES_IN_2M 512
#define HA_CHUNK (32 * 1024)
struct hv_dynmem_device {
struct hv_device *dev;
enum hv_dm_state state;
struct completion host_event;
struct completion config_event;
/*
* Number of pages we have currently ballooned out.
*/
unsigned int num_pages_ballooned;
unsigned int num_pages_onlined;
unsigned int num_pages_added;
/*
* State to manage the ballooning (up) operation.
*/
struct balloon_state balloon_wrk;
/*
* State to execute the "hot-add" operation.
*/
struct hot_add_wrk ha_wrk;
/*
* This state tracks if the host has specified a hot-add
* region.
*/
bool host_specified_ha_region;
/*
* State to synchronize hot-add.
*/
struct completion ol_waitevent;
bool ha_waiting;
/*
* This thread handles hot-add
* requests from the host as well as notifying
* the host with regards to memory pressure in
* the guest.
*/
struct task_struct *thread;
struct mutex ha_region_mutex;
/*
* A list of hot-add regions.
*/
struct list_head ha_region_list;
/*
* We start with the highest version we can support
* and downgrade based on the host; we save here the
* next version to try.
*/
__u32 next_version;
};
static struct hv_dynmem_device dm_device;
static void post_status(struct hv_dynmem_device *dm);
#ifdef CONFIG_MEMORY_HOTPLUG
static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
void *v)
{
struct memory_notify *mem = (struct memory_notify *)v;
switch (val) {
case MEM_GOING_ONLINE:
Drivers: hv: hv_balloon: eliminate the trylock path in acquire/release_region_mutex When many memory regions are being added and automatically onlined the following lockup is sometimes observed: INFO: task udevd:1872 blocked for more than 120 seconds. ... Call Trace: [<ffffffff816ec0bc>] schedule_timeout+0x22c/0x350 [<ffffffff816eb98f>] wait_for_common+0x10f/0x160 [<ffffffff81067650>] ? default_wake_function+0x0/0x20 [<ffffffff816eb9fd>] wait_for_completion+0x1d/0x20 [<ffffffff8144cb9c>] hv_memory_notifier+0xdc/0x120 [<ffffffff816f298c>] notifier_call_chain+0x4c/0x70 ... When several memory blocks are going online simultaneously we got several hv_memory_notifier() trying to acquire the ha_region_mutex. When this mutex is being held by hot_add_req() all these competing acquire_region_mutex() do mutex_trylock, fail, and queue themselves into wait_for_completion(..). However when we do complete() from release_region_mutex() only one of them wakes up. This could be solved by changing complete() -> complete_all() memory onlining can be delayed as well, in that case we can still get several hv_memory_notifier() runners at the same time trying to grab the mutex. Only one of them will succeed and the others will hang for forever as complete() is not being called. We don't see this issue often because we have 5sec onlining timeout in hv_mem_hot_add() and usually all udev events arrive in this time frame. Get rid of the trylock path, waiting on the mutex is supposed to provide the required serialization. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-02-28 19:38:58 +00:00
mutex_lock(&dm_device.ha_region_mutex);
break;
case MEM_ONLINE:
dm_device.num_pages_onlined += mem->nr_pages;
case MEM_CANCEL_ONLINE:
Drivers: hv: balloon: check if ha_region_mutex was acquired in MEM_CANCEL_ONLINE case Memory notifiers are being executed in a sequential order and when one of them fails returning something different from NOTIFY_OK the remainder of the notification chain is not being executed. When a memory block is being onlined in online_pages() we do memory_notify(MEM_GOING_ONLINE, ) and if one of the notifiers in the chain fails we end up doing memory_notify(MEM_CANCEL_ONLINE, ) so it is possible for a notifier to see MEM_CANCEL_ONLINE without seeing the corresponding MEM_GOING_ONLINE event. E.g. when CONFIG_KASAN is enabled the kasan_mem_notifier() is being used to prevent memory hotplug, it returns NOTIFY_BAD for all MEM_GOING_ONLINE events. As kasan_mem_notifier() comes before the hv_memory_notifier() in the notification chain we don't see the MEM_GOING_ONLINE event and we do not take the ha_region_mutex. We, however, see the MEM_CANCEL_ONLINE event and unconditionally try to release the lock, the following is observed: [ 110.850927] ===================================== [ 110.850927] [ BUG: bad unlock balance detected! ] [ 110.850927] 4.1.0-rc3_bugxxxxxxx_test_xxxx #595 Not tainted [ 110.850927] ------------------------------------- [ 110.850927] systemd-udevd/920 is trying to release lock (&dm_device.ha_region_mutex) at: [ 110.850927] [<ffffffff81acda0e>] mutex_unlock+0xe/0x10 [ 110.850927] but there are no more locks to release! At the same time we can have the ha_region_mutex taken when we get the MEM_CANCEL_ONLINE event in case one of the memory notifiers after the hv_memory_notifier() in the notification chain failed so we need to add the mutex_is_locked() check. In case of MEM_ONLINE we are always supposed to have the mutex locked. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-05-29 18:18:02 +00:00
if (val == MEM_ONLINE ||
mutex_is_locked(&dm_device.ha_region_mutex))
mutex_unlock(&dm_device.ha_region_mutex);
if (dm_device.ha_waiting) {
dm_device.ha_waiting = false;
complete(&dm_device.ol_waitevent);
}
break;
case MEM_OFFLINE:
mutex_lock(&dm_device.ha_region_mutex);
dm_device.num_pages_onlined -= mem->nr_pages;
mutex_unlock(&dm_device.ha_region_mutex);
break;
case MEM_GOING_OFFLINE:
case MEM_CANCEL_OFFLINE:
break;
}
return NOTIFY_OK;
}
static struct notifier_block hv_memory_nb = {
.notifier_call = hv_memory_notifier,
.priority = 0
};
static void hv_bring_pgs_online(unsigned long start_pfn, unsigned long size)
{
int i;
for (i = 0; i < size; i++) {
struct page *pg;
pg = pfn_to_page(start_pfn + i);
__online_page_set_limits(pg);
__online_page_increment_counters(pg);
__online_page_free(pg);
}
}
static void hv_mem_hot_add(unsigned long start, unsigned long size,
unsigned long pfn_count,
struct hv_hotadd_state *has)
{
int ret = 0;
int i, nid;
unsigned long start_pfn;
unsigned long processed_pfn;
unsigned long total_pfn = pfn_count;
for (i = 0; i < (size/HA_CHUNK); i++) {
start_pfn = start + (i * HA_CHUNK);
has->ha_end_pfn += HA_CHUNK;
if (total_pfn > HA_CHUNK) {
processed_pfn = HA_CHUNK;
total_pfn -= HA_CHUNK;
} else {
processed_pfn = total_pfn;
total_pfn = 0;
}
has->covered_end_pfn += processed_pfn;
init_completion(&dm_device.ol_waitevent);
dm_device.ha_waiting = true;
Drivers: hv: hv_balloon: eliminate the trylock path in acquire/release_region_mutex When many memory regions are being added and automatically onlined the following lockup is sometimes observed: INFO: task udevd:1872 blocked for more than 120 seconds. ... Call Trace: [<ffffffff816ec0bc>] schedule_timeout+0x22c/0x350 [<ffffffff816eb98f>] wait_for_common+0x10f/0x160 [<ffffffff81067650>] ? default_wake_function+0x0/0x20 [<ffffffff816eb9fd>] wait_for_completion+0x1d/0x20 [<ffffffff8144cb9c>] hv_memory_notifier+0xdc/0x120 [<ffffffff816f298c>] notifier_call_chain+0x4c/0x70 ... When several memory blocks are going online simultaneously we got several hv_memory_notifier() trying to acquire the ha_region_mutex. When this mutex is being held by hot_add_req() all these competing acquire_region_mutex() do mutex_trylock, fail, and queue themselves into wait_for_completion(..). However when we do complete() from release_region_mutex() only one of them wakes up. This could be solved by changing complete() -> complete_all() memory onlining can be delayed as well, in that case we can still get several hv_memory_notifier() runners at the same time trying to grab the mutex. Only one of them will succeed and the others will hang for forever as complete() is not being called. We don't see this issue often because we have 5sec onlining timeout in hv_mem_hot_add() and usually all udev events arrive in this time frame. Get rid of the trylock path, waiting on the mutex is supposed to provide the required serialization. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-02-28 19:38:58 +00:00
mutex_unlock(&dm_device.ha_region_mutex);
nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
ret = add_memory(nid, PFN_PHYS((start_pfn)),
(HA_CHUNK << PAGE_SHIFT));
if (ret) {
pr_info("hot_add memory failed error is %d\n", ret);
if (ret == -EEXIST) {
/*
* This error indicates that the error
* is not a transient failure. This is the
* case where the guest's physical address map
* precludes hot adding memory. Stop all further
* memory hot-add.
*/
do_hot_add = false;
}
has->ha_end_pfn -= HA_CHUNK;
has->covered_end_pfn -= processed_pfn;
mutex_lock(&dm_device.ha_region_mutex);
break;
}
/*
* Wait for the memory block to be onlined.
* Since the hot add has succeeded, it is ok to
* proceed even if the pages in the hot added region
* have not been "onlined" within the allowed time.
*/
wait_for_completion_timeout(&dm_device.ol_waitevent, 5*HZ);
Drivers: hv: hv_balloon: eliminate the trylock path in acquire/release_region_mutex When many memory regions are being added and automatically onlined the following lockup is sometimes observed: INFO: task udevd:1872 blocked for more than 120 seconds. ... Call Trace: [<ffffffff816ec0bc>] schedule_timeout+0x22c/0x350 [<ffffffff816eb98f>] wait_for_common+0x10f/0x160 [<ffffffff81067650>] ? default_wake_function+0x0/0x20 [<ffffffff816eb9fd>] wait_for_completion+0x1d/0x20 [<ffffffff8144cb9c>] hv_memory_notifier+0xdc/0x120 [<ffffffff816f298c>] notifier_call_chain+0x4c/0x70 ... When several memory blocks are going online simultaneously we got several hv_memory_notifier() trying to acquire the ha_region_mutex. When this mutex is being held by hot_add_req() all these competing acquire_region_mutex() do mutex_trylock, fail, and queue themselves into wait_for_completion(..). However when we do complete() from release_region_mutex() only one of them wakes up. This could be solved by changing complete() -> complete_all() memory onlining can be delayed as well, in that case we can still get several hv_memory_notifier() runners at the same time trying to grab the mutex. Only one of them will succeed and the others will hang for forever as complete() is not being called. We don't see this issue often because we have 5sec onlining timeout in hv_mem_hot_add() and usually all udev events arrive in this time frame. Get rid of the trylock path, waiting on the mutex is supposed to provide the required serialization. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-02-28 19:38:58 +00:00
mutex_lock(&dm_device.ha_region_mutex);
post_status(&dm_device);
}
return;
}
static void hv_online_page(struct page *pg)
{
struct list_head *cur;
struct hv_hotadd_state *has;
unsigned long cur_start_pgp;
unsigned long cur_end_pgp;
list_for_each(cur, &dm_device.ha_region_list) {
has = list_entry(cur, struct hv_hotadd_state, list);
Drivers: hv: hv_balloon: don't lose memory when onlining order is not natural Memory blocks can be onlined in random order. When this order is not natural some memory pages are not onlined because of the redundant check in hv_online_page(). Here is a real world scenario: 1) Host tries to hot-add the following (process_hot_add): pg_start=rg_start=0x48000, pfn_cnt=111616, rg_size=262144 2) This results in adding 4 memory blocks: [ 109.057866] init_memory_mapping: [mem 0x48000000-0x4fffffff] [ 114.102698] init_memory_mapping: [mem 0x50000000-0x57ffffff] [ 119.168039] init_memory_mapping: [mem 0x58000000-0x5fffffff] [ 124.233053] init_memory_mapping: [mem 0x60000000-0x67ffffff] The last one is incomplete but we have special has->covered_end_pfn counter to avoid onlining non-backed frames and hv_bring_pgs_online() function to bring them online later on. 3) Now we have 4 offline memory blocks: /sys/devices/system/memory/memory9-12 $ for f in /sys/devices/system/memory/memory*/state; do echo $f `cat $f`; done | grep -v onlin /sys/devices/system/memory/memory10/state offline /sys/devices/system/memory/memory11/state offline /sys/devices/system/memory/memory12/state offline /sys/devices/system/memory/memory9/state offline 4) We bring them online in non-natural order: $grep MemTotal /proc/meminfo MemTotal: 966348 kB $echo online > /sys/devices/system/memory/memory12/state && grep MemTotal /proc/meminfo MemTotal: 1019596 kB $echo online > /sys/devices/system/memory/memory11/state && grep MemTotal /proc/meminfo MemTotal: 1150668 kB $echo online > /sys/devices/system/memory/memory9/state && grep MemTotal /proc/meminfo MemTotal: 1150668 kB As you can see memory9 block gives us zero additional memory. We can also observe a huge discrepancy between host- and guest-reported memory sizes. The root cause of the issue is the redundant pg >= covered_start_pfn check (and covered_start_pfn advancing) in hv_online_page(). When upper memory block in being onlined before the lower one (memory12 and memory11 in the above case) we advance the covered_start_pfn pointer and all memory9 pages do not pass the check. If the assumption that host always gives us requests in sequential order and pg_start always equals rg_start when the first request for the new HA region is received (that's the case in my testing) is correct than we can get rid of covered_start_pfn and pg >= start_pfn check in hv_online_page() is sufficient. The current char-next branch is broken and this patch fixes the bug. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-03-18 19:29:23 +00:00
cur_start_pgp = (unsigned long)pfn_to_page(has->start_pfn);
cur_end_pgp = (unsigned long)pfn_to_page(has->covered_end_pfn);
if (((unsigned long)pg >= cur_start_pgp) &&
((unsigned long)pg < cur_end_pgp)) {
/*
* This frame is currently backed; online the
* page.
*/
__online_page_set_limits(pg);
__online_page_increment_counters(pg);
__online_page_free(pg);
}
}
}
static bool pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
{
struct list_head *cur;
struct hv_hotadd_state *has;
unsigned long residual, new_inc;
if (list_empty(&dm_device.ha_region_list))
return false;
list_for_each(cur, &dm_device.ha_region_list) {
has = list_entry(cur, struct hv_hotadd_state, list);
/*
* If the pfn range we are dealing with is not in the current
* "hot add block", move on.
*/
if ((start_pfn >= has->end_pfn))
continue;
/*
* If the current hot add-request extends beyond
* our current limit; extend it.
*/
if ((start_pfn + pfn_cnt) > has->end_pfn) {
residual = (start_pfn + pfn_cnt - has->end_pfn);
/*
* Extend the region by multiples of HA_CHUNK.
*/
new_inc = (residual / HA_CHUNK) * HA_CHUNK;
if (residual % HA_CHUNK)
new_inc += HA_CHUNK;
has->end_pfn += new_inc;
}
/*
* If the current start pfn is not where the covered_end
* is, update it.
*/
Drivers: hv: hv_balloon: don't lose memory when onlining order is not natural Memory blocks can be onlined in random order. When this order is not natural some memory pages are not onlined because of the redundant check in hv_online_page(). Here is a real world scenario: 1) Host tries to hot-add the following (process_hot_add): pg_start=rg_start=0x48000, pfn_cnt=111616, rg_size=262144 2) This results in adding 4 memory blocks: [ 109.057866] init_memory_mapping: [mem 0x48000000-0x4fffffff] [ 114.102698] init_memory_mapping: [mem 0x50000000-0x57ffffff] [ 119.168039] init_memory_mapping: [mem 0x58000000-0x5fffffff] [ 124.233053] init_memory_mapping: [mem 0x60000000-0x67ffffff] The last one is incomplete but we have special has->covered_end_pfn counter to avoid onlining non-backed frames and hv_bring_pgs_online() function to bring them online later on. 3) Now we have 4 offline memory blocks: /sys/devices/system/memory/memory9-12 $ for f in /sys/devices/system/memory/memory*/state; do echo $f `cat $f`; done | grep -v onlin /sys/devices/system/memory/memory10/state offline /sys/devices/system/memory/memory11/state offline /sys/devices/system/memory/memory12/state offline /sys/devices/system/memory/memory9/state offline 4) We bring them online in non-natural order: $grep MemTotal /proc/meminfo MemTotal: 966348 kB $echo online > /sys/devices/system/memory/memory12/state && grep MemTotal /proc/meminfo MemTotal: 1019596 kB $echo online > /sys/devices/system/memory/memory11/state && grep MemTotal /proc/meminfo MemTotal: 1150668 kB $echo online > /sys/devices/system/memory/memory9/state && grep MemTotal /proc/meminfo MemTotal: 1150668 kB As you can see memory9 block gives us zero additional memory. We can also observe a huge discrepancy between host- and guest-reported memory sizes. The root cause of the issue is the redundant pg >= covered_start_pfn check (and covered_start_pfn advancing) in hv_online_page(). When upper memory block in being onlined before the lower one (memory12 and memory11 in the above case) we advance the covered_start_pfn pointer and all memory9 pages do not pass the check. If the assumption that host always gives us requests in sequential order and pg_start always equals rg_start when the first request for the new HA region is received (that's the case in my testing) is correct than we can get rid of covered_start_pfn and pg >= start_pfn check in hv_online_page() is sufficient. The current char-next branch is broken and this patch fixes the bug. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-03-18 19:29:23 +00:00
if (has->covered_end_pfn != start_pfn)
has->covered_end_pfn = start_pfn;
Drivers: hv: hv_balloon: don't lose memory when onlining order is not natural Memory blocks can be onlined in random order. When this order is not natural some memory pages are not onlined because of the redundant check in hv_online_page(). Here is a real world scenario: 1) Host tries to hot-add the following (process_hot_add): pg_start=rg_start=0x48000, pfn_cnt=111616, rg_size=262144 2) This results in adding 4 memory blocks: [ 109.057866] init_memory_mapping: [mem 0x48000000-0x4fffffff] [ 114.102698] init_memory_mapping: [mem 0x50000000-0x57ffffff] [ 119.168039] init_memory_mapping: [mem 0x58000000-0x5fffffff] [ 124.233053] init_memory_mapping: [mem 0x60000000-0x67ffffff] The last one is incomplete but we have special has->covered_end_pfn counter to avoid onlining non-backed frames and hv_bring_pgs_online() function to bring them online later on. 3) Now we have 4 offline memory blocks: /sys/devices/system/memory/memory9-12 $ for f in /sys/devices/system/memory/memory*/state; do echo $f `cat $f`; done | grep -v onlin /sys/devices/system/memory/memory10/state offline /sys/devices/system/memory/memory11/state offline /sys/devices/system/memory/memory12/state offline /sys/devices/system/memory/memory9/state offline 4) We bring them online in non-natural order: $grep MemTotal /proc/meminfo MemTotal: 966348 kB $echo online > /sys/devices/system/memory/memory12/state && grep MemTotal /proc/meminfo MemTotal: 1019596 kB $echo online > /sys/devices/system/memory/memory11/state && grep MemTotal /proc/meminfo MemTotal: 1150668 kB $echo online > /sys/devices/system/memory/memory9/state && grep MemTotal /proc/meminfo MemTotal: 1150668 kB As you can see memory9 block gives us zero additional memory. We can also observe a huge discrepancy between host- and guest-reported memory sizes. The root cause of the issue is the redundant pg >= covered_start_pfn check (and covered_start_pfn advancing) in hv_online_page(). When upper memory block in being onlined before the lower one (memory12 and memory11 in the above case) we advance the covered_start_pfn pointer and all memory9 pages do not pass the check. If the assumption that host always gives us requests in sequential order and pg_start always equals rg_start when the first request for the new HA region is received (that's the case in my testing) is correct than we can get rid of covered_start_pfn and pg >= start_pfn check in hv_online_page() is sufficient. The current char-next branch is broken and this patch fixes the bug. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-03-18 19:29:23 +00:00
return true;
}
return false;
}
static unsigned long handle_pg_range(unsigned long pg_start,
unsigned long pg_count)
{
unsigned long start_pfn = pg_start;
unsigned long pfn_cnt = pg_count;
unsigned long size;
struct list_head *cur;
struct hv_hotadd_state *has;
unsigned long pgs_ol = 0;
unsigned long old_covered_state;
if (list_empty(&dm_device.ha_region_list))
return 0;
list_for_each(cur, &dm_device.ha_region_list) {
has = list_entry(cur, struct hv_hotadd_state, list);
/*
* If the pfn range we are dealing with is not in the current
* "hot add block", move on.
*/
if ((start_pfn >= has->end_pfn))
continue;
old_covered_state = has->covered_end_pfn;
if (start_pfn < has->ha_end_pfn) {
/*
* This is the case where we are backing pages
* in an already hot added region. Bring
* these pages online first.
*/
pgs_ol = has->ha_end_pfn - start_pfn;
if (pgs_ol > pfn_cnt)
pgs_ol = pfn_cnt;
/*
* Check if the corresponding memory block is already
* online by checking its last previously backed page.
* In case it is we need to bring rest (which was not
* backed previously) online too.
*/
if (start_pfn > has->start_pfn &&
!PageReserved(pfn_to_page(start_pfn - 1)))
hv_bring_pgs_online(start_pfn, pgs_ol);
has->covered_end_pfn += pgs_ol;
pfn_cnt -= pgs_ol;
}
if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
/*
* We have some residual hot add range
* that needs to be hot added; hot add
* it now. Hot add a multiple of
* of HA_CHUNK that fully covers the pages
* we have.
*/
size = (has->end_pfn - has->ha_end_pfn);
if (pfn_cnt <= size) {
size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
if (pfn_cnt % HA_CHUNK)
size += HA_CHUNK;
} else {
pfn_cnt = size;
}
hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
}
/*
* If we managed to online any pages that were given to us,
* we declare success.
*/
return has->covered_end_pfn - old_covered_state;
}
return 0;
}
static unsigned long process_hot_add(unsigned long pg_start,
unsigned long pfn_cnt,
unsigned long rg_start,
unsigned long rg_size)
{
struct hv_hotadd_state *ha_region = NULL;
if (pfn_cnt == 0)
return 0;
if (!dm_device.host_specified_ha_region)
if (pfn_covered(pg_start, pfn_cnt))
goto do_pg_range;
/*
* If the host has specified a hot-add range; deal with it first.
*/
if (rg_size != 0) {
ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
if (!ha_region)
return 0;
INIT_LIST_HEAD(&ha_region->list);
list_add_tail(&ha_region->list, &dm_device.ha_region_list);
ha_region->start_pfn = rg_start;
ha_region->ha_end_pfn = rg_start;
ha_region->covered_end_pfn = pg_start;
ha_region->end_pfn = rg_start + rg_size;
}
do_pg_range:
/*
* Process the page range specified; bringing them
* online if possible.
*/
return handle_pg_range(pg_start, pfn_cnt);
}
#endif
static void hot_add_req(struct work_struct *dummy)
{
struct dm_hot_add_response resp;
#ifdef CONFIG_MEMORY_HOTPLUG
unsigned long pg_start, pfn_cnt;
unsigned long rg_start, rg_sz;
#endif
struct hv_dynmem_device *dm = &dm_device;
memset(&resp, 0, sizeof(struct dm_hot_add_response));
resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
resp.hdr.size = sizeof(struct dm_hot_add_response);
#ifdef CONFIG_MEMORY_HOTPLUG
Drivers: hv: hv_balloon: eliminate the trylock path in acquire/release_region_mutex When many memory regions are being added and automatically onlined the following lockup is sometimes observed: INFO: task udevd:1872 blocked for more than 120 seconds. ... Call Trace: [<ffffffff816ec0bc>] schedule_timeout+0x22c/0x350 [<ffffffff816eb98f>] wait_for_common+0x10f/0x160 [<ffffffff81067650>] ? default_wake_function+0x0/0x20 [<ffffffff816eb9fd>] wait_for_completion+0x1d/0x20 [<ffffffff8144cb9c>] hv_memory_notifier+0xdc/0x120 [<ffffffff816f298c>] notifier_call_chain+0x4c/0x70 ... When several memory blocks are going online simultaneously we got several hv_memory_notifier() trying to acquire the ha_region_mutex. When this mutex is being held by hot_add_req() all these competing acquire_region_mutex() do mutex_trylock, fail, and queue themselves into wait_for_completion(..). However when we do complete() from release_region_mutex() only one of them wakes up. This could be solved by changing complete() -> complete_all() memory onlining can be delayed as well, in that case we can still get several hv_memory_notifier() runners at the same time trying to grab the mutex. Only one of them will succeed and the others will hang for forever as complete() is not being called. We don't see this issue often because we have 5sec onlining timeout in hv_mem_hot_add() and usually all udev events arrive in this time frame. Get rid of the trylock path, waiting on the mutex is supposed to provide the required serialization. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-02-28 19:38:58 +00:00
mutex_lock(&dm_device.ha_region_mutex);
pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
unsigned long region_size;
unsigned long region_start;
/*
* The host has not specified the hot-add region.
* Based on the hot-add page range being specified,
* compute a hot-add region that can cover the pages
* that need to be hot-added while ensuring the alignment
* and size requirements of Linux as it relates to hot-add.
*/
region_start = pg_start;
region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
if (pfn_cnt % HA_CHUNK)
region_size += HA_CHUNK;
region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
rg_start = region_start;
rg_sz = region_size;
}
if (do_hot_add)
resp.page_count = process_hot_add(pg_start, pfn_cnt,
rg_start, rg_sz);
dm->num_pages_added += resp.page_count;
Drivers: hv: hv_balloon: eliminate the trylock path in acquire/release_region_mutex When many memory regions are being added and automatically onlined the following lockup is sometimes observed: INFO: task udevd:1872 blocked for more than 120 seconds. ... Call Trace: [<ffffffff816ec0bc>] schedule_timeout+0x22c/0x350 [<ffffffff816eb98f>] wait_for_common+0x10f/0x160 [<ffffffff81067650>] ? default_wake_function+0x0/0x20 [<ffffffff816eb9fd>] wait_for_completion+0x1d/0x20 [<ffffffff8144cb9c>] hv_memory_notifier+0xdc/0x120 [<ffffffff816f298c>] notifier_call_chain+0x4c/0x70 ... When several memory blocks are going online simultaneously we got several hv_memory_notifier() trying to acquire the ha_region_mutex. When this mutex is being held by hot_add_req() all these competing acquire_region_mutex() do mutex_trylock, fail, and queue themselves into wait_for_completion(..). However when we do complete() from release_region_mutex() only one of them wakes up. This could be solved by changing complete() -> complete_all() memory onlining can be delayed as well, in that case we can still get several hv_memory_notifier() runners at the same time trying to grab the mutex. Only one of them will succeed and the others will hang for forever as complete() is not being called. We don't see this issue often because we have 5sec onlining timeout in hv_mem_hot_add() and usually all udev events arrive in this time frame. Get rid of the trylock path, waiting on the mutex is supposed to provide the required serialization. Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-02-28 19:38:58 +00:00
mutex_unlock(&dm_device.ha_region_mutex);
#endif
/*
* The result field of the response structure has the
* following semantics:
*
* 1. If all or some pages hot-added: Guest should return success.
*
* 2. If no pages could be hot-added:
*
* If the guest returns success, then the host
* will not attempt any further hot-add operations. This
* signifies a permanent failure.
*
* If the guest returns failure, then this failure will be
* treated as a transient failure and the host may retry the
* hot-add operation after some delay.
*/
if (resp.page_count > 0)
resp.result = 1;
else if (!do_hot_add)
resp.result = 1;
else
resp.result = 0;
if (!do_hot_add || (resp.page_count == 0))
pr_info("Memory hot add failed\n");
dm->state = DM_INITIALIZED;
resp.hdr.trans_id = atomic_inc_return(&trans_id);
vmbus_sendpacket(dm->dev->channel, &resp,
sizeof(struct dm_hot_add_response),
(unsigned long)NULL,
VM_PKT_DATA_INBAND, 0);
}
static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
{
struct dm_info_header *info_hdr;
info_hdr = (struct dm_info_header *)msg->info;
switch (info_hdr->type) {
case INFO_TYPE_MAX_PAGE_CNT:
pr_info("Received INFO_TYPE_MAX_PAGE_CNT\n");
pr_info("Data Size is %d\n", info_hdr->data_size);
break;
default:
pr_info("Received Unknown type: %d\n", info_hdr->type);
}
}
static unsigned long compute_balloon_floor(void)
{
unsigned long min_pages;
#define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
/* Simple continuous piecewiese linear function:
* max MiB -> min MiB gradient
* 0 0
* 16 16
* 32 24
* 128 72 (1/2)
* 512 168 (1/4)
* 2048 360 (1/8)
* 8192 744 (1/16)
* 32768 1512 (1/32)
*/
if (totalram_pages < MB2PAGES(128))
min_pages = MB2PAGES(8) + (totalram_pages >> 1);
else if (totalram_pages < MB2PAGES(512))
min_pages = MB2PAGES(40) + (totalram_pages >> 2);
else if (totalram_pages < MB2PAGES(2048))
min_pages = MB2PAGES(104) + (totalram_pages >> 3);
else if (totalram_pages < MB2PAGES(8192))
min_pages = MB2PAGES(232) + (totalram_pages >> 4);
else
min_pages = MB2PAGES(488) + (totalram_pages >> 5);
#undef MB2PAGES
return min_pages;
}
/*
* Post our status as it relates memory pressure to the
* host. Host expects the guests to post this status
* periodically at 1 second intervals.
*
* The metrics specified in this protocol are very Windows
* specific and so we cook up numbers here to convey our memory
* pressure.
*/
static void post_status(struct hv_dynmem_device *dm)
{
struct dm_status status;
struct sysinfo val;
unsigned long now = jiffies;
unsigned long last_post = last_post_time;
if (pressure_report_delay > 0) {
--pressure_report_delay;
return;
}
if (!time_after(now, (last_post_time + HZ)))
return;
si_meminfo(&val);
memset(&status, 0, sizeof(struct dm_status));
status.hdr.type = DM_STATUS_REPORT;
status.hdr.size = sizeof(struct dm_status);
status.hdr.trans_id = atomic_inc_return(&trans_id);
/*
* The host expects the guest to report free and committed memory.
* Furthermore, the host expects the pressure information to include
* the ballooned out pages. For a given amount of memory that we are
* managing we need to compute a floor below which we should not
* balloon. Compute this and add it to the pressure report.
* We also need to report all offline pages (num_pages_added -
* num_pages_onlined) as committed to the host, otherwise it can try
* asking us to balloon them out.
*/
status.num_avail = val.freeram;
status.num_committed = vm_memory_committed() +
dm->num_pages_ballooned +
(dm->num_pages_added > dm->num_pages_onlined ?
dm->num_pages_added - dm->num_pages_onlined : 0) +
compute_balloon_floor();
/*
* If our transaction ID is no longer current, just don't
* send the status. This can happen if we were interrupted
* after we picked our transaction ID.
*/
if (status.hdr.trans_id != atomic_read(&trans_id))
return;
/*
* If the last post time that we sampled has changed,
* we have raced, don't post the status.
*/
if (last_post != last_post_time)
return;
last_post_time = jiffies;
vmbus_sendpacket(dm->dev->channel, &status,
sizeof(struct dm_status),
(unsigned long)NULL,
VM_PKT_DATA_INBAND, 0);
}
static void free_balloon_pages(struct hv_dynmem_device *dm,
union dm_mem_page_range *range_array)
{
int num_pages = range_array->finfo.page_cnt;
__u64 start_frame = range_array->finfo.start_page;
struct page *pg;
int i;
for (i = 0; i < num_pages; i++) {
pg = pfn_to_page(i + start_frame);
__free_page(pg);
dm->num_pages_ballooned--;
}
}
static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
unsigned int num_pages,
struct dm_balloon_response *bl_resp,
int alloc_unit)
{
unsigned int i = 0;
struct page *pg;
if (num_pages < alloc_unit)
return 0;
for (i = 0; (i * alloc_unit) < num_pages; i++) {
if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
PAGE_SIZE)
return i * alloc_unit;
/*
* We execute this code in a thread context. Furthermore,
* we don't want the kernel to try too hard.
*/
pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
__GFP_NOMEMALLOC | __GFP_NOWARN,
get_order(alloc_unit << PAGE_SHIFT));
Drivers: hv: hv_balloon: survive ballooning request with num_pages=0 ... and simplify alloc_balloon_pages() interface by removing redundant alloc_error from it. If we happen to enter balloon_up() with balloon_wrk.num_pages = 0 we will enter infinite 'while (!done)' loop as alloc_balloon_pages() will be always returning 0 and not setting alloc_error. We will also be sending a meaningless message to the host on every iteration. The 'alloc_unit == 1 && alloc_error -> num_ballooned == 0' change and alloc_error elimination requires a special comment. We do alloc_balloon_pages() with 2 different alloc_unit values and there are 4 different alloc_balloon_pages() results, let's check them all. alloc_unit = 512: 1) num_ballooned = 0, alloc_error = 0: we do 'alloc_unit=1' and retry pre- and post-patch. 2) num_ballooned > 0, alloc_error = 0: we check 'num_ballooned == num_pages' and act accordingly, pre- and post-patch. 3) num_ballooned > 0, alloc_error > 0: we report this chunk and remain within the loop, no changes here. 4) num_ballooned = 0, alloc_error > 0: we do 'alloc_unit=1' and retry pre- and post-patch. alloc_unit = 1: 1) num_ballooned = 0, alloc_error = 0: this can happen in two cases: when we passed 'num_pages=0' to alloc_balloon_pages() or when there was no space in bl_resp to place a single response. The second option is not possible as bl_resp is of PAGE_SIZE size and single response 'union dm_mem_page_range' is 8 bytes, but the first one is (in theory, I think that Hyper-V host never places such requests). Pre-patch code loops forever, post-patch code sends a reply with more_pages = 0 and finishes. 2) num_ballooned > 0, alloc_error = 0: we ran out of space in bl_resp, we report partial success and remain within the loop, no changes pre- and post-patch. 3) num_ballooned > 0, alloc_error > 0: pre-patch code finishes, post-patch code does one more try and if there is no progress (we finish with 'num_ballooned = 0') we finish. So we try a bit harder with this patch. 4) num_ballooned = 0, alloc_error > 0: both pre- and post-patch code enter 'more_pages = 0' branch and finish. So this patch has two real effects: 1) We reply with an empty response to 'num_pages=0' request. 2) We try a bit harder on alloc_unit=1 allocations (and reply with an empty tail reply in case we fail). An empty reply should be supported by host as we were able to send it even with pre-patch code when we were not able to allocate a single page. Suggested-by: Laszlo Ersek <lersek@redhat.com> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-03-27 16:10:13 +00:00
if (!pg)
return i * alloc_unit;
dm->num_pages_ballooned += alloc_unit;
/*
* If we allocatted 2M pages; split them so we
* can free them in any order we get.
*/
if (alloc_unit != 1)
split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
bl_resp->range_count++;
bl_resp->range_array[i].finfo.start_page =
page_to_pfn(pg);
bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
bl_resp->hdr.size += sizeof(union dm_mem_page_range);
}
return num_pages;
}
static void balloon_up(struct work_struct *dummy)
{
unsigned int num_pages = dm_device.balloon_wrk.num_pages;
unsigned int num_ballooned = 0;
struct dm_balloon_response *bl_resp;
int alloc_unit;
int ret;
bool done = false;
int i;
struct sysinfo val;
unsigned long floor;
/* The host balloons pages in 2M granularity. */
WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
/*
* We will attempt 2M allocations. However, if we fail to
* allocate 2M chunks, we will go back to 4k allocations.
*/
alloc_unit = 512;
si_meminfo(&val);
floor = compute_balloon_floor();
/* Refuse to balloon below the floor, keep the 2M granularity. */
if (val.freeram < num_pages || val.freeram - num_pages < floor) {
num_pages = val.freeram > floor ? (val.freeram - floor) : 0;
num_pages -= num_pages % PAGES_IN_2M;
}
while (!done) {
bl_resp = (struct dm_balloon_response *)send_buffer;
memset(send_buffer, 0, PAGE_SIZE);
bl_resp->hdr.type = DM_BALLOON_RESPONSE;
bl_resp->hdr.size = sizeof(struct dm_balloon_response);
bl_resp->more_pages = 1;
num_pages -= num_ballooned;
num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
Drivers: hv: hv_balloon: survive ballooning request with num_pages=0 ... and simplify alloc_balloon_pages() interface by removing redundant alloc_error from it. If we happen to enter balloon_up() with balloon_wrk.num_pages = 0 we will enter infinite 'while (!done)' loop as alloc_balloon_pages() will be always returning 0 and not setting alloc_error. We will also be sending a meaningless message to the host on every iteration. The 'alloc_unit == 1 && alloc_error -> num_ballooned == 0' change and alloc_error elimination requires a special comment. We do alloc_balloon_pages() with 2 different alloc_unit values and there are 4 different alloc_balloon_pages() results, let's check them all. alloc_unit = 512: 1) num_ballooned = 0, alloc_error = 0: we do 'alloc_unit=1' and retry pre- and post-patch. 2) num_ballooned > 0, alloc_error = 0: we check 'num_ballooned == num_pages' and act accordingly, pre- and post-patch. 3) num_ballooned > 0, alloc_error > 0: we report this chunk and remain within the loop, no changes here. 4) num_ballooned = 0, alloc_error > 0: we do 'alloc_unit=1' and retry pre- and post-patch. alloc_unit = 1: 1) num_ballooned = 0, alloc_error = 0: this can happen in two cases: when we passed 'num_pages=0' to alloc_balloon_pages() or when there was no space in bl_resp to place a single response. The second option is not possible as bl_resp is of PAGE_SIZE size and single response 'union dm_mem_page_range' is 8 bytes, but the first one is (in theory, I think that Hyper-V host never places such requests). Pre-patch code loops forever, post-patch code sends a reply with more_pages = 0 and finishes. 2) num_ballooned > 0, alloc_error = 0: we ran out of space in bl_resp, we report partial success and remain within the loop, no changes pre- and post-patch. 3) num_ballooned > 0, alloc_error > 0: pre-patch code finishes, post-patch code does one more try and if there is no progress (we finish with 'num_ballooned = 0') we finish. So we try a bit harder with this patch. 4) num_ballooned = 0, alloc_error > 0: both pre- and post-patch code enter 'more_pages = 0' branch and finish. So this patch has two real effects: 1) We reply with an empty response to 'num_pages=0' request. 2) We try a bit harder on alloc_unit=1 allocations (and reply with an empty tail reply in case we fail). An empty reply should be supported by host as we were able to send it even with pre-patch code when we were not able to allocate a single page. Suggested-by: Laszlo Ersek <lersek@redhat.com> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-03-27 16:10:13 +00:00
bl_resp, alloc_unit);
if (alloc_unit != 1 && num_ballooned == 0) {
alloc_unit = 1;
continue;
}
Drivers: hv: hv_balloon: survive ballooning request with num_pages=0 ... and simplify alloc_balloon_pages() interface by removing redundant alloc_error from it. If we happen to enter balloon_up() with balloon_wrk.num_pages = 0 we will enter infinite 'while (!done)' loop as alloc_balloon_pages() will be always returning 0 and not setting alloc_error. We will also be sending a meaningless message to the host on every iteration. The 'alloc_unit == 1 && alloc_error -> num_ballooned == 0' change and alloc_error elimination requires a special comment. We do alloc_balloon_pages() with 2 different alloc_unit values and there are 4 different alloc_balloon_pages() results, let's check them all. alloc_unit = 512: 1) num_ballooned = 0, alloc_error = 0: we do 'alloc_unit=1' and retry pre- and post-patch. 2) num_ballooned > 0, alloc_error = 0: we check 'num_ballooned == num_pages' and act accordingly, pre- and post-patch. 3) num_ballooned > 0, alloc_error > 0: we report this chunk and remain within the loop, no changes here. 4) num_ballooned = 0, alloc_error > 0: we do 'alloc_unit=1' and retry pre- and post-patch. alloc_unit = 1: 1) num_ballooned = 0, alloc_error = 0: this can happen in two cases: when we passed 'num_pages=0' to alloc_balloon_pages() or when there was no space in bl_resp to place a single response. The second option is not possible as bl_resp is of PAGE_SIZE size and single response 'union dm_mem_page_range' is 8 bytes, but the first one is (in theory, I think that Hyper-V host never places such requests). Pre-patch code loops forever, post-patch code sends a reply with more_pages = 0 and finishes. 2) num_ballooned > 0, alloc_error = 0: we ran out of space in bl_resp, we report partial success and remain within the loop, no changes pre- and post-patch. 3) num_ballooned > 0, alloc_error > 0: pre-patch code finishes, post-patch code does one more try and if there is no progress (we finish with 'num_ballooned = 0') we finish. So we try a bit harder with this patch. 4) num_ballooned = 0, alloc_error > 0: both pre- and post-patch code enter 'more_pages = 0' branch and finish. So this patch has two real effects: 1) We reply with an empty response to 'num_pages=0' request. 2) We try a bit harder on alloc_unit=1 allocations (and reply with an empty tail reply in case we fail). An empty reply should be supported by host as we were able to send it even with pre-patch code when we were not able to allocate a single page. Suggested-by: Laszlo Ersek <lersek@redhat.com> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2015-03-27 16:10:13 +00:00
if (num_ballooned == 0 || num_ballooned == num_pages) {
bl_resp->more_pages = 0;
done = true;
dm_device.state = DM_INITIALIZED;
}
/*
* We are pushing a lot of data through the channel;
* deal with transient failures caused because of the
* lack of space in the ring buffer.
*/
do {
bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
ret = vmbus_sendpacket(dm_device.dev->channel,
bl_resp,
bl_resp->hdr.size,
(unsigned long)NULL,
VM_PKT_DATA_INBAND, 0);
if (ret == -EAGAIN)
msleep(20);
post_status(&dm_device);
} while (ret == -EAGAIN);
if (ret) {
/*
* Free up the memory we allocatted.
*/
pr_info("Balloon response failed\n");
for (i = 0; i < bl_resp->range_count; i++)
free_balloon_pages(&dm_device,
&bl_resp->range_array[i]);
done = true;
}
}
}
static void balloon_down(struct hv_dynmem_device *dm,
struct dm_unballoon_request *req)
{
union dm_mem_page_range *range_array = req->range_array;
int range_count = req->range_count;
struct dm_unballoon_response resp;
int i;
for (i = 0; i < range_count; i++) {
free_balloon_pages(dm, &range_array[i]);
complete(&dm_device.config_event);
}
if (req->more_pages == 1)
return;
memset(&resp, 0, sizeof(struct dm_unballoon_response));
resp.hdr.type = DM_UNBALLOON_RESPONSE;
resp.hdr.trans_id = atomic_inc_return(&trans_id);
resp.hdr.size = sizeof(struct dm_unballoon_response);
vmbus_sendpacket(dm_device.dev->channel, &resp,
sizeof(struct dm_unballoon_response),
(unsigned long)NULL,
VM_PKT_DATA_INBAND, 0);
dm->state = DM_INITIALIZED;
}
static void balloon_onchannelcallback(void *context);
static int dm_thread_func(void *dm_dev)
{
struct hv_dynmem_device *dm = dm_dev;
while (!kthread_should_stop()) {
wait_for_completion_interruptible_timeout(
&dm_device.config_event, 1*HZ);
/*
* The host expects us to post information on the memory
* pressure every second.
*/
reinit_completion(&dm_device.config_event);
post_status(dm);
}
return 0;
}
static void version_resp(struct hv_dynmem_device *dm,
struct dm_version_response *vresp)
{
struct dm_version_request version_req;
int ret;
if (vresp->is_accepted) {
/*
* We are done; wakeup the
* context waiting for version
* negotiation.
*/
complete(&dm->host_event);
return;
}
/*
* If there are more versions to try, continue
* with negotiations; if not
* shutdown the service since we are not able
* to negotiate a suitable version number
* with the host.
*/
if (dm->next_version == 0)
goto version_error;
memset(&version_req, 0, sizeof(struct dm_version_request));
version_req.hdr.type = DM_VERSION_REQUEST;
version_req.hdr.size = sizeof(struct dm_version_request);
version_req.hdr.trans_id = atomic_inc_return(&trans_id);
version_req.version.version = dm->next_version;
/*
* Set the next version to try in case current version fails.
* Win7 protocol ought to be the last one to try.
*/
switch (version_req.version.version) {
case DYNMEM_PROTOCOL_VERSION_WIN8:
dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
version_req.is_last_attempt = 0;
break;
default:
dm->next_version = 0;
version_req.is_last_attempt = 1;
}
ret = vmbus_sendpacket(dm->dev->channel, &version_req,
sizeof(struct dm_version_request),
(unsigned long)NULL,
VM_PKT_DATA_INBAND, 0);
if (ret)
goto version_error;
return;
version_error:
dm->state = DM_INIT_ERROR;
complete(&dm->host_event);
}
static void cap_resp(struct hv_dynmem_device *dm,
struct dm_capabilities_resp_msg *cap_resp)
{
if (!cap_resp->is_accepted) {
pr_info("Capabilities not accepted by host\n");
dm->state = DM_INIT_ERROR;
}
complete(&dm->host_event);
}
static void balloon_onchannelcallback(void *context)
{
struct hv_device *dev = context;
u32 recvlen;
u64 requestid;
struct dm_message *dm_msg;
struct dm_header *dm_hdr;
struct hv_dynmem_device *dm = hv_get_drvdata(dev);
struct dm_balloon *bal_msg;
struct dm_hot_add *ha_msg;
union dm_mem_page_range *ha_pg_range;
union dm_mem_page_range *ha_region;
memset(recv_buffer, 0, sizeof(recv_buffer));
vmbus_recvpacket(dev->channel, recv_buffer,
PAGE_SIZE, &recvlen, &requestid);
if (recvlen > 0) {
dm_msg = (struct dm_message *)recv_buffer;
dm_hdr = &dm_msg->hdr;
switch (dm_hdr->type) {
case DM_VERSION_RESPONSE:
version_resp(dm,
(struct dm_version_response *)dm_msg);
break;
case DM_CAPABILITIES_RESPONSE:
cap_resp(dm,
(struct dm_capabilities_resp_msg *)dm_msg);
break;
case DM_BALLOON_REQUEST:
if (dm->state == DM_BALLOON_UP)
pr_warn("Currently ballooning\n");
bal_msg = (struct dm_balloon *)recv_buffer;
dm->state = DM_BALLOON_UP;
dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
schedule_work(&dm_device.balloon_wrk.wrk);
break;
case DM_UNBALLOON_REQUEST:
dm->state = DM_BALLOON_DOWN;
balloon_down(dm,
(struct dm_unballoon_request *)recv_buffer);
break;
case DM_MEM_HOT_ADD_REQUEST:
if (dm->state == DM_HOT_ADD)
pr_warn("Currently hot-adding\n");
dm->state = DM_HOT_ADD;
ha_msg = (struct dm_hot_add *)recv_buffer;
if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
/*
* This is a normal hot-add request specifying
* hot-add memory.
*/
ha_pg_range = &ha_msg->range;
dm->ha_wrk.ha_page_range = *ha_pg_range;
dm->ha_wrk.ha_region_range.page_range = 0;
} else {
/*
* Host is specifying that we first hot-add
* a region and then partially populate this
* region.
*/
dm->host_specified_ha_region = true;
ha_pg_range = &ha_msg->range;
ha_region = &ha_pg_range[1];
dm->ha_wrk.ha_page_range = *ha_pg_range;
dm->ha_wrk.ha_region_range = *ha_region;
}
schedule_work(&dm_device.ha_wrk.wrk);
break;
case DM_INFO_MESSAGE:
process_info(dm, (struct dm_info_msg *)dm_msg);
break;
default:
pr_err("Unhandled message: type: %d\n", dm_hdr->type);
}
}
}
static int balloon_probe(struct hv_device *dev,
const struct hv_vmbus_device_id *dev_id)
{
int ret;
unsigned long t;
struct dm_version_request version_req;
struct dm_capabilities cap_msg;
do_hot_add = hot_add;
/*
* First allocate a send buffer.
*/
send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
if (!send_buffer)
return -ENOMEM;
ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
balloon_onchannelcallback, dev);
if (ret)
goto probe_error0;
dm_device.dev = dev;
dm_device.state = DM_INITIALIZING;
dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
init_completion(&dm_device.host_event);
init_completion(&dm_device.config_event);
INIT_LIST_HEAD(&dm_device.ha_region_list);
mutex_init(&dm_device.ha_region_mutex);
INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
dm_device.host_specified_ha_region = false;
dm_device.thread =
kthread_run(dm_thread_func, &dm_device, "hv_balloon");
if (IS_ERR(dm_device.thread)) {
ret = PTR_ERR(dm_device.thread);
goto probe_error1;
}
#ifdef CONFIG_MEMORY_HOTPLUG
set_online_page_callback(&hv_online_page);
register_memory_notifier(&hv_memory_nb);
#endif
hv_set_drvdata(dev, &dm_device);
/*
* Initiate the hand shake with the host and negotiate
* a version that the host can support. We start with the
* highest version number and go down if the host cannot
* support it.
*/
memset(&version_req, 0, sizeof(struct dm_version_request));
version_req.hdr.type = DM_VERSION_REQUEST;
version_req.hdr.size = sizeof(struct dm_version_request);
version_req.hdr.trans_id = atomic_inc_return(&trans_id);
version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
version_req.is_last_attempt = 0;
ret = vmbus_sendpacket(dev->channel, &version_req,
sizeof(struct dm_version_request),
(unsigned long)NULL,
VM_PKT_DATA_INBAND, 0);
if (ret)
goto probe_error2;
t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
if (t == 0) {
ret = -ETIMEDOUT;
goto probe_error2;
}
/*
* If we could not negotiate a compatible version with the host
* fail the probe function.
*/
if (dm_device.state == DM_INIT_ERROR) {
ret = -ETIMEDOUT;
goto probe_error2;
}
/*
* Now submit our capabilities to the host.
*/
memset(&cap_msg, 0, sizeof(struct dm_capabilities));
cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
cap_msg.hdr.size = sizeof(struct dm_capabilities);
cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
cap_msg.caps.cap_bits.balloon = 1;
cap_msg.caps.cap_bits.hot_add = 1;
/*
* Specify our alignment requirements as it relates
* memory hot-add. Specify 128MB alignment.
*/
cap_msg.caps.cap_bits.hot_add_alignment = 7;
/*
* Currently the host does not use these
* values and we set them to what is done in the
* Windows driver.
*/
cap_msg.min_page_cnt = 0;
cap_msg.max_page_number = -1;
ret = vmbus_sendpacket(dev->channel, &cap_msg,
sizeof(struct dm_capabilities),
(unsigned long)NULL,
VM_PKT_DATA_INBAND, 0);
if (ret)
goto probe_error2;
t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
if (t == 0) {
ret = -ETIMEDOUT;
goto probe_error2;
}
/*
* If the host does not like our capabilities,
* fail the probe function.
*/
if (dm_device.state == DM_INIT_ERROR) {
ret = -ETIMEDOUT;
goto probe_error2;
}
dm_device.state = DM_INITIALIZED;
return 0;
probe_error2:
#ifdef CONFIG_MEMORY_HOTPLUG
restore_online_page_callback(&hv_online_page);
#endif
kthread_stop(dm_device.thread);
probe_error1:
vmbus_close(dev->channel);
probe_error0:
kfree(send_buffer);
return ret;
}
static int balloon_remove(struct hv_device *dev)
{
struct hv_dynmem_device *dm = hv_get_drvdata(dev);
struct list_head *cur, *tmp;
struct hv_hotadd_state *has;
if (dm->num_pages_ballooned != 0)
pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
cancel_work_sync(&dm->balloon_wrk.wrk);
cancel_work_sync(&dm->ha_wrk.wrk);
vmbus_close(dev->channel);
kthread_stop(dm->thread);
kfree(send_buffer);
#ifdef CONFIG_MEMORY_HOTPLUG
restore_online_page_callback(&hv_online_page);
unregister_memory_notifier(&hv_memory_nb);
#endif
list_for_each_safe(cur, tmp, &dm->ha_region_list) {
has = list_entry(cur, struct hv_hotadd_state, list);
list_del(&has->list);
kfree(has);
}
return 0;
}
static const struct hv_vmbus_device_id id_table[] = {
/* Dynamic Memory Class ID */
/* 525074DC-8985-46e2-8057-A307DC18A502 */
{ HV_DM_GUID, },
{ },
};
MODULE_DEVICE_TABLE(vmbus, id_table);
static struct hv_driver balloon_drv = {
.name = "hv_balloon",
.id_table = id_table,
.probe = balloon_probe,
.remove = balloon_remove,
};
static int __init init_balloon_drv(void)
{
return vmbus_driver_register(&balloon_drv);
}
module_init(init_balloon_drv);
MODULE_DESCRIPTION("Hyper-V Balloon");
MODULE_LICENSE("GPL");