migration/postcopy: Sync faulted addresses after network recovered
v2 cover migration/postcopy: Sync faulted addresses after network recovered
v2:
- add r-bs for Dave
- add patch "migration: Properly destroy variables on incoming side" as patch 1
- destroy page_request_mutex in migration_incoming_state_destroy() too [Dave]
- use WITH_QEMU_LOCK_GUARD in two places where we can [Dave]
We've seen conditional guest hangs on destination VM after postcopy recovered. However the hang will resolve itself after a few minutes.
The problem is: after a postcopy recovery, the prioritized postcopy queue on the source VM is actually missing. So all the faulted threads before the postcopy recovery happened will keep halted until (accidentally) the page got copied by the background precopy migration stream.
The solution is to also refresh this information after postcopy recovery. To achieve this, we need to maintain a list of faulted addresses on the destination node, so that we can resend the list when necessary. This work is done via patch 2-5.
With that, the last thing we need to do is to send this extra information to source VM after recovered. Very luckily, this synchronization can be "emulated" by sending a bunch of page requests (although these pages have been sent previously!) to source VM just like when we've got a page fault. Even in the 1st version of the postcopy code we'll handle duplicated pages well. So this fix does not even need a new capability bit and it'll work smoothly on old QEMUs when we migrate from them to the new QEMUs.
Please review, thanks.
Peter Xu (6):
migration: Properly destroy variables on incoming side
migration: Rework migrate_send_rp_req_pages() function
migration: Pass incoming state into qemu_ufd_copy_ioctl()
migration: Introduce migrate_send_rp_message_req_pages()
migration: Maintain postcopy faulted addresses
migration: Sync requested pages after postcopy recovery
migration/migration.c | 79 +++++++++++++++++++++++++++++++++++----- migration/migration.h | 23 +++++++++++- migration/postcopy-ram.c | 46 ++++++++++------------- migration/savevm.c | 57 +++++++++++++++++++++++++++++ migration/trace-events | 3 ++ 5 files changed, 171 insertions(+), 37 deletions(-)
Signature · 1 line
2.26.2
PATCH v2 1/6 migration: Properly destroy variables on incoming side
In migration_incoming_state_destroy(), we've got a few variables that aren't destroyed properly, namely:
main_thread_load_event
postcopy_pause_sem_dst
postcopy_pause_sem_fault
rp_mutex
Destroy them properly.
--- migration/migration.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)
migration/migration.c +5 -2
@@ -238,12 +238,15 @@ void migration_incoming_state_destroy(void) mis->postcopy_remote_fds =NULL ; } -qemu_event_reset (&mis->main_thread_load_event); -if (mis->socket_address_list) {qapi_free_SocketAddressList (mis->socket_address_list); mis->socket_address_list =NULL ; } + +qemu_event_destroy (&mis->main_thread_load_event); +qemu_sem_destroy (&mis->postcopy_pause_sem_dst); +qemu_sem_destroy (&mis->postcopy_pause_sem_fault); +qemu_mutex_destroy (&mis->rp_mutex); }static void migrate_generate_event (int new_state)
Signature · 1 line
2.26.2
Re: [PATCH v2 1/6] migration: Properly destroy variables on incoming side
Quoted 11 lines · * Peter Xu ([email protected]) wrote:
In
migration_incoming_state_destroy(), we've got a few variables that aren't destroyed properly, namely:
main_thread_load_event
postcopy_pause_sem_dst
postcopy_pause_sem_fault
rp_mutexDestroy them properly.
Signed-off-by Peter Xu <[email protected]>
Quoted 29 lines
--- migration/migration.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)migration/migration.c +5
-2@@ -238,12 +238,15 @@ void migration_incoming_state_destroy(void) mis->postcopy_remote_fds =NULL ; } -qemu_event_reset (&mis->main_thread_load_event); -if (mis->socket_address_list) {qapi_free_SocketAddressList (mis->socket_address_list); mis->socket_address_list =NULL ; } + +qemu_event_destroy (&mis->main_thread_load_event); +qemu_sem_destroy (&mis->postcopy_pause_sem_dst); +qemu_sem_destroy (&mis->postcopy_pause_sem_fault); +qemu_mutex_destroy (&mis->rp_mutex); }static void migrate_generate_event (int new_state)Signature · 1 line
2.26.2
Signature · 1 line
Dr. David Alan Gilbert / [email protected] / Manchester, UK
PATCH v2 2/6 migration: Rework migrate_send_rp_req_pages() function
In reply to Peter XuWe duplicated the logic of maintaining the last_rb variable at both callers of this function. Pass *rb pointer into the function so that we can avoid duplicating the logic. Also, when we have the rb pointer, it's also easier to remove the original 2nd & 4th parameters, because both of them (name of the ramblock when needed, or the page size) can be fetched from the ramblock pointer too.
--- migration/migration.c | 26 ++++++++++++++++++-------- migration/migration.h | 4 ++-- migration/postcopy-ram.c | 24 ++---------------------- 3 files changed, 22 insertions(+), 32 deletions(-)
migration/migration.c +18 -8
@@ -314,25 +314,35 @@ error:return ret; } -/* Request a range of pages from the source VM at the given -* start address. -* rbname: Name of the RAMBlock to request the page in, if NULL it's the same -* as the last request (a name must have been given previously) +/* Request one page from the source VM at the given start address. +* rb: the RAMBlock to request the page in * Start: Address offset within the RB * Len: Length in bytes required - must be a multiple of pagesize */ -int migrate_send_rp_req_pages (MigrationIncomingState *mis,const char *rbname, -ram_addr_t start,size_t len) +int migrate_send_rp_req_pages (MigrationIncomingState *mis,RAMBlock *rb, +ram_addr_t start) {uint8_t bufc[12 +1 +255 ];/* start (8), len (4), rbname up to 256 */ size_t msglen =12 ;/* start + len */ +size_t len =qemu_ram_pagesize (rb);enum mig_rp_message_type msg_type; +const char *rbname; +int rbname_len; *(uint64_t *)bufc =cpu_to_be64 ((uint64_t )start); *(uint32_t *)(bufc +8 ) =cpu_to_be32 ((uint32_t )len); -if (rbname) { -int rbname_len =strlen (rbname); +/* +* We maintain the last ramblock that we requested for page. Note that we +* don't need locking because this function will only be called within the +* postcopy ram fault thread. +*/ +if (rb != mis->last_rb) { + mis->last_rb = rb; + + rbname =qemu_ram_get_idstr (rb); + rbname_len =strlen (rbname); +assert (rbname_len <256 ); bufc[msglen++] = rbname_len;
migration/migration.h +2 -2
@@ -328,8 +328,8 @@ void migrate_send_rp_shut(MigrationIncomingState *mis, uint32_t value);void migrate_send_rp_pong (MigrationIncomingState *mis,uint32_t value); -int migrate_send_rp_req_pages (MigrationIncomingState *mis,const char * rbname, -ram_addr_t start,size_t len); +int migrate_send_rp_req_pages (MigrationIncomingState *mis,RAMBlock *rb, +ram_addr_t start);void migrate_send_rp_recv_bitmap (MigrationIncomingState *mis,char *block_name);void migrate_send_rp_resume_ack (MigrationIncomingState *mis,uint32_t value);
migration/postcopy-ram.c +2 -22
@@ -684,14 +684,7 @@ int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, qemu_ram_get_idstr(rb), rb_offset);return postcopy_wake_shared (pcfd, client_addr, rb); } -if (rb != mis->last_rb) { - mis->last_rb = rb; -migrate_send_rp_req_pages (mis,qemu_ram_get_idstr (rb), - aligned_rbo, pagesize); - }else { -/* Save some space */ -migrate_send_rp_req_pages (mis,NULL , aligned_rbo, pagesize); - } +migrate_send_rp_req_pages (mis, rb, aligned_rbo);return 0 ; }
@@ -986,20 +979,7 @@ retry:* Send the request to the source - we want to request one * of our host page sizes (which is >= TPS) */ -if (rb != mis->last_rb) { - mis->last_rb = rb; - ret =migrate_send_rp_req_pages (mis, -qemu_ram_get_idstr (rb), - rb_offset, -qemu_ram_pagesize (rb)); - }else { -/* Save some space */ - ret =migrate_send_rp_req_pages (mis, -NULL , - rb_offset, -qemu_ram_pagesize (rb)); - } - + ret =migrate_send_rp_req_pages (mis, rb, rb_offset);if (ret) {/* May be network failure, try to wait for recovery */ if (ret == -EIO &&postcopy_pause_fault_thread (mis)) {
Signature · 1 line
2.26.2
PATCH v2 3/6 migration: Pass incoming state into qemu_ufd_copy_ioctl()
In reply to Peter XuIt'll be used in follow up patches to access more fields out of it. Meanwhile fetch the userfaultfd inside the function.
--- migration/postcopy-ram.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
migration/postcopy-ram.c +5 -3
@@ -1128,10 +1128,12 @@ int postcopy_ram_incoming_setup(MigrationIncomingState *mis)return 0 ; } -static int qemu_ufd_copy_ioctl (int userfault_fd,void *host_addr, +static int qemu_ufd_copy_ioctl (MigrationIncomingState *mis,void *host_addr,void *from_addr,uint64_t pagesize,RAMBlock *rb) { +int userfault_fd = mis->userfault_fd;int ret; +if (from_addr) {struct uffdio_copy copy_struct; copy_struct.dst = (uint64_t )(uintptr_t )host_addr;
@@ -1185,7 +1187,7 @@ int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,* which would be slightly cheaper, but we'd have to be careful * of the order of updating our page state. */ -if (qemu_ufd_copy_ioctl (mis->userfault_fd, host, from, pagesize, rb)) { +if (qemu_ufd_copy_ioctl (mis, host, from, pagesize, rb)) {int e = errno;error_report ("%s: %s copy host: %p from: %p (size: %zd)" , __func__,strerror (e), host, from, pagesize);
@@ -1212,7 +1214,7 @@ int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,* but it's not available for everything (e.g. hugetlbpages) */ if (qemu_ram_is_uf_zeroable (rb)) { -if (qemu_ufd_copy_ioctl (mis->userfault_fd, host,NULL , pagesize, rb)) { +if (qemu_ufd_copy_ioctl (mis, host,NULL , pagesize, rb)) {int e = errno;error_report ("%s: %s zero host: %p" , __func__,strerror (e), host);
Signature · 1 line
2.26.2
PATCH v2 4/6 migration: Introduce migrate_send_rp_message_req_pages()
In reply to Peter XuThis is another layer wrapper for sending a page request to the source VM. The new migrate_send_rp_message_req_pages() will be used elsewhere in coming patches.
--- migration/migration.c | 10 ++++++++-- migration/migration.h | 2 ++ 2 files changed, 10 insertions(+), 2 deletions(-)
migration/migration.c +8 -2
@@ -319,8 +319,8 @@ error:* Start: Address offset within the RB * Len: Length in bytes required - must be a multiple of pagesize */ -int migrate_send_rp_req_pages (MigrationIncomingState *mis,RAMBlock *rb, -ram_addr_t start) +int migrate_send_rp_message_req_pages (MigrationIncomingState *mis, +RAMBlock *rb,ram_addr_t start) {uint8_t bufc[12 +1 +255 ];/* start (8), len (4), rbname up to 256 */ size_t msglen =12 ;/* start + len */
@@ -356,6 +356,12 @@ int migrate_send_rp_req_pages(MigrationIncomingState *mis, RAMBlock *rb,return migrate_send_rp_message (mis, msg_type, msglen, bufc); } +int migrate_send_rp_req_pages (MigrationIncomingState *mis, +RAMBlock *rb,ram_addr_t start) +{ +return migrate_send_rp_message_req_pages (mis, rb, start); +} +static bool migration_colo_enabled;bool migration_incoming_colo_enabled (void ) {
migration/migration.h +2 -0
@@ -330,6 +330,8 @@ void migrate_send_rp_pong(MigrationIncomingState *mis, uint32_t value);int migrate_send_rp_req_pages (MigrationIncomingState *mis,RAMBlock *rb,ram_addr_t start); +int migrate_send_rp_message_req_pages (MigrationIncomingState *mis, +RAMBlock *rb,ram_addr_t start);void migrate_send_rp_recv_bitmap (MigrationIncomingState *mis,char *block_name);void migrate_send_rp_resume_ack (MigrationIncomingState *mis,uint32_t value);
Signature · 1 line
2.26.2
PATCH v2 5/6 migration: Maintain postcopy faulted addresses
In reply to Peter XuMaintain a list of faulted addresses on the destination host for which we're waiting on. This is implemented using a GTree rather than a real list to make sure even there're plenty of vCPUs/threads that are faulting, the lookup will still be fast with O(log(N)) (because we'll do that after placing each page). It should bring a slight overhead, but ideally that shouldn't be a big problem simply because in most cases the requested page list will be short.
Actually we did similar things for postcopy blocktime measurements. This patch didn't use that simply because:
(1) blocktime measurement is towards vcpu threads only, but here we need to record all faulted addresses, including main thread and external thread (like, DPDK via vhost-user).
(2) blocktime measurement will require UFFD_FEATURE_THREAD_ID, but here we don't want to add that extra dependency on the kernel version since not necessary. E.g., we don't need to know which thread faulted on which page, we also don't care about multiple threads faulting on the same page. But we only care about what addresses are faulted so waiting for a page copying from src.
(3) blocktime measurement is not enabled by default. However we need this by default especially for postcopy recover.
Another thing to mention is that this patch introduced a new mutex to serialize the receivedmap and the page_requested tree, however that serialization does not cover other procedures like UFFDIO_COPY.
--- migration/migration.c | 44 +++++++++++++++++++++++++++++++++++++++- migration/migration.h | 19 ++++++++++++++++- migration/postcopy-ram.c | 18 +++++++++++++--- migration/trace-events | 2 ++ 4 files changed, 78 insertions(+), 5 deletions(-)
migration/migration.c +43 -1
@@ -143,6 +143,13 @@ static int migration_maybe_pause(MigrationState *s, int new_state);static void migrate_fd_cancel (MigrationState *s); +static gint page_request_addr_cmp (gconstpointer ap,gconstpointer bp) +{ +uint64_t a = (uint64_t ) ap, b = (uint64_t ) bp; + +return (a > b) - (a < b); +} +void migration_object_init (void ) {MachineState *ms =MACHINE (qdev_get_machine ());
@@ -165,6 +172,8 @@ void migration_object_init(void)qemu_event_init (¤t_incoming->main_thread_load_event, false);qemu_sem_init (¤t_incoming->postcopy_pause_sem_dst,0 );qemu_sem_init (¤t_incoming->postcopy_pause_sem_fault,0 ); +qemu_mutex_init (¤t_incoming->page_request_mutex); + current_incoming->page_requested =g_tree_new (page_request_addr_cmp);if (!migration_object_check (current_migration, &err)) {error_report_err (err);
@@ -238,6 +247,11 @@ void migration_incoming_state_destroy(void) mis->postcopy_remote_fds =NULL ; } +if (mis->page_requested) { +g_tree_destroy (mis->page_requested); + mis->page_requested =NULL ; + } +if (mis->socket_address_list) {qapi_free_SocketAddressList (mis->socket_address_list); mis->socket_address_list =NULL ;
@@ -247,6 +261,7 @@ void migration_incoming_state_destroy(void)qemu_sem_destroy (&mis->postcopy_pause_sem_dst);qemu_sem_destroy (&mis->postcopy_pause_sem_fault);qemu_mutex_destroy (&mis->rp_mutex); +qemu_mutex_destroy (&mis->page_request_mutex); }static void migrate_generate_event (int new_state)
@@ -357,8 +372,35 @@ int migrate_send_rp_message_req_pages(MigrationIncomingState *mis, }int migrate_send_rp_req_pages (MigrationIncomingState *mis, -RAMBlock *rb,ram_addr_t start) +RAMBlock *rb,ram_addr_t start,uint64_t haddr) { +uint64_t aligned = haddr & (-qemu_target_page_size ()); +bool received; + + WITH_QEMU_LOCK_GUARD(&mis->page_request_mutex) { + received =ramblock_recv_bitmap_test_byte_offset (rb, start); +if (!received && !g_tree_lookup (mis->page_requested, + (gpointer ) aligned)) { +/* +* The page has not been received, and it's not yet in the page +* request list. Queue it. Set the value of element to 1, so that +* things like g_tree_lookup() will return TRUE (1) when found. +*/ +g_tree_insert (mis->page_requested, (gpointer ) aligned, + (gpointer )1 ); + mis->page_requested_count++; +trace_postcopy_page_req_add (aligned, mis->page_requested_count); + } + } + +/* +* If the page is there, skip sending the message. We don't even need the +* lock because as long as the page arrived, it'll be there forever. +*/ +if (received) { +return 0 ; + } +return migrate_send_rp_message_req_pages (mis, rb, start); }
migration/migration.h +18 -1
@@ -103,6 +103,23 @@ struct MigrationIncomingState {/* List of listening socket addresses */ SocketAddressList *socket_address_list; + +/* A tree of pages that we requested to the source VM */ +GTree *page_requested; +/* For debugging purpose only, but would be nice to keep */ +int page_requested_count; +/* +* The mutex helps to maintain the requested pages that we sent to the +* source, IOW, to guarantee coherent between the page_requests tree and +* the per-ramblock receivedmap. Note! This does not guarantee consistency +* of the real page copy procedures (using UFFDIO_[ZERO]COPY). E.g., even +* if one bit in receivedmap is cleared, UFFDIO_COPY could have happened +* for that page already. This is intended so that the mutex won't +* serialize and blocked by slow operations like UFFDIO_* ioctls. However +* this should be enough to make sure the page_requested tree always +* contains valid information. +*/ +QemuMutex page_request_mutex; };MigrationIncomingState *migration_incoming_get_current (void );
@@ -329,7 +346,7 @@ void migrate_send_rp_shut(MigrationIncomingState *mis,void migrate_send_rp_pong (MigrationIncomingState *mis,uint32_t value);int migrate_send_rp_req_pages (MigrationIncomingState *mis,RAMBlock *rb, -ram_addr_t start); +ram_addr_t start,uint64_t haddr);int migrate_send_rp_message_req_pages (MigrationIncomingState *mis,RAMBlock *rb,ram_addr_t start); void migrate_send_rp_recv_bitmap(MigrationIncomingState *mis,
migration/postcopy-ram.c +15 -3
@@ -684,7 +684,7 @@ int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, qemu_ram_get_idstr(rb), rb_offset);return postcopy_wake_shared (pcfd, client_addr, rb); } -migrate_send_rp_req_pages (mis, rb, aligned_rbo); +migrate_send_rp_req_pages (mis, rb, aligned_rbo, client_addr);return 0 ; }
@@ -979,7 +979,8 @@ retry:* Send the request to the source - we want to request one * of our host page sizes (which is >= TPS) */ - ret =migrate_send_rp_req_pages (mis, rb, rb_offset); + ret =migrate_send_rp_req_pages (mis, rb, rb_offset, + msg.arg.pagefault.address);if (ret) {/* May be network failure, try to wait for recovery */ if (ret == -EIO &&postcopy_pause_fault_thread (mis)) {
@@ -1149,10 +1150,21 @@ static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr, ret =ioctl (userfault_fd,UFFDIO_ZEROPAGE , &zero_struct); }if (!ret) { +qemu_mutex_lock (&mis->page_request_mutex);ramblock_recv_bitmap_set_range (rb, host_addr, pagesize /qemu_target_page_size ()); +/* +* If this page resolves a page fault for a previous recorded faulted +* address, take a special note to maintain the requested page list. +*/ +if (g_tree_lookup (mis->page_requested, (gconstpointer )host_addr)) { +g_tree_remove (mis->page_requested, (gconstpointer )host_addr); + mis->page_requested_count--; +trace_postcopy_page_req_del ((uint64_t )host_addr, + mis->page_requested_count); + } +qemu_mutex_unlock (&mis->page_request_mutex);mark_postcopy_blocktime_end ((uintptr_t )host_addr); - }return ret; }
migration/trace-events +2 -0
@@ -157,6 +157,7 @@ postcopy_pause_return_path(void) "" postcopy_pause_return_path_continued(void) "" postcopy_pause_continued(void) "" postcopy_start_set_run(void) "" +postcopy_page_req_add(uint64_t addr, int count) "new page req 0x%lx total %d" source_return_path_thread_bad_end(void) "" source_return_path_thread_end(void) "" source_return_path_thread_entry(void) ""
@@ -267,6 +268,7 @@ postcopy_ram_incoming_cleanup_blocktime(uint64_t total) "total blocktime %" PRIu postcopy_request_shared_page(const char *sharer, const char *rb, uint64_t rb_offset) "for %s in %s offset 0x%"PRIx64 postcopy_request_shared_page_present(const char *sharer, const char *rb, uint64_t rb_offset) "%s already %s offset 0x%"PRIx64 postcopy_wake_shared(uint64_t client_addr, const char *rb) "at 0x%"PRIx64" in %s" +postcopy_page_req_del(uint64_t addr, int count) "resolved page req 0x%lx total %d" get_mem_fault_cpu_index(int cpu, uint32_t pid) "cpu: %d, pid: %u"
Signature · 1 line
2.26.2
Re: [PATCH v2 5/6] migration: Maintain postcopy faulted addresses
Quoted 21 lines · * Peter Xu ([email protected]) wrote:
Maintain a list of faulted addresses on the destination host for which we're waiting on. This is implemented using a GTree rather than a real list to make sure even there're plenty of vCPUs/threads that are faulting, the lookup will still be fast with O(log(N)) (because we'll do that after placing each page). It should bring a slight overhead, but ideally that shouldn't be a big problem simply because in most cases the requested page list will be short.
Actually we did similar things for postcopy blocktime measurements. This patch didn't use that simply because:
(1) blocktime measurement is towards vcpu threads only, but here we need to record all faulted addresses, including main thread and external thread (like, DPDK via vhost-user).
(2) blocktime measurement will require
UFFD_FEATURE_THREAD_ID, but here we don't want to add that extra dependency on the kernel version since not necessary. E.g., we don't need to know which thread faulted on which page, we also don't care about multiple threads faulting on the same page. But we only care about what addresses are faulted so waiting for a page copying from src.
Quoted 8 lines
(3) blocktime measurement is not enabled by default. However we need this by default especially for postcopy recover.
Another thing to mention is that this patch introduced a new mutex to serialize the receivedmap and the
page_requestedtree, however that serialization does not cover other procedures likeUFFDIO_COPY.Signed-off-by Peter Xu <[email protected]>
Quoted 197 lines
--- migration/migration.c | 44 +++++++++++++++++++++++++++++++++++++++- migration/migration.h | 19 ++++++++++++++++- migration/postcopy-ram.c | 18 +++++++++++++--- migration/trace-events | 2 ++ 4 files changed, 78 insertions(+), 5 deletions(-)migration/migration.c +43
-1@@ -143,6 +143,13 @@ static int migration_maybe_pause(MigrationState *s, int new_state);static void migrate_fd_cancel (MigrationState *s); +static gint page_request_addr_cmp (gconstpointer ap,gconstpointer bp) +{ +uint64_t a = (uint64_t ) ap, b = (uint64_t ) bp; + +return (a > b) - (a < b); +} +void migration_object_init (void ) {MachineState *ms =MACHINE (qdev_get_machine ());@@ -165,6 +172,8 @@ void migration_object_init(void)qemu_event_init (¤t_incoming->main_thread_load_event, false);qemu_sem_init (¤t_incoming->postcopy_pause_sem_dst,0 );qemu_sem_init (¤t_incoming->postcopy_pause_sem_fault,0 ); +qemu_mutex_init (¤t_incoming->page_request_mutex); + current_incoming->page_requested =g_tree_new (page_request_addr_cmp);if (!migration_object_check (current_migration, &err)) {error_report_err (err);@@ -238,6 +247,11 @@ void migration_incoming_state_destroy(void) mis->postcopy_remote_fds =NULL ; } +if (mis->page_requested) { +g_tree_destroy (mis->page_requested); + mis->page_requested =NULL ; + } +if (mis->socket_address_list) {qapi_free_SocketAddressList (mis->socket_address_list); mis->socket_address_list =NULL ;@@ -247,6 +261,7 @@ void migration_incoming_state_destroy(void)qemu_sem_destroy (&mis->postcopy_pause_sem_dst);qemu_sem_destroy (&mis->postcopy_pause_sem_fault);qemu_mutex_destroy (&mis->rp_mutex); +qemu_mutex_destroy (&mis->page_request_mutex); }static void migrate_generate_event (int new_state)@@ -357,8 +372,35 @@ int migrate_send_rp_message_req_pages(MigrationIncomingState *mis, }int migrate_send_rp_req_pages (MigrationIncomingState *mis, -RAMBlock *rb,ram_addr_t start) +RAMBlock *rb,ram_addr_t start,uint64_t haddr) { +uint64_t aligned = haddr & (-qemu_target_page_size ()); +bool received; + + WITH_QEMU_LOCK_GUARD(&mis->page_request_mutex) { + received =ramblock_recv_bitmap_test_byte_offset (rb, start); +if (!received && !g_tree_lookup (mis->page_requested, + (gpointer ) aligned)) { +/* +* The page has not been received, and it's not yet in the page +* request list. Queue it. Set the value of element to 1, so that +* things like g_tree_lookup() will return TRUE (1) when found. +*/ +g_tree_insert (mis->page_requested, (gpointer ) aligned, + (gpointer )1 ); + mis->page_requested_count++; +trace_postcopy_page_req_add (aligned, mis->page_requested_count); + } + } + +/* +* If the page is there, skip sending the message. We don't even need the +* lock because as long as the page arrived, it'll be there forever. +*/ +if (received) { +return 0 ; + } +return migrate_send_rp_message_req_pages (mis, rb, start); }migration/migration.h +18
-1@@ -103,6 +103,23 @@ struct MigrationIncomingState {/* List of listening socket addresses */ SocketAddressList *socket_address_list; + +/* A tree of pages that we requested to the source VM */ +GTree *page_requested; +/* For debugging purpose only, but would be nice to keep */ +int page_requested_count; +/* +* The mutex helps to maintain the requested pages that we sent to the +* source, IOW, to guarantee coherent between the page_requests tree and +* the per-ramblock receivedmap. Note! This does not guarantee consistency +* of the real page copy procedures (using UFFDIO_[ZERO]COPY). E.g., even +* if one bit in receivedmap is cleared, UFFDIO_COPY could have happened +* for that page already. This is intended so that the mutex won't +* serialize and blocked by slow operations like UFFDIO_* ioctls. However +* this should be enough to make sure the page_requested tree always +* contains valid information. +*/ +QemuMutex page_request_mutex; };MigrationIncomingState *migration_incoming_get_current (void );@@ -329,7 +346,7 @@ void migrate_send_rp_shut(MigrationIncomingState *mis,void migrate_send_rp_pong (MigrationIncomingState *mis,uint32_t value);int migrate_send_rp_req_pages (MigrationIncomingState *mis,RAMBlock *rb, -ram_addr_t start); +ram_addr_t start,uint64_t haddr);int migrate_send_rp_message_req_pages (MigrationIncomingState *mis,RAMBlock *rb,ram_addr_t start); void migrate_send_rp_recv_bitmap(MigrationIncomingState *mis,migration/postcopy-ram.c +15
-3@@ -684,7 +684,7 @@ int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, qemu_ram_get_idstr(rb), rb_offset);return postcopy_wake_shared (pcfd, client_addr, rb); } -migrate_send_rp_req_pages (mis, rb, aligned_rbo); +migrate_send_rp_req_pages (mis, rb, aligned_rbo, client_addr);return 0 ; }@@ -979,7 +979,8 @@ retry:* Send the request to the source - we want to request one * of our host page sizes (which is >= TPS) */ - ret =migrate_send_rp_req_pages (mis, rb, rb_offset); + ret =migrate_send_rp_req_pages (mis, rb, rb_offset, + msg.arg.pagefault.address);if (ret) {/* May be network failure, try to wait for recovery */ if (ret == -EIO &&postcopy_pause_fault_thread (mis)) {@@ -1149,10 +1150,21 @@ static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr, ret =ioctl (userfault_fd,UFFDIO_ZEROPAGE , &zero_struct); }if (!ret) { +qemu_mutex_lock (&mis->page_request_mutex);ramblock_recv_bitmap_set_range (rb, host_addr, pagesize /qemu_target_page_size ()); +/* +* If this page resolves a page fault for a previous recorded faulted +* address, take a special note to maintain the requested page list. +*/ +if (g_tree_lookup (mis->page_requested, (gconstpointer )host_addr)) { +g_tree_remove (mis->page_requested, (gconstpointer )host_addr); + mis->page_requested_count--; +trace_postcopy_page_req_del ((uint64_t )host_addr, + mis->page_requested_count); + } +qemu_mutex_unlock (&mis->page_request_mutex);mark_postcopy_blocktime_end ((uintptr_t )host_addr); - }return ret; }migration/trace-events +2
-0@@ -157,6 +157,7 @@ postcopy_pause_return_path(void) "" postcopy_pause_return_path_continued(void) "" postcopy_pause_continued(void) "" postcopy_start_set_run(void) "" +postcopy_page_req_add(uint64_t addr, int count) "new page req 0x%lx total %d" source_return_path_thread_bad_end(void) "" source_return_path_thread_end(void) "" source_return_path_thread_entry(void) ""@@ -267,6 +268,7 @@ postcopy_ram_incoming_cleanup_blocktime(uint64_t total) "total blocktime %" PRIu postcopy_request_shared_page(const char *sharer, const char *rb, uint64_t rb_offset) "for %s in %s offset 0x%"PRIx64 postcopy_request_shared_page_present(const char *sharer, const char *rb, uint64_t rb_offset) "%s already %s offset 0x%"PRIx64 postcopy_wake_shared(uint64_t client_addr, const char *rb) "at 0x%"PRIx64" in %s" +postcopy_page_req_del(uint64_t addr, int count) "resolved page req 0x%lx total %d" get_mem_fault_cpu_index(int cpu, uint32_t pid) "cpu: %d, pid: %u"Signature · 1 line
2.26.2
Signature · 1 line
Dr. David Alan Gilbert / [email protected] / Manchester, UK
PATCH v2 6/6 migration: Sync requested pages after postcopy recovery
In reply to Peter XuWe synchronize the requested pages right after a postcopy recovery happens. This helps to synchronize the prioritized pages on source so that the faulted threads can be served faster.
--- migration/savevm.c | 57 ++++++++++++++++++++++++++++++++++++++++++ migration/trace-events | 1 + 2 files changed, 58 insertions(+)
migration/savevm.c +57 -0
@@ -2011,6 +2011,49 @@ static int loadvm_postcopy_handle_run(MigrationIncomingState *mis)return LOADVM_QUIT ; } +/* We must be with page_request_mutex held */ +static gboolean postcopy_sync_page_req (gpointer key,gpointer value, +gpointer data) +{ +MigrationIncomingState *mis = data; +void *host_addr = (void *) key; +ram_addr_t rb_offset; +RAMBlock *rb; +int ret; + + rb =qemu_ram_block_from_host (host_addr, true, &rb_offset); +if (!rb) { +/* +* This should _never_ happen. However be nice for a migrating VM to +* not crash/assert. Post an error (note: intended to not use *_once +* because we do want to see all the illegal addresses; and this can +* never be triggered by the guest so we're safe) and move on next. +*/ +error_report ("%s: illegal host addr %p" , __func__, host_addr); +/* Try the next entry */ +return FALSE; + } + + ret =migrate_send_rp_message_req_pages (mis, rb, rb_offset); +if (ret) { +/* Please refer to above comment. */ +error_report ("%s: send rp message failed for addr %p" , + __func__, host_addr); +return FALSE; + } + +trace_postcopy_page_req_sync ((uint64_t )host_addr); + +return FALSE; +} + +static void migrate_send_rp_req_pages_pending (MigrationIncomingState *mis) +{ + WITH_QEMU_LOCK_GUARD(&mis->page_request_mutex) { +g_tree_foreach (mis->page_requested, postcopy_sync_page_req, mis); + } +} +static int loadvm_postcopy_handle_resume (MigrationIncomingState *mis) {if (mis->state !=MIGRATION_STATUS_POSTCOPY_RECOVER ) {
@@ -2033,6 +2076,20 @@ static int loadvm_postcopy_handle_resume(MigrationIncomingState *mis)/* Tell source that "we are ready" */ migrate_send_rp_resume_ack (mis,MIGRATION_RESUME_ACK_VALUE ); +/* +* After a postcopy recovery, the source should have lost the postcopy +* queue, or potentially the requested pages could have been lost during +* the network down phase. Let's re-sync with the source VM by re-sending +* all the pending pages that we eagerly need, so these threads won't get +* blocked too long due to the recovery. +* +* Without this procedure, the faulted destination VM threads (waiting for +* page requests right before the postcopy is interrupted) can keep hanging +* until the pages are sent by the source during the background copying of +* pages, or another thread faulted on the same address accidentally. +*/ +migrate_send_rp_req_pages_pending (mis); +return 0 ; }
migration/trace-events +1 -0
@@ -49,6 +49,7 @@ vmstate_save(const char *idstr, const char *vmsd_name) "%s, %s" vmstate_load(const char *idstr, const char *vmsd_name) "%s, %s" postcopy_pause_incoming(void) "" postcopy_pause_incoming_continued(void) "" +postcopy_page_req_sync(uint64_t host_addr) "sync page req 0x%"PRIx64 # vmstate.c vmstate_load_field_error(const char *field, int ret) "field \"%s\" load failed, ret = %d"
Signature · 1 line
2.26.2
Quoted 1 line · * Peter Xu ([email protected]) wrote:
v2:
Queued
Quoted 48 lines
- add r-bs for Dave
- add patch "migration: Properly destroy variables on incoming side" as patch 1
- destroypage_request_mutexinmigration_incoming_state_destroy()too [Dave]
- useWITH_QEMU_LOCK_GUARDin two places where we can [Dave]We've seen conditional guest hangs on destination VM after postcopy recovered. However the hang will resolve itself after a few minutes.
The problem is: after a postcopy recovery, the prioritized postcopy queue on the source VM is actually missing. So all the faulted threads before the postcopy recovery happened will keep halted until (accidentally) the page got copied by the background precopy migration stream.
The solution is to also refresh this information after postcopy recovery. To achieve this, we need to maintain a list of faulted addresses on the destination node, so that we can resend the list when necessary. This work is done via patch 2-5.
With that, the last thing we need to do is to send this extra information to source VM after recovered. Very luckily, this synchronization can be "emulated" by sending a bunch of page requests (although these pages have been sent previously!) to source VM just like when we've got a page fault. Even in the 1st version of the postcopy code we'll handle duplicated pages well. So this fix does not even need a new capability bit and it'll work smoothly on old QEMUs when we migrate from them to the new QEMUs.
Please review, thanks.
Peter Xu (6):
migration: Properly destroy variables on incoming side
migration: Reworkmigrate_send_rp_req_pages()function
migration: Pass incoming state intoqemu_ufd_copy_ioctl()
migration: Introducemigrate_send_rp_message_req_pages()
migration: Maintain postcopy faulted addresses
migration: Sync requested pages after postcopy recoverymigration/migration.c | 79 +++++++++++++++++++++++++++++++++++----- migration/migration.h | 23 +++++++++++- migration/postcopy-ram.c | 46 ++++++++++------------- migration/savevm.c | 57 +++++++++++++++++++++++++++++ migration/trace-events | 3 ++ 5 files changed, 171 insertions(+), 37 deletions(-)Signature · 1 line
2.26.2
Signature · 1 line
Dr. David Alan Gilbert / [email protected] / Manchester, UK
Quoted 4 lines · * Dr. David Alan Gilbert ([email protected]) wrote:
Quoted 1 line · * Peter Xu ([email protected]) wrote:
v2:
Queued
Hi Peter,
I've had to unqueue most of this; it doesn't like building on 32bit.
I fixed up the trace_ stuff easily (_del can take a void*, add just needs to use PRIX64) but there are other places where it doesn't like the casting from pointers to uint64_t's etc.
I've kept the first couple of commits.
Dave
Quoted 50 lines
Quoted 48 lines
- add r-bs for Dave
- add patch "migration: Properly destroy variables on incoming side" as patch 1
- destroypage_request_mutexinmigration_incoming_state_destroy()too [Dave]
- useWITH_QEMU_LOCK_GUARDin two places where we can [Dave]We've seen conditional guest hangs on destination VM after postcopy recovered. However the hang will resolve itself after a few minutes.
The problem is: after a postcopy recovery, the prioritized postcopy queue on the source VM is actually missing. So all the faulted threads before the postcopy recovery happened will keep halted until (accidentally) the page got copied by the background precopy migration stream.
The solution is to also refresh this information after postcopy recovery. To achieve this, we need to maintain a list of faulted addresses on the destination node, so that we can resend the list when necessary. This work is done via patch 2-5.
With that, the last thing we need to do is to send this extra information to source VM after recovered. Very luckily, this synchronization can be "emulated" by sending a bunch of page requests (although these pages have been sent previously!) to source VM just like when we've got a page fault. Even in the 1st version of the postcopy code we'll handle duplicated pages well. So this fix does not even need a new capability bit and it'll work smoothly on old QEMUs when we migrate from them to the new QEMUs.
Please review, thanks.
Peter Xu (6):
migration: Properly destroy variables on incoming side
migration: Reworkmigrate_send_rp_req_pages()function
migration: Pass incoming state intoqemu_ufd_copy_ioctl()
migration: Introducemigrate_send_rp_message_req_pages()
migration: Maintain postcopy faulted addresses
migration: Sync requested pages after postcopy recoverymigration/migration.c | 79 +++++++++++++++++++++++++++++++++++----- migration/migration.h | 23 +++++++++++- migration/postcopy-ram.c | 46 ++++++++++------------- migration/savevm.c | 57 +++++++++++++++++++++++++++++ migration/trace-events | 3 ++ 5 files changed, 171 insertions(+), 37 deletions(-)Signature · 1 line
2.26.2Signature · 1 line
Dr. David Alan Gilbert / [email protected] / Manchester, UK
Signature · 1 line
Dr. David Alan Gilbert / [email protected] / Manchester, UK
Quoted 6 lines · Dr. David Alan Gilbert
Quoted 4 lines · * Dr. David Alan Gilbert ([email protected]) wrote:
Quoted 1 line · * Peter Xu ([email protected]) wrote:
v2:
Queued
Quoted 7 lines
Hi Peter,
I've had to unqueue most of this; it doesn't like building on 32bit.
I fixed up thetrace_stuff easily (_delcan take a void*, add just needs to use PRIX64) but there are other places where it doesn't like the casting from pointers to uint64_t's etc.I've kept the first couple of commits.
Thanks, Dave. I'll have a look later and repost.
Signature · 1 line
Peter Xu