kvm/migration: support KVM_CLEAR_DIRTY_LOG
cover kvm/migration: support KVM_CLEAR_DIRTY_LOG
Summary
=====================
Based-on: <[email protected]>
(It's "[PATCH] checkpatch: allow SPDX-License-Identifier", not a big
deal, just to make sure no spoil message since one patch used SPDX license identifier and checkpatch doesn't like it...)
This series allows QEMU to start using the new KVM_CLEAR_DIRTY_LOG interface. For more on KVM_CLEAR_DIRTY_LOG itself, please refer to:
https://github.com/torvalds/linux/blob/master/Documentation/virtual/kvm/api.txt#L3810
The QEMU work (which is this series) is pushed too, please find the tree here:
https://github.com/xzpeter/qemu/tree/kvm-clear-dirty-log
Meanwhile, For anyone who really wants to try this out, please also upgrade the host kernel and use the tree here:
https://github.com/xzpeter/linux/tree/kvm-clear-dirty-fixes
The new kernel is required because there are still some fixes to make the whole thing work, which have not yet been finalized and they haven't reached Linux master.
Design
===================
I started with a naive/stupid design that I always pass all 1's to the KVM for a memory range to clear all the dirty bits within that memory range, but then I encountered guest oops - it's simply because we can't clear any dirty bit from QEMU if we are not _sure_ that the bit is dirty in the kernel. Otherwise we might accidentally clear a bit that we don't even know of (e.g., the bit was clear in migration's dirty bitmap in QEMU) but actually that page was just being written so QEMU will never remember to migrate that new page again.
The new design is focused on a dirty bitmap cache within the QEMU kvm layer (which is per kvm memory slot). With that we know what's dirty in the kernel previously (note! the kernel bitmap is still growing all the time so the cache will only be a subset of the realtime kernel bitmap but that's far enough for us) and with that we'll be sure to not accidentally clear unknown dirty pages.
With this method, we can also avoid race when multiple users (e.g., DIRTY_MEMORY_VGA and DIRTY_MEMORY_MIGRATION) want to clear the bit for multiple time. If without the kvm memory slot cached dirty bitmap we won't be able to know which bit has been cleared and then if we send the CLEAR operation upon the same bit twice (or more) we can still face the same issue to clear something accidentally while we shouldn't.
Summary: we really need to be careful on what bit to clear otherwise we can face anything after the migration completes. And I hope this series has considered all about this.
Besides the new KVM cache layer and the new ioctl support, this series introduced the memory_region_clear_dirty_bitmap() in the memory API layer to allow clearing dirty bits of a specific memory range within the memory region.
Implementations
============================
Patch 1-3: these should be nothing directly related to the series but
they are things I found during working on it. They can be picked even earlier if reviewers are happy with them.
Patch 4: pre-work on bitmap operations, and within the patch I added
the first unit test for utils/bitmap.c.
Patch 5-6: the new memory API interface. Since no one is providing
log_clear() yet so it's not working yet. Note that this only splits the dirty clear operation from sync but it hasn't yet been splitted into smaller chunk so it's not really helpful for us yet.
Patch 7-10: kvm support of KVM_CLEAR_DIRTY_LOG.
Patch 11: do the log_clear() splitting for the case of migration.
Also a new parameter is introduced to define the block size of the small chunks (the unit to clear dirty bits)
Tests
===========================
- make check
- migrate idle/memory-heavy guests
(Not yet tested with huge guests but it'll be more than welcomed if
someone has the resource and wants to give it a shot)
Please have a look, thanks.
Peter Xu (11):
migration: No need to take rcu during sync_dirty_bitmap
memory: Remove memory_region_get_dirty()
memory: Don't set migration bitmap when without migration
bitmap: Add bitmap_copy_with_{src|dst}_offset()
memory: Pass mr into snapshot_and_clear_dirty
memory: Introduce memory listener hook log_clear()
kvm: Update comments for sync_dirty_bitmap
kvm: Persistent per kvmslot dirty bitmap
kvm: Introduce slots lock for memory listener
kvm: Support KVM_CLEAR_DIRTY_LOG
migration: Split log_clear() into smaller chunks
accel/kvm/kvm-all.c | 287 ++++++++++++++++++++++++++++++++++----- accel/kvm/trace-events | 1 + exec.c | 15 +- include/exec/memory.h | 36 ++--- include/exec/ram_addr.h | 91 ++++++++++++- include/qemu/bitmap.h | 9 ++ include/sysemu/kvm_int.h | 4 + memory.c | 64 +++++++-- migration/migration.c | 4 + migration/migration.h | 27 ++++ migration/ram.c | 45 ++++++ migration/trace-events | 1 + tests/Makefile.include | 2 + tests/test-bitmap.c | 81 +++++++++++ util/bitmap.c | 73 ++++++++++ 15 files changed, 671 insertions(+), 69 deletions(-) create mode 100644 tests/test-bitmap.c
Signature · 1 line
2.17.1
PATCH 1/11 migration: No need to take rcu during sync_dirty_bitmap
cpu_physical_memory_sync_dirty_bitmap() has one RAMBlock* as parameter, which means that it must be with RCU read lock held already. Taking it again inside seems redundant. Removing it. Instead comment on the functions about the RCU read lock.
--- include/exec/ram_addr.h | 5 +---- migration/ram.c | 1 + 2 files changed, 2 insertions(+), 4 deletions(-)
include/exec/ram_addr.h +1 -4
@@ -408,6 +408,7 @@ static inline void cpu_physical_memory_clear_dirty_range(ram_addr_t start, } +/* Must be with rcu read lock held */ static inline uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb, ram_addr_t start,
@@ -431,8 +432,6 @@ uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb, DIRTY_MEMORY_BLOCK_SIZE);unsigned long page =BIT_WORD (start >>TARGET_PAGE_BITS ); -rcu_read_lock (); - src =atomic_rcu_read ( &ram_list.dirty_memory[DIRTY_MEMORY_MIGRATION ])->blocks;
@@ -452,8 +451,6 @@ uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb, idx++; } } - -rcu_read_unlock (); }else {ram_addr_t offset = rb->offset;
migration/ram.c +1 -0
@@ -1680,6 +1680,7 @@ static inline bool migration_bitmap_clear_dirty(RAMState *rs,return ret; } +/* Must be with rcu read lock held */ static void migration_bitmap_sync_range (RAMState *rs,RAMBlock *rb,ram_addr_t start,ram_addr_t length) {
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH 01/11] migration: No need to take rcu during sync_dirty_bitmap
Quoted 47 lines · Peter Xu
cpu_physical_memory_sync_dirty_bitmap()has one RAMBlock* as parameter, which means that it must be with RCU read lock held already. Taking it again inside seems redundant. Removing it. Instead comment on the functions about the RCU read lock.Signed-off-by Peter Xu <[email protected]>--- include/exec/ram_addr.h | 5 +---- migration/ram.c | 1 + 2 files changed, 2 insertions(+), 4 deletions(-)include/exec/ram_addr.h +1
-4@@ -408,6 +408,7 @@ static inline void cpu_physical_memory_clear_dirty_range(ram_addr_t start, } +/* Must be with rcu read lock held */ static inline uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb, ram_addr_t start,@@ -431,8 +432,6 @@ uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb, DIRTY_MEMORY_BLOCK_SIZE);unsigned long page =BIT_WORD (start >>TARGET_PAGE_BITS ); -rcu_read_lock (); - src =atomic_rcu_read ( &ram_list.dirty_memory[DIRTY_MEMORY_MIGRATION ])->blocks;@@ -452,8 +451,6 @@ uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb, idx++; } } - -rcu_read_unlock (); }else {ram_addr_t offset = rb->offset;migration/ram.c +0
-0
migration/ram.c static inline bool migration_bitmap_clear_dirty(RAMState *rs, · line 1686
Quoted 7 lines
migration/ram.c +1
-0@@ -1680,6 +1680,7 @@ static inline bool migration_bitmap_clear_dirty(RAMState *rs,return ret; } +/* Must be with rcu read lock held */ static void migration_bitmap_sync_range (RAMState *rs,RAMBlock *rb,ram_addr_t start,ram_addr_t length) {
PATCH 2/11 memory: Remove memory_region_get_dirty()
In reply to Peter XuIt's never used anywhere.
--- include/exec/memory.h | 17 ----------------- memory.c | 8 -------- 2 files changed, 25 deletions(-)
include/exec/memory.h +0 -17
@@ -1254,23 +1254,6 @@ void memory_region_ram_resize(MemoryRegion *mr, ram_addr_t newsize,*/ void memory_region_set_log (MemoryRegion *mr,bool log,unsigned client); -/** -* memory_region_get_dirty: Check whether a range of bytes is dirty -* for a specified client. -* -* Checks whether a range of bytes has been written to since the last -* call to memory_region_reset_dirty() with the same @client. Dirty logging -* must be enabled. -* -* @mr: the memory region being queried. -* @addr: the address (relative to the start of the region) being queried. -* @size: the size of the range being queried. -* @client: the user of the logging information; %DIRTY_MEMORY_MIGRATION or -* %DIRTY_MEMORY_VGA. -*/ -bool memory_region_get_dirty (MemoryRegion *mr,hwaddr addr, -hwaddr size,unsigned client); -/** * memory_region_set_dirty: Mark a range of bytes as dirty in a memory region. *
memory.c +0 -8
@@ -2027,14 +2027,6 @@ void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client)memory_region_transaction_commit (); } -bool memory_region_get_dirty (MemoryRegion *mr,hwaddr addr, -hwaddr size,unsigned client) -{ -assert (mr->ram_block); -return cpu_physical_memory_get_dirty (memory_region_get_ram_addr (mr) + addr, - size, client); -} -void memory_region_set_dirty (MemoryRegion *mr,hwaddr addr,hwaddr size) {
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH 02/11] memory: Remove memory_region_get_dirty()
Quoted 46 lines · Peter Xu
It's never used anywhere.
Signed-off-by Peter Xu <[email protected]>--- include/exec/memory.h | 17 ----------------- memory.c | 8 -------- 2 files changed, 25 deletions(-)include/exec/memory.h +0
-17@@ -1254,23 +1254,6 @@ void memory_region_ram_resize(MemoryRegion *mr, ram_addr_t newsize,*/ void memory_region_set_log (MemoryRegion *mr,bool log,unsigned client); -/** -* memory_region_get_dirty: Check whether a range of bytes is dirty -* for a specified client. -* -* Checks whether a range of bytes has been written to since the last -* call to memory_region_reset_dirty() with the same @client. Dirty logging -* must be enabled. -* -* @mr: the memory region being queried. -* @addr: the address (relative to the start of the region) being queried. -* @size: the size of the range being queried. -* @client: the user of the logging information; %DIRTY_MEMORY_MIGRATION or -* %DIRTY_MEMORY_VGA. -*/ -bool memory_region_get_dirty (MemoryRegion *mr,hwaddr addr, -hwaddr size,unsigned client); -/** * memory_region_set_dirty: Mark a range of bytes as dirty in a memory region. * memory.c +0
-1@@ -2027,14 +2027,6 @@ void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client)memory_region_transaction_commit (); } -bool memory_region_get_dirty(MemoryRegion *mr, hwaddr addr,
memory.c void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client) · line 2032
Quoted 10 lines
memory.c +0
-7- hwaddr size, unsigned client) -{ -assert (mr->ram_block); -return cpu_physical_memory_get_dirty (memory_region_get_ram_addr (mr) + addr, - size, client); -} -void memory_region_set_dirty (MemoryRegion *mr,hwaddr addr,hwaddr size) {
PATCH 3/11 memory: Don't set migration bitmap when without migration
In reply to Peter XuSimilar to 9460dee4b2 ("memory: do not touch code dirty bitmap unless TCG is enabled", 2015-06-05) but for the migration bitmap - we can skip the MIGRATION bitmap update if migration not enabled.
--- include/exec/memory.h | 2 ++ include/exec/ram_addr.h | 12 +++++++++++- memory.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)
include/exec/memory.h +2 -0
@@ -46,6 +46,8 @@OBJECT_GET_CLASS (IOMMUMemoryRegionClass, (obj), \ TYPE_IOMMU_MEMORY_REGION) +extern bool global_dirty_log; +typedef struct MemoryRegionOps MemoryRegionOps ;typedef struct MemoryRegionMmio MemoryRegionMmio ;
include/exec/ram_addr.h +11 -1
@@ -348,8 +348,13 @@ static inline void cpu_physical_memory_set_dirty_lebitmap(unsigned long *bitmap,if (bitmap[k]) {unsigned long temp =leul_to_cpu (bitmap[k]); -atomic_or (&blocks[DIRTY_MEMORY_MIGRATION ][idx][offset], temp);atomic_or (&blocks[DIRTY_MEMORY_VGA ][idx][offset], temp); + +if (global_dirty_log) { +atomic_or (&blocks[DIRTY_MEMORY_MIGRATION ][idx][offset], + temp); + } +if (tcg_enabled ()) {atomic_or (&blocks[DIRTY_MEMORY_CODE ][idx][offset], temp); }
@@ -366,6 +371,11 @@ static inline void cpu_physical_memory_set_dirty_lebitmap(unsigned long *bitmap,xen_hvm_modified_memory (start, pages <<TARGET_PAGE_BITS ); }else {uint8_t clients =tcg_enabled () ?DIRTY_CLIENTS_ALL :DIRTY_CLIENTS_NOCODE ; + +if (!global_dirty_log) { + clients &= ~(1 <<DIRTY_MEMORY_MIGRATION ); + } +/* * bitmap-traveling is faster than memory-traveling (for addr...) * especially when most of the memory is not dirty.
memory.c +1 -1
@@ -38,7 +38,7 @@static unsigned memory_region_transaction_depth;static bool memory_region_update_pending;static bool ioeventfd_update_pending; -static bool global_dirty_log = false; +bool global_dirty_log; static QTAILQ_HEAD(, MemoryListener) memory_listeners =QTAILQ_HEAD_INITIALIZER (memory_listeners);
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH 03/11] memory: Don't set migration bitmap when without migration
Quoted 61 lines · Peter Xu
Similar to
9460dee4b2("memory: do not touch code dirty bitmap unless TCG is enabled", 2015-06-05) but for the migration bitmap - we can skip the MIGRATION bitmap update if migration not enabled.Signed-off-by Peter Xu <[email protected]>--- include/exec/memory.h | 2 ++ include/exec/ram_addr.h | 12 +++++++++++- memory.c | 2 +- 3 files changed, 14 insertions(+), 2 deletions(-)include/exec/memory.h +2
-0@@ -46,6 +46,8 @@OBJECT_GET_CLASS (IOMMUMemoryRegionClass, (obj), \ TYPE_IOMMU_MEMORY_REGION) +extern bool global_dirty_log; +typedef struct MemoryRegionOps MemoryRegionOps ;typedef struct MemoryRegionMmio MemoryRegionMmio ;include/exec/ram_addr.h +11
-1@@ -348,8 +348,13 @@ static inline void cpu_physical_memory_set_dirty_lebitmap(unsigned long *bitmap,if (bitmap[k]) {unsigned long temp =leul_to_cpu (bitmap[k]); -atomic_or (&blocks[DIRTY_MEMORY_MIGRATION ][idx][offset], temp);atomic_or (&blocks[DIRTY_MEMORY_VGA ][idx][offset], temp); + +if (global_dirty_log) { +atomic_or (&blocks[DIRTY_MEMORY_MIGRATION ][idx][offset], + temp); + } +if (tcg_enabled ()) {atomic_or (&blocks[DIRTY_MEMORY_CODE ][idx][offset], temp); }@@ -366,6 +371,11 @@ static inline void cpu_physical_memory_set_dirty_lebitmap(unsigned long *bitmap,xen_hvm_modified_memory (start, pages <<TARGET_PAGE_BITS ); }else {uint8_t clients =tcg_enabled () ?DIRTY_CLIENTS_ALL :DIRTY_CLIENTS_NOCODE ; + +if (!global_dirty_log) { + clients &= ~(1 <<DIRTY_MEMORY_MIGRATION ); + } +/* * bitmap-traveling is faster than memory-traveling (for addr...) * especially when most of the memory is not dirty. memory.c +0
-0
memory.c line 44
Quoted 8 lines
memory.c +1
-1@@ -38,7 +38,7 @@static unsigned memory_region_transaction_depth;static bool memory_region_update_pending;static bool ioeventfd_update_pending; -static bool global_dirty_log = false; +bool global_dirty_log; static QTAILQ_HEAD(, MemoryListener) memory_listeners =QTAILQ_HEAD_INITIALIZER (memory_listeners);
PATCH 4/11 bitmap: Add bitmap_copy_with_{src|dst}_offset()
In reply to Peter XuThese helpers copy the source bitmap to destination bitmap with a shift either on the src or dst bitmap.
Meanwhile, we never have bitmap tests but we should.
This patch also introduces the initial test cases for utils/bitmap.c but it only tests the newly introduced functions.
--- include/qemu/bitmap.h | 9 +++++ tests/Makefile.include | 2 ++ tests/test-bitmap.c | 81 ++++++++++++++++++++++++++++++++++++++++++ util/bitmap.c | 73 +++++++++++++++++++++++++++++++++++++ 4 files changed, 165 insertions(+) create mode 100644 tests/test-bitmap.c
include/qemu/bitmap.h +9 -0
@@ -41,6 +41,10 @@* bitmap_find_next_zero_area(buf, len, pos, n, mask) Find bit free area * bitmap_to_le(dst, src, nbits) Convert bitmap to little endian * bitmap_from_le(dst, src, nbits) Convert bitmap from little endian +* bitmap_copy_with_src_offset(dst, src, offset, nbits) +* *dst = *src (with an offset upon src) +* bitmap_copy_with_dst_offset(dst, src, offset, nbits) +* *dst = *src (with an offset upon dst) */ /*
@@ -271,4 +275,9 @@ void bitmap_to_le(unsigned long *dst, const unsigned long *src,void bitmap_from_le (unsigned long *dst,const unsigned long *src,long nbits); +void bitmap_copy_with_src_offset (unsigned long *dst,const unsigned long *src, +long offset,long nbits); +void bitmap_copy_with_dst_offset (unsigned long *dst,const unsigned long *src, +long shift,long nbits); +#endif /* BITMAP_H */
tests/Makefile.include +2 -0
@@ -69,6 +69,7 @@ check-unit-y += tests/test-aio-multithread$(EXESUF) check-unit-y += tests/test-throttle$(EXESUF) check-unit-y += tests/test-thread-pool$(EXESUF) check-unit-y += tests/test-hbitmap$(EXESUF) +check-unit-y += tests/test-bitmap$(EXESUF) check-unit-y += tests/test-bdrv-drain$(EXESUF) check-unit-y += tests/test-bdrv-graph-mod$(EXESUF) check-unit-y += tests/test-blockjob$(EXESUF)
@@ -548,6 +549,7 @@ tests/test-image-locking$(EXESUF): tests/test-image-locking.o $(test-block-obj-y tests/test-thread-pool$(EXESUF): tests/test-thread-pool.o $(test-block-obj-y) tests/test-iov$(EXESUF): tests/test-iov.o $(test-util-obj-y) tests/test-hbitmap$(EXESUF): tests/test-hbitmap.o $(test-util-obj-y) $(test-crypto-obj-y) +tests/test-bitmap$(EXESUF): tests/test-bitmap.o $(test-util-obj-y) tests/test-x86-cpuid$(EXESUF): tests/test-x86-cpuid.o tests/test-xbzrle$(EXESUF): tests/test-xbzrle.o migration/xbzrle.o migration/page_cache.o $(test-util-obj-y) tests/test-cutils$(EXESUF): tests/test-cutils.o util/cutils.o $(test-util-obj-y)
tests/test-bitmap.c new file +81 -0
81 lines not on this page · raw message
util/bitmap.c +73 -0
76 lines not on this page · raw message
Signature · 1 line
2.17.1
PATCH 5/11 memory: Pass mr into snapshot_and_clear_dirty
In reply to Peter XuAlso we change the 2nd parameter of it to be the relative offset within the memory region. This is to be used in follow up patches.
--- exec.c | 3 ++- include/exec/ram_addr.h | 2 +- memory.c | 3 +-- 3 files changed, 4 insertions(+), 4 deletions(-)
exec.c +2 -1
@@ -1387,9 +1387,10 @@ bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start, }DirtyBitmapSnapshot *cpu_physical_memory_snapshot_and_clear_dirty - (ram_addr_t start,ram_addr_t length,unsigned client) +(MemoryRegion *mr,hwaddr addr,hwaddr length,unsigned client) {DirtyMemoryBlocks *blocks; +ram_addr_t start =memory_region_get_ram_addr (mr) + addr;unsigned long align =1UL << (TARGET_PAGE_BITS +BITS_PER_LEVEL );ram_addr_t first =QEMU_ALIGN_DOWN (start, align);ram_addr_t last =QEMU_ALIGN_UP (start + length, align);
include/exec/ram_addr.h +1 -1
@@ -403,7 +403,7 @@ bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start, unsigned client);DirtyBitmapSnapshot *cpu_physical_memory_snapshot_and_clear_dirty - (ram_addr_t start,ram_addr_t length,unsigned client); +(MemoryRegion *mr,ram_addr_t start,hwaddr length,unsigned client); bool cpu_physical_memory_snapshot_get_dirty(DirtyBitmapSnapshot *snap, ram_addr_t start,
memory.c +1 -2
@@ -2071,8 +2071,7 @@ DirtyBitmapSnapshot *memory_region_snapshot_and_clear_dirty(MemoryRegion *mr, {assert (mr->ram_block);memory_region_sync_dirty_bitmap (mr); -return cpu_physical_memory_snapshot_and_clear_dirty ( -memory_region_get_ram_addr (mr) + addr, size, client); +return cpu_physical_memory_snapshot_and_clear_dirty (mr, addr, size, client); } bool memory_region_snapshot_get_dirty(MemoryRegion *mr, DirtyBitmapSnapshot *snap,
Signature · 1 line
2.17.1
PATCH 6/11 memory: Introduce memory listener hook log_clear()
In reply to Peter XuIntroduce a new memory region listener hook log_clear() to allow the listeners to hook onto the points where the dirty bitmap is cleared by the bitmap users.
Previously log_sync() contains two operations:
- dirty bitmap collection, and,
- dirty bitmap clear on remote site.
Let's take KVM as example - log_sync() for KVM will first copy the kernel dirty bitmap to userspace, and at the same time we'll clear the dirty bitmap there along with re-protecting all the guest pages again.
We add this new log_clear() interface only to split the old log_sync() into two separated procedures:
- use log_sync() to collect the collection only, and,
- use log_clear() to clear the remote dirty bitmap.
With the new interface, the memory listener users will still be able to decide how to implement the log synchronization procedure, e.g., they can still only provide log_sync() method only and put all the two procedures within log_sync() (that's how the old KVM works before KVM_CAP_MANUAL_DIRTY_LOG_PROTECT is introduced). However with this new interface the memory listener users will start to have a chance to postpone the log clear operation explicitly if the module supports. That can really benefit users like KVM at least for host kernels that support KVM_CAP_MANUAL_DIRTY_LOG_PROTECT.
There are three places that can clear dirty bits in any one of the dirty bitmap in the ram_list.dirty_memory[3] array:
cpu_physical_memory_snapshot_and_clear_dirty cpu_physical_memory_test_and_clear_dirty cpu_physical_memory_sync_dirty_bitmap
Currently we hook directly into each of the functions to notify about the log_clear().
--- exec.c | 12 ++++++++++ include/exec/memory.h | 17 ++++++++++++++ include/exec/ram_addr.h | 3 +++ memory.c | 51 +++++++++++++++++++++++++++++++++++++++++ 4 files changed, 83 insertions(+)
exec.c +12 -0
@@ -1355,6 +1355,8 @@ bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start,DirtyMemoryBlocks *blocks;unsigned long end, page;bool dirty = false; +RAMBlock *ramblock; +uint64_t mr_offset, mr_size;if (length ==0 ) {return false;
@@ -1366,6 +1368,10 @@ bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start,rcu_read_lock (); blocks =atomic_rcu_read (&ram_list.dirty_memory[client]); + ramblock =qemu_get_ram_block (start); +/* Range sanity check on the ramblock */ +assert (start >= ramblock->offset && + start + length <= ramblock->offset + ramblock->used_length);while (page < end) {unsigned long idx = page /DIRTY_MEMORY_BLOCK_SIZE ;
@@ -1377,6 +1383,10 @@ bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start, page += num; } + mr_offset = (ram_addr_t)(page <<TARGET_PAGE_BITS ) - ramblock->offset; + mr_size = (end - page) <<TARGET_PAGE_BITS ; +memory_region_clear_dirty_bitmap (ramblock->mr, mr_offset, mr_size); +rcu_read_unlock ();if (dirty &&tcg_enabled ()) {
@@ -1432,6 +1442,8 @@ DirtyBitmapSnapshot *cpu_physical_memory_snapshot_and_clear_dirtytlb_reset_dirty_range_all (start, length); } +memory_region_clear_dirty_bitmap (mr, addr, length); +return snap; }
include/exec/memory.h +17 -0
@@ -416,6 +416,7 @@ struct MemoryListener {void (*log_stop)(MemoryListener *listener,MemoryRegionSection *section,int old,int new);void (*log_sync)(MemoryListener *listener,MemoryRegionSection *section); +void (*log_clear)(MemoryListener *listener,MemoryRegionSection *section);void (*log_global_start)(MemoryListener *listener);void (*log_global_stop)(MemoryListener *listener); void (*eventfd_add)(MemoryListener *listener, MemoryRegionSection *section,
@@ -1269,6 +1270,22 @@ void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client);void memory_region_set_dirty (MemoryRegion *mr,hwaddr addr,hwaddr size); +/** +* memory_region_clear_dirty_bitmap - clear dirty bitmap for memory range +* +* This function is called when the caller wants to clear the remote +* dirty bitmap of a memory range within the memory region. This can +* be used by e.g. KVM to manually clear dirty log when +* KVM_CAP_MANUAL_DIRTY_LOG_PROTECT is declared support by the host +* kernel. +* +* @mr: the memory region to clear the dirty log upon +* @start: start address offset within the memory region +* @len: length of the memory region to clear dirty bitmap +*/ +void memory_region_clear_dirty_bitmap (MemoryRegion *mr,hwaddr start, +hwaddr len); +/** * memory_region_snapshot_and_clear_dirty: Get a snapshot of the dirty * bitmap and clear it.
include/exec/ram_addr.h +3 -0
@@ -461,6 +461,9 @@ uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb, idx++; } } + +/* TODO: split the huge bitmap into smaller chunks */ +memory_region_clear_dirty_bitmap (rb->mr, start, length); }else {ram_addr_t offset = rb->offset;
memory.c +51 -0
57 lines not on this page · raw message
Signature · 1 line
2.17.1
PATCH 7/11 kvm: Update comments for sync_dirty_bitmap
In reply to Peter XuIt's obviously obsolete. Do some update.
--- accel/kvm/kvm-all.c | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-)
accel/kvm/kvm-all.c +6 -6
@@ -473,13 +473,13 @@ static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,#define ALIGN (x, y) (((x)+(y)-1) & ~((y)-1))/** -* kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space -* This function updates qemu's dirty bitmap using -* memory_region_set_dirty(). This means all bits are set -* to dirty. +* kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space * -* @start_add: start of logged region. -* @end_addr: end of logged region. +* This function will first try to fetch dirty bitmap from the kernel, +* and then updates qemu's dirty bitmap. +* +* @kml: the KVM memory listener object +* @section: the memory section to sync the dirty bitmap with */ static int kvm_physical_sync_dirty_bitmap (KVMMemoryListener *kml,MemoryRegionSection *section)
Signature · 1 line
2.17.1
PATCH 8/11 kvm: Persistent per kvmslot dirty bitmap
In reply to Peter XuWhen synchronizing dirty bitmap from kernel KVM we do it in a per-kvmslot fashion and we allocate the userspace bitmap for each of the ioctl. This patch instead make the bitmap cache be persistent then we don't need to g_malloc0() every time.
More importantly, the cached per-kvmslot dirty bitmap will be further used when we want to add support for the KVM_CLEAR_DIRTY_LOG and this cached bitmap will be used to guarantee we won't clear any unknown dirty bits otherwise that can be a severe data loss issue for migration code.
--- accel/kvm/kvm-all.c | 39 +++++++++++++++++++++------------------ include/sysemu/kvm_int.h | 2 ++ 2 files changed, 23 insertions(+), 18 deletions(-)
accel/kvm/kvm-all.c +21 -18
@@ -497,31 +497,14 @@ static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,return 0 ; } -/* XXX bad kernel interface alert -* For dirty bitmap, kernel allocates array of size aligned to -* bits-per-long. But for case when the kernel is 64bits and -* the userspace is 32bits, userspace can't align to the same -* bits-per-long, since sizeof(long) is different between kernel -* and user space. This way, userspace will provide buffer which -* may be 4 bytes less than the kernel will use, resulting in -* userspace memory corruption (which is not detectable by valgrind -* too, in most cases). -* So for now, let's align to 64 instead of HOST_LONG_BITS here, in -* a hope that sizeof(long) won't become >8 any time soon. -*/ - size =ALIGN (((mem->memory_size) >>TARGET_PAGE_BITS ), -/*HOST_LONG_BITS*/ 64 ) /8 ; - d.dirty_bitmap =g_malloc0 (size); - + d.dirty_bitmap = mem->dirty_bmap; d.slot = mem->slot | (kml->as_id <<16 );if (kvm_vm_ioctl (s,KVM_GET_DIRTY_LOG , &d) ==-1 ) {DPRINTF ("ioctl failed %d\n" , errno); -g_free (d.dirty_bitmap);return -1 ; }kvm_get_dirty_pages_log_range (section, d.dirty_bitmap); -g_free (d.dirty_bitmap); }return 0 ;
@@ -765,6 +748,7 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml,MemoryRegion *mr = section->mr;bool writeable = !mr->readonly && !mr->rom_device;hwaddr start_addr, size; +unsigned long bmap_size;void *ram;if (!memory_region_is_ram (mr)) {
@@ -796,6 +780,8 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml, }/* unregister the slot */ +g_free (mem->dirty_bmap); + mem->dirty_bmap =NULL ; mem->memory_size =0 ; mem->flags =0 ; err =kvm_set_user_memory_region (kml, mem, false);
@@ -807,12 +793,29 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml,return ; } +/* +* XXX bad kernel interface alert For dirty bitmap, kernel +* allocates array of size aligned to bits-per-long. But for case +* when the kernel is 64bits and the userspace is 32bits, +* userspace can't align to the same bits-per-long, since +* sizeof(long) is different between kernel and user space. This +* way, userspace will provide buffer which may be 4 bytes less +* than the kernel will use, resulting in userspace memory +* corruption (which is not detectable by valgrind too, in most +* cases). So for now, let's align to 64 instead of +* HOST_LONG_BITS here, in a hope that sizeof(long) won't become +* >8 any time soon. +*/ + bmap_size =ALIGN ((size >>TARGET_PAGE_BITS ), +/*HOST_LONG_BITS*/ 64 ) /8 ; +/* register the new slot */ mem =kvm_alloc_slot (kml); mem->memory_size = size; mem->start_addr = start_addr; mem->ram = ram; mem->flags =kvm_mem_flags (mr); + mem->dirty_bmap =g_malloc0 (bmap_size); err =kvm_set_user_memory_region (kml, mem, true);if (err) {
include/sysemu/kvm_int.h +2 -0
@@ -21,6 +21,8 @@ typedef struct KVMSlotint slot;int flags;int old_flags; +/* Dirty bitmap cache for the slot */ +unsigned long *dirty_bmap; } KVMSlot; typedef struct KVMMemoryListener {
Signature · 1 line
2.17.1
PATCH 9/11 kvm: Introduce slots lock for memory listener
In reply to Peter XuIntroduce KVMMemoryListener.slots_lock to protect the slots inside the kvm memory listener. Currently it is close to useless because all the KVM code path now is always protected by the BQL. But it'll start to make sense in follow up patches where we might do remote dirty bitmap clear and also we'll update the per-slot cached dirty bitmap even without the BQL. So let's prepare for it.
We can also use per-slot lock for above reason but it seems to be an overkill. Let's just use this bigger one (which covers all the slots of a single address space) but anyway this lock is still much smaller than the BQL.
--- accel/kvm/kvm-all.c | 58 +++++++++++++++++++++++++++++++--------- include/sysemu/kvm_int.h | 2 ++ 2 files changed, 48 insertions(+), 12 deletions(-)
accel/kvm/kvm-all.c +46 -12
@@ -138,6 +138,9 @@ static const KVMCapabilityInfo kvm_required_capabilites[] = { KVM_CAP_LAST_INFO }; +#define kvm_slots_lock (kml) qemu_mutex_lock(&(kml)->slots_lock) +#define kvm_slots_unlock (kml) qemu_mutex_unlock(&(kml)->slots_lock) +int kvm_get_max_memslots (void ) {KVMState *s =KVM_STATE (current_machine->accelerator);
@@ -165,6 +168,7 @@ int kvm_memcrypt_encrypt_data(uint8_t *ptr, uint64_t len)return 1 ; } +/* Must be with slots_lock held */ static KVMSlot *kvm_get_free_slot (KVMMemoryListener *kml) {KVMState *s = kvm_state;
@@ -182,10 +186,17 @@ static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)bool kvm_has_free_slot (MachineState *ms) {KVMState *s =KVM_STATE (ms->accelerator); +bool result; +KVMMemoryListener *kml = &s->memory_listener; + +kvm_slots_lock (kml); + result = !!kvm_get_free_slot (kml); +kvm_slots_unlock (kml); -return kvm_get_free_slot (&s->memory_listener); +return result; } +/* Must be with slots_lock held */ static KVMSlot *kvm_alloc_slot (KVMMemoryListener *kml) {KVMSlot *slot =kvm_get_free_slot (kml);
@@ -244,18 +255,21 @@ int kvm_physical_memory_addr_from_host(KVMState *s, void *ram, hwaddr *phys_addr) {KVMMemoryListener *kml = &s->memory_listener; -int i; +int i, ret =0 ; +kvm_slots_lock (kml);for (i =0 ; i < s->nr_slots; i++) {KVMSlot *mem = &kml->slots[i];if (ram >= mem->ram && ram < mem->ram + mem->memory_size) { *phys_addr = mem->start_addr + (ram - mem->ram); -return 1 ; + ret =1 ; +break ; } } +kvm_slots_unlock (kml); -return 0 ; +return ret; }static int kvm_set_user_memory_region (KVMMemoryListener *kml,KVMSlot *slot,bool new)
@@ -391,6 +405,7 @@ static int kvm_mem_flags(MemoryRegion *mr)return flags; } +/* Must be with slots_lock held */ static int kvm_slot_update_flags (KVMMemoryListener *kml,KVMSlot *mem,MemoryRegion *mr) {
@@ -409,19 +424,26 @@ static int kvm_section_update_flags(KVMMemoryListener *kml, {hwaddr start_addr, size;KVMSlot *mem; +int ret =0 ; size =kvm_align_section (section, &start_addr);if (!size) {return 0 ; } +kvm_slots_lock (kml); + mem =kvm_lookup_matching_slot (kml, start_addr, size);if (!mem) {/* We don't have a slot if we want to trap every access. */ -return 0 ; + goto out; } -return kvm_slot_update_flags (kml, mem, section->mr); + ret =kvm_slot_update_flags (kml, mem, section->mr); + +out: +kvm_slots_unlock (kml); +return ret; } static void kvm_log_start(MemoryListener *listener,
@@ -478,6 +500,8 @@ static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,* This function will first try to fetch dirty bitmap from the kernel, * and then updates qemu's dirty bitmap. * +* NOTE: caller must be with kml->slots_lock held. +* * @kml: the KVM memory listener object * @section: the memory section to sync the dirty bitmap with */
@@ -488,26 +512,28 @@ static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,struct kvm_dirty_log d = {};KVMSlot *mem;hwaddr start_addr, size; +int ret =0 ; size =kvm_align_section (section, &start_addr);if (size) { mem =kvm_lookup_matching_slot (kml, start_addr, size);if (!mem) {/* We don't have a slot if we want to trap every access. */ -return 0 ; + goto out; } d.dirty_bitmap = mem->dirty_bmap; d.slot = mem->slot | (kml->as_id <<16 );if (kvm_vm_ioctl (s,KVM_GET_DIRTY_LOG , &d) ==-1 ) {DPRINTF ("ioctl failed %d\n" , errno); -return -1 ; + ret =-1 ; + goto out; }kvm_get_dirty_pages_log_range (section, d.dirty_bitmap); } - -return 0 ; +out: +return ret; } static void kvm_coalesce_mmio_region(MemoryListener *listener,
@@ -770,10 +796,12 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml, ram =memory_region_get_ram_ptr (mr) + section->offset_within_region + (start_addr - section->offset_within_address_space); +kvm_slots_lock (kml); +if (!add) { mem =kvm_lookup_matching_slot (kml, start_addr, size);if (!mem) { -return ; + goto out; }if (mem->flags &KVM_MEM_LOG_DIRTY_PAGES ) {kvm_physical_sync_dirty_bitmap (kml, section);
@@ -790,7 +818,7 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml, __func__, strerror(-err));abort (); } -return ; + goto out; }/*
@@ -823,6 +851,9 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml, strerror(-err));abort (); } + +out: +kvm_slots_unlock (kml); } static void kvm_region_add(MemoryListener *listener,
@@ -849,7 +880,9 @@ static void kvm_log_sync(MemoryListener *listener,KVMMemoryListener *kml =container_of (listener, KVMMemoryListener, listener);int r; +kvm_slots_lock (kml); r =kvm_physical_sync_dirty_bitmap (kml, section); +kvm_slots_unlock (kml);if (r <0 ) {abort (); }
@@ -929,6 +962,7 @@ void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml, {int i; +qemu_mutex_init (&kml->slots_lock); kml->slots =g_malloc0 (s->nr_slots *sizeof (KVMSlot)); kml->as_id = as_id;
include/sysemu/kvm_int.h +2 -0
@@ -27,6 +27,8 @@ typedef struct KVMSlottypedef struct KVMMemoryListener {MemoryListener listener; +/* Protects the slots and all inside them */ +QemuMutex slots_lock;KVMSlot *slots;int as_id; }KVMMemoryListener ;
Signature · 1 line
2.17.1
PATCH 10/11 kvm: Support KVM_CLEAR_DIRTY_LOG
In reply to Peter XuFirstly detect the interface using KVM_CAP_MANUAL_DIRTY_LOG_PROTECT and mark it. When failed to enable the new feature we'll fall back to the old sync.
Provide the log_clear() hook for the memory listeners for both address spaces of KVM (normal system memory, and SMM) and deliever the clear message to kernel.
--- accel/kvm/kvm-all.c | 180 +++++++++++++++++++++++++++++++++++++++++ accel/kvm/trace-events | 1 + 2 files changed, 181 insertions(+)
accel/kvm/kvm-all.c +180 -0
@@ -91,6 +91,7 @@ struct KVMStateint many_ioeventfds;int intx_set_mask;bool sync_mmu; +bool manual_dirty_log_protect;/* The man page (and posix) say ioctl numbers are signed int, but * they're not. Linux, glibc and *BSD all treat ioctl numbers as * unsigned, and treating them as signed here can break things */
@@ -536,6 +537,157 @@ out:return ret; } +/* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */ +#define KVM_CLEAR_LOG_SHIFT 6 +#define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT) +#define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN) + +/** +* kvm_physical_log_clear - Clear the kernel's dirty bitmap for range +* +* NOTE: this will be a no-op if we haven't enabled manual dirty log +* protection in the host kernel because in that case this operation +* will be done within log_sync(). +* +* @kml: the kvm memory listener +* @section: the memory range to clear dirty bitmap +*/ +static int kvm_physical_log_clear (KVMMemoryListener *kml, +MemoryRegionSection *section) +{ +KVMState *s = kvm_state; +struct kvm_clear_dirty_log d; +uint64_t start, end, bmap_start, start_delta, bmap_npages, size; +unsigned long *bmap_clear =NULL , psize = qemu_real_host_page_size; +KVMSlot *mem; +int ret, i; + +if (!s->manual_dirty_log_protect) { +/* No need to do explicit clear */ +return 0 ; + } + + start = section->offset_within_address_space; + size =int128_get64 (section->size); + +if (!size) { +/* Nothing more we can do... */ +return 0 ; + } + +kvm_slots_lock (kml); + +/* Find any possible slot that covers the section */ +for (i =0 ; i < s->nr_slots; i++) { + mem = &kml->slots[i]; +if (mem->start_addr <= start && + start + size <= mem->start_addr + mem->memory_size) { +break ; + } + } + +/* +* We should always find one memslot until this point, otherwise +* there could be something wrong from the upper layer +*/ +assert (i != s->nr_slots); + +/* +* We need to extend either the start or the size or both to +* satisfy the KVM interface requirement. Firstly, do the start +* page alignment on 64 host pages +*/ + bmap_start = (start - mem->start_addr) &KVM_CLEAR_LOG_MASK ; + start_delta = start - mem->start_addr - bmap_start; + bmap_start /= psize; + +/* +* The kernel interface has restriction on the size too, that either: +* +* (1) the size is 64 host pages aligned (just like the start), or +* (2) the size fills up until the end of the KVM memslot. +*/ + bmap_npages =DIV_ROUND_UP (size + start_delta,KVM_CLEAR_LOG_ALIGN ) + <<KVM_CLEAR_LOG_SHIFT ; + end = mem->memory_size / psize; +if (bmap_npages > end - bmap_start) { + bmap_npages = end - bmap_start; + } + start_delta /= psize; + +/* +* Prepare the bitmap to clear dirty bits. Here we must guarantee +* that we won't clear any unknown dirty bits otherwise we might +* accidentally clear some set bits which are not yet synced from +* the kernel into QEMU's bitmap, then we'll lose track of the +* guest modifications upon those pages (which can directly lead +* to guest data loss or panic after migration). +* +* Layout of the KVMSlot.dirty_bmap: +* +* |<-------- bmap_npages -----------..>| +* [1] +* start_delta size +* |----------------|-------------|------------------|------------| +* ^ ^ ^ ^ +* | | | | +* start bmap_start (start) end +* of memslot of memslot +* +* [1] bmap_npages can be aligned to either 64 pages or the end of slot +*/ + +assert (bmap_start %BITS_PER_LONG ==0 ); +if (start_delta) { +/* Slow path - we need to manipulate a temp bitmap */ + bmap_clear =bitmap_new (bmap_npages); +bitmap_copy_with_src_offset (bmap_clear, mem->dirty_bmap, + bmap_start, start_delta + size / psize); +/* +* We need to fill the holes at start because that was not +* specified by the caller and we extended the bitmap only for +* 64 pages alignment +*/ +bitmap_clear (bmap_clear,0 , start_delta); + d.dirty_bitmap = bmap_clear; + }else { +/* Fast path - start address aligns well with BITS_PER_LONG */ + d.dirty_bitmap = mem->dirty_bmap +BIT_WORD (bmap_start); + } + + d.first_page = bmap_start; +/* It should never overflow. If it happens, say something */ +assert (bmap_npages <=UINT32_MAX ); + d.num_pages = bmap_npages; + d.slot = mem->slot | (kml->as_id <<16 ); + +if (kvm_vm_ioctl (s,KVM_CLEAR_DIRTY_LOG , &d) ==-1 ) { + ret = -errno; +error_report ("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, " +"start=0x%" PRIx64", size=0x%" PRIx32", errno=%d" , + __func__, d.slot, (uint64_t )d.first_page, + (uint32_t )d.num_pages, ret); + }else { + ret =0 ; +trace_kvm_clear_dirty_log (d.slot, d.first_page, d.num_pages); + } + +/* +* After we have updated the remote dirty bitmap, we update the +* cached bitmap as well for the memslot, then if another user +* clears the same region we know we shouldn't clear it again on +* the remote otherwise it's data loss as well. +*/ +bitmap_clear (mem->dirty_bmap, bmap_start + start_delta, + size / psize); +/* This handles the NULL case well */ +g_free (bmap_clear); + +kvm_slots_unlock (kml); + +return ret; +} +static void kvm_coalesce_mmio_region (MemoryListener *listener,MemoryRegionSection *secion,hwaddr start,hwaddr size)
@@ -888,6 +1040,22 @@ static void kvm_log_sync(MemoryListener *listener, } } +static void kvm_log_clear (MemoryListener *listener, +MemoryRegionSection *section) +{ +KVMMemoryListener *kml =container_of (listener, KVMMemoryListener, listener); +int r; + + r =kvm_physical_log_clear (kml, section); +if (r <0 ) { +error_report_once ("%s: kvm log clear failed: mr=%s " +"offset=%" HWADDR_PRIx" size=%" PRIx64, __func__, + section->mr->name, section->offset_within_region, +int128_get64 (section->size)); +abort (); + } +} + static void kvm_mem_ioeventfd_add(MemoryListener *listener, MemoryRegionSection *section, bool match_data, uint64_t data,
@@ -975,6 +1143,7 @@ void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml, kml->listener.log_start = kvm_log_start; kml->listener.log_stop = kvm_log_stop; kml->listener.log_sync = kvm_log_sync; + kml->listener.log_clear = kvm_log_clear; kml->listener.priority =10 ;memory_listener_register (&kml->listener, as);
@@ -1699,6 +1868,17 @@ static int kvm_init(MachineState *ms) s->coalesced_pio = s->coalesced_mmio &&kvm_check_extension (s,KVM_CAP_COALESCED_PIO ); + s->manual_dirty_log_protect = +kvm_check_extension (s,KVM_CAP_MANUAL_DIRTY_LOG_PROTECT ); +if (s->manual_dirty_log_protect) { + ret =kvm_vm_enable_cap (s,KVM_CAP_MANUAL_DIRTY_LOG_PROTECT ,0 ,1 ); +if (ret) { +warn_report ("Trying to enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT " +"but failed. Falling back to the legacy mode. " ); + s->manual_dirty_log_protect = false; + } + } +#ifdef KVM_CAP_VCPU_EVENTS s->vcpu_events =kvm_check_extension (s,KVM_CAP_VCPU_EVENTS );#endif
accel/kvm/trace-events +1 -0
@@ -15,4 +15,5 @@ kvm_irqchip_release_virq(int virq) "virq %d" kvm_set_ioeventfd_mmio(int fd, uint64_t addr, uint32_t val, bool assign, uint32_t size, bool datamatch) "fd: %d @0x%" PRIx64 " val=0x%x assign: %d size: %d match: %d" kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint32_t val, bool assign, uint32_t size, bool datamatch) "fd: %d @0x%x val=0x%x assign: %d size: %d match: %d" kvm_set_user_memory(uint32_t slot, uint32_t flags, uint64_t guest_phys_addr, uint64_t memory_size, uint64_t userspace_addr, int ret) "Slot#%d flags=0x%x gpa=0x%"PRIx64 " size=0x%"PRIx64 " ua=0x%"PRIx64 " ret=%d" +kvm_clear_dirty_log(uint32_t slot, uint64_t start, uint32_t size) "slot#%"PRId32" start 0x%"PRIx64" size 0x%"PRIx32
Signature · 1 line
2.17.1
PATCH 11/11 migration: Split log_clear() into smaller chunks
In reply to Peter XuCurrently we are doing log_clear() right after log_sync() which mostly keeps the old behavior when log_clear() was still part of log_sync().
This patch tries to further optimize the migration log_clear() code path to split huge log_clear()s into smaller chunks.
We do this by spliting the whole guest memory region into memory chunks, whose size is decided by MigrationState.clear_bitmap_shift (an example will be given below). With that, we don't do the dirty bitmap clear operation on the remote node (e.g., KVM) when we fetch the dirty bitmap, instead we explicitly clear the dirty bitmap for the memory chunk for each of the first time we send a page in that chunk.
Here comes an example.
Assuming the guest has 64G memory, then before this patch the KVM ioctl KVM_CLEAR_DIRTY_LOG will be a single one covering 64G memory. If after the patch, let's assume when the clear bitmap shift is 18, then the memory chunk size on x86_64 will be 1UL<<18 * 4K = 1GB. Then instead of sending a big 64G ioctl, we'll send 64 small ioctls, each of the ioctl will cover 1G of the guest memory. For each of the 64 small ioctls, we'll only send if any of the page in that small chunk was going to be sent right away.
--- include/exec/ram_addr.h | 75 +++++++++++++++++++++++++++++++++++++++-- migration/migration.c | 4 +++ migration/migration.h | 27 +++++++++++++++ migration/ram.c | 44 ++++++++++++++++++++++++ migration/trace-events | 1 + 5 files changed, 149 insertions(+), 2 deletions(-)
include/exec/ram_addr.h +73 -2
@@ -50,8 +50,69 @@ struct RAMBlock {unsigned long *unsentmap;/* bitmap of already received pages in postcopy */ unsigned long *receivedmap; + +/* +* bitmap of already cleared dirty bitmap. Set this up to +* non-NULL to enable the capability to postpone and split +* clearing of dirty bitmap on the remote node (e.g., KVM). The +* bitmap will be set only when doing global sync. +* +* NOTE: this bitmap is different comparing to the other bitmaps +* in that one bit can represent multiple guest pages (which is +* decided by the `clear_bmap_shift' variable below). On +* destination side, this should always be NULL, and the variable +* `clear_bmap_shift' is meaningless. +*/ +unsigned long *clear_bmap; +uint8_t clear_bmap_shift; }; +/** +* clear_bmap_size: calculate clear bitmap size +* +* @pages: number of guest pages +* @shift: guest page number shift +* +* Returns: number of bits for the clear bitmap +*/ +static inline long clear_bmap_size (uint64_t pages,uint8_t shift) +{ +return DIV_ROUND_UP (pages,1UL << shift); +} + +/** +* clear_bmap_set: set clear bitmap for the page range +* +* @rb: the ramblock to operate on +* @start: the start page number +* @size: number of pages to set in the bitmap +* +* Returns: None +*/ +static inline void clear_bmap_set (RAMBlock *rb,uint64_t start, +uint64_t npages) +{ +uint8_t shift = rb->clear_bmap_shift; + +bitmap_set_atomic (rb->clear_bmap, start >> shift, +clear_bmap_size (npages, shift)); +} + +/** +* clear_bmap_test_and_clear: test clear bitmap for the page, clear if set +* +* @rb: the ramblock to operate on +* @page: the page number to check +* +* Returns: true if the bit was set, false otherwise +*/ +static inline bool clear_bmap_test_and_clear (RAMBlock *rb,uint64_t page) +{ +uint8_t shift = rb->clear_bmap_shift; + +return bitmap_test_and_clear_atomic (rb->clear_bmap, page >> shift,1 ); +} +static inline bool offset_in_ramblock (RAMBlock *b,ram_addr_t offset) {return (b && b->host && offset < b->used_length) ? true : false;
@@ -462,8 +523,18 @@ uint64_t cpu_physical_memory_sync_dirty_bitmap(RAMBlock *rb, } } -/* TODO: split the huge bitmap into smaller chunks */ -memory_region_clear_dirty_bitmap (rb->mr, start, length); +if (rb->clear_bmap) { +/* +* Postpone the dirty bitmap clear to the point before we +* really send the pages, also we will split the clear +* dirty procedure into smaller chunks. +*/ +clear_bmap_set (rb, start >>TARGET_PAGE_BITS , + length >>TARGET_PAGE_BITS ); + }else { +/* Slow path - still do that in a huge chunk */ +memory_region_clear_dirty_bitmap (rb->mr, start, length); + } }else {ram_addr_t offset = rb->offset;
migration/migration.c +4 -0
@@ -3348,6 +3348,8 @@ void migration_global_dump(Monitor *mon) ms->send_section_footer ? "on" : "off");monitor_printf (mon,"decompress-error-check: %s\n" , ms->decompress_error_check ?"on" :"off" ); +monitor_printf (mon,"clear-bitmap-shift: %u\n" , + ms->clear_bitmap_shift); }#define DEFINE_PROP_MIG_CAP (name, x) \
@@ -3362,6 +3364,8 @@ static Property migration_properties[] = { send_section_footer, true),DEFINE_PROP_BOOL ("decompress-error-check" , MigrationState, decompress_error_check, true), +DEFINE_PROP_UINT8 ("x-clear-bitmap-shift" , MigrationState, + clear_bitmap_shift,CLEAR_BITMAP_SHIFT_DEFAULT ),/* Migration parameters */ DEFINE_PROP_UINT8("x-compress-level", MigrationState,
migration/migration.h +27 -0
@@ -27,6 +27,23 @@ struct PostcopyBlocktimeContext;#define MIGRATION_RESUME_ACK_VALUE (1) +/* +* 1<<6=64 pages -> 256K chunk when page size is 4K. This gives us +* the benefit that all the chunks are 64 pages aligned then the +* bitmaps are always aligned to LONG. +*/ +#define CLEAR_BITMAP_SHIFT_MIN 6 +/* +* 1<<18=256K pages -> 1G chunk when page size is 4K. This is the +* default value to use if no one specified. +*/ +#define CLEAR_BITMAP_SHIFT_DEFAULT 18 +/* +* 1<<31=2G pages -> 8T chunk when page size is 4K. This should be +* big enough and make sure we won't overflow easily. +*/ +#define CLEAR_BITMAP_SHIFT_MAX 31 +/* State for the incoming migration */ struct MigrationIncomingState {QEMUFile *from_src_file;
@@ -234,6 +251,16 @@ struct MigrationState* do not trigger spurious decompression errors. */ bool decompress_error_check; + +/* +* This decides the size of guest memory chunk that will be used +* to track dirty bitmap clearing. The size of memory chunk will +* be GUEST_PAGE_SIZE << N. Say, N=0 means we will clear dirty +* bitmap for each page to send (1<<0=1); N=10 means we will clear +* dirty bitmap only once for 1<<10=1K continuous guest pages +* (which is in 4M chunk). +*/ +uint8_t clear_bitmap_shift; };void migrate_set_state (int *state,int old_state,int new_state);
migration/ram.c +44 -0
71 lines not on this page · raw message
migration/trace-events +1 -0
@@ -79,6 +79,7 @@ get_queued_page(const char *block_name, uint64_t tmp_offset, unsigned long page_ get_queued_page_not_dirty(const char *block_name, uint64_t tmp_offset, unsigned long page_abs, int sent) "%s/0x%" PRIx64 " page_abs=0x%lx (sent=%d)" migration_bitmap_sync_start(void) "" migration_bitmap_sync_end(uint64_t dirty_pages) "dirty_pages %" PRIu64 +migration_bitmap_clear_dirty(char *str, uint64_t start, uint64_t size, unsigned long page) "rb %s start 0x%"PRIx64" size 0x%"PRIx64" page 0x%lx" migration_throttle(void) "" multifd_recv(uint8_t id, uint64_t packet_num, uint32_t used, uint32_t flags, uint32_t next_packet_size) "channel %d packet number %" PRIu64 " pages %d flags 0x%x next packet size %d" multifd_recv_sync_main(long packet_num) "packet num %ld"
Signature · 1 line
2.17.1
Quoted 27 lines · Peter Xu
Design
===================I started with a naive/stupid design that I always pass all 1's to the KVM for a memory range to clear all the dirty bits within that memory range, but then I encountered guest oops - it's simply because we can't clear any dirty bit from QEMU if we are not _sure_ that the bit is dirty in the kernel. Otherwise we might accidentally clear a bit that we don't even know of (e.g., the bit was clear in migration's dirty bitmap in QEMU) but actually that page was just being written so QEMU will never remember to migrate that new page again.
The new design is focused on a dirty bitmap cache within the QEMU kvm layer (which is per kvm memory slot). With that we know what's dirty in the kernel previously (note! the kernel bitmap is still growing all the time so the cache will only be a subset of the realtime kernel bitmap but that's far enough for us) and with that we'll be sure to not accidentally clear unknown dirty pages.
With this method, we can also avoid race when multiple users (e.g.,
DIRTY_MEMORY_VGAandDIRTY_MEMORY_MIGRATION) want to clear the bit for multiple time. If without the kvm memory slot cached dirty bitmap we won't be able to know which bit has been cleared and then if we send the CLEAR operation upon the same bit twice (or more) we can still face the same issue to clear something accidentally while we shouldn't.
Quoted 3 lines
Summary: we really need to be careful on what bit to clear otherwise we can face anything after the migration completes. And I hope this series has considered all about this.
The disadvantage of this is that you won't clear in the kernel those dirty bits that come from other sources (e.g. vhost or address_space_map). This can lead to double-copying of pages.
Migration already makes a local copy in rb->bmap, and memory_region_snapshot_and_clear_dirty can also do the clear. Would it be possible to invoke the clear using rb->bmap instead of the KVMSlot's new bitmap?
Thanks,
Paolo
Quoted 4 lines
Besides the new KVM cache layer and the new ioctl support, this series introduced the
memory_region_clear_dirty_bitmap()in the memory API layer to allow clearing dirty bits of a specific memory range within the memory region.
Quoted 32 lines · Paolo Bonzini
Quoted 27 lines · Peter Xu
Design
===================I started with a naive/stupid design that I always pass all 1's to the KVM for a memory range to clear all the dirty bits within that memory range, but then I encountered guest oops - it's simply because we can't clear any dirty bit from QEMU if we are not _sure_ that the bit is dirty in the kernel. Otherwise we might accidentally clear a bit that we don't even know of (e.g., the bit was clear in migration's dirty bitmap in QEMU) but actually that page was just being written so QEMU will never remember to migrate that new page again.
The new design is focused on a dirty bitmap cache within the QEMU kvm layer (which is per kvm memory slot). With that we know what's dirty in the kernel previously (note! the kernel bitmap is still growing all the time so the cache will only be a subset of the realtime kernel bitmap but that's far enough for us) and with that we'll be sure to not accidentally clear unknown dirty pages.
With this method, we can also avoid race when multiple users (e.g.,
DIRTY_MEMORY_VGAandDIRTY_MEMORY_MIGRATION) want to clear the bit for multiple time. If without the kvm memory slot cached dirty bitmap we won't be able to know which bit has been cleared and then if we send the CLEAR operation upon the same bit twice (or more) we can still face the same issue to clear something accidentally while we shouldn't.Quoted 3 lines
Summary: we really need to be careful on what bit to clear otherwise we can face anything after the migration completes. And I hope this series has considered all about this.
Quoted 8 lines
The disadvantage of this is that you won't clear in the kernel those dirty bits that come from other sources (e.g. vhost or
address_space_map). This can lead to double-copying of pages.Migration already makes a local copy in
rb->bmap, andmemory_region_snapshot_and_clear_dirtycan also do the clear. Would it be possible to invoke the clear usingrb->bmapinstead of the KVMSlot's new bitmap?
Do you mean to invoke the clear notifications and deliver the bitmap from the very top of the stack (for VGA it would probably be the DirtyBitmapSnapshot, for MIGRATION it's the rb->bmap)?
Actually that's what I did in the first version before I post the work but I noticed that there seems to have a race condition with the design. The problem is we have multiple copies of the same dirty bitmap from KVM and the race can happen with those multiple users (bitmaps of the users can be a merged version containing KVM and other sources like vhost, address_space_map, etc. but let's just make it simpler to not have them yet). Here's one possible race condition (assuming migration has started):
(1) page P is written by the guest (let's version its data as P1), its dirty bit D is set in KVM
(2) QEMU sync dirty log, fetch D for page P (which is set). D is not cleared in KVM due to MANUAL_PROTECT cap.
(3) QEMU copies the D bit to all users of dirty bmap (MIGRATION and VGA as example).
(4) MIGRATION code collects bit D in migration bitmap, clear it, send CLEAR to KVM to clear the bit on remote (then D in KVM is cleared too), send the page P with content P1 to destination.
CAUTION: when reach here VGA bitmap still has the D bit set.
(5) page P is written by the guest again (let's version its data as P2), bit D set again in KVM.
(6) VGA code collectes bit D (note! we haven't synced the log again so this is the _old_ dirty bit came from step 3 above) in VGA bitmap, clear it, send CLEAR to KVM to clear the bit on remote. Then D bit in KVM is cleared again. Until here, D bit in all bitmaps are cleared (MIGRATION, VGA, KVM).
(7) page P is never written again until migration completes (no matter whether we sync again D bit will be cleared).
(8) On destination VM page P will contain content P1 rather than P2, this is data loss...
After I noticed it, I added the kvm bitmap layer with the lock to make sure we won't send the 2nd clear at step (6) above because with the per memslot bitmap we will be able to know that we've already cleared a bit by one user so we shouldn't clear it again (as long as we don't SYNC again). However later on I just noticed maybe I don't even need the top layer bitmap at all, because it seems to not bring much benefit but instead it'll bring lots of more complexity - I'll need to do bitmap convertions (because upper layer bitmaps are per ramblock, and kvm bitmaps are per memslot), not to mention I'll need to AND the two bitmaps (upper layer bitmap, and the QEMU kvm cached bitmap) to avoid the above race. So at last I simply removed the top layer bitmap and only keep the per memslot bitmap as what this series did.
Thanks,
Signature · 1 line
Peter Xu
Quoted 9 lines · Peter Xu
Quoted 8 lines
The disadvantage of this is that you won't clear in the kernel those dirty bits that come from other sources (e.g. vhost or
address_space_map). This can lead to double-copying of pages.Migration already makes a local copy in
rb->bmap, andmemory_region_snapshot_and_clear_dirtycan also do the clear. Would it be possible to invoke the clear usingrb->bmapinstead of the KVMSlot's new bitmap?
Quoted 7 lines
Actually that's what I did in the first version before I post the work but I noticed that there seems to have a race condition with the design. The problem is we have multiple copies of the same dirty bitmap from KVM and the race can happen with those multiple users (bitmaps of the users can be a merged version containing KVM and other sources like vhost,
address_space_map, etc. but let's just make it simpler to not have them yet).
I see now. And in fact the same double-copying inefficiency happens already without this series, so you are improving the situation anyway.
Have you done any kind of benchmarking already?
Thanks,
Paolo
Quoted 28 lines
(1) page P is written by the guest (let's version its data as P1), its dirty bit D is set in KVM
(2) QEMU sync dirty log, fetch D for page P (which is set). D is not cleared in KVM due to
MANUAL_PROTECTcap.(3) QEMU copies the D bit to all users of dirty bmap (MIGRATION and VGA as example).
(4) MIGRATION code collects bit D in migration bitmap, clear it, send CLEAR to KVM to clear the bit on remote (then D in KVM is cleared too), send the page P with content P1 to destination.
CAUTION: when reach here VGA bitmap still has the D bit set.(5) page P is written by the guest again (let's version its data as P2), bit D set again in KVM.
(6) VGA code collectes bit D (note! we haven't synced the log again so this is the _old_ dirty bit came from step 3 above) in VGA bitmap, clear it, send CLEAR to KVM to clear the bit on remote. Then D bit in KVM is cleared again. Until here, D bit in all bitmaps are cleared (MIGRATION, VGA, KVM).
(7) page P is never written again until migration completes (no matter whether we sync again D bit will be cleared).
(8) On destination VM page P will contain content P1 rather than P2, this is data loss...
Quoted 18 lines · Paolo Bonzini
Quoted 9 lines · Peter Xu
Quoted 8 lines
The disadvantage of this is that you won't clear in the kernel those dirty bits that come from other sources (e.g. vhost or
address_space_map). This can lead to double-copying of pages.Migration already makes a local copy in
rb->bmap, andmemory_region_snapshot_and_clear_dirtycan also do the clear. Would it be possible to invoke the clear usingrb->bmapinstead of the KVMSlot's new bitmap?Quoted 7 lines
Actually that's what I did in the first version before I post the work but I noticed that there seems to have a race condition with the design. The problem is we have multiple copies of the same dirty bitmap from KVM and the race can happen with those multiple users (bitmaps of the users can be a merged version containing KVM and other sources like vhost,
address_space_map, etc. but let's just make it simpler to not have them yet).
Quoted 4 lines
I see now. And in fact the same double-copying inefficiency happens already without this series, so you are improving the situation anyway.
Have you done any kind of benchmarking already?
Not yet. I posted the series for some initial reviews first before moving on with performance tests.
My plan of the test scenario could be:
- find a guest with relatively large memory (I would guess it is better to have memory like 64G or even more to make some big difference)
- run random dirty memory workload upon most of the mem, with dirty rate X Bps.
- setup the migration bandwidth to Y Bps (Y should be bigger than X but not that big. One could be X=800M and Y=1G to emulate 10G nic with a workload that we can still converge with precopy only) and start precopy migration.
- measure total migration time with CLEAR_LOG on & off. We should expect the guest to have these with CLEAR_LOG: (1) not hang during log_sync, and (2) migration should complete faster.
Does above test plan makes sense?
If both the QEMU/KVM changes looks ok in general, I can at least try this on some smaller guests (I can manage ~10G mem guests with my own hosts, but I can also try to find some bigger ones).
Thanks,
Signature · 1 line
Peter Xu
Quoted 36 lines · Peter Xu
Quoted 18 lines · Paolo Bonzini
Quoted 9 lines · Peter Xu
Quoted 8 lines
The disadvantage of this is that you won't clear in the kernel those dirty bits that come from other sources (e.g. vhost or
address_space_map). This can lead to double-copying of pages.Migration already makes a local copy in
rb->bmap, andmemory_region_snapshot_and_clear_dirtycan also do the clear. Would it be possible to invoke the clear usingrb->bmapinstead of the KVMSlot's new bitmap?Quoted 7 lines
Actually that's what I did in the first version before I post the work but I noticed that there seems to have a race condition with the design. The problem is we have multiple copies of the same dirty bitmap from KVM and the race can happen with those multiple users (bitmaps of the users can be a merged version containing KVM and other sources like vhost,
address_space_map, etc. but let's just make it simpler to not have them yet).Quoted 4 lines
I see now. And in fact the same double-copying inefficiency happens already without this series, so you are improving the situation anyway.
Have you done any kind of benchmarking already?
Not yet. I posted the series for some initial reviews first before moving on with performance tests.
My plan of the test scenario could be:
- find a guest with relatively large memory (I would guess it is better to have memory like 64G or even more to make some big difference)
- run random dirty memory workload upon most of the mem, with dirty rate X Bps.
Quoted 8 lines
- setup the migration bandwidth to Y Bps (Y should be bigger than X but not that big. One could be X=800M and Y=1G to emulate 10G nic with a workload that we can still converge with precopy only) and start precopy migration.
- measure total migration time with
CLEAR_LOGon & off. We should expect the guest to have these withCLEAR_LOG: (1) not hang duringlog_sync, and (2) migration should complete faster.
Some updates on performance numbers.
Summary: the ideal case below shows ~40% or even more time reduced to migrate the same VM with the same workload. In other words, it could be seen as ~40% faster than before.
Test environment: 13G guest, 10G test mem (so I leave 3G untouched), dirty rate 900MB/s, BW 10Gbps to emulate ixgbe, downtime 100ms.
IO pattern: I pre-fault all the 10G mem then I do random writes (with command "mig_mon mm_dirty 10240 900 random" [1]) upon the test memory with a constant dirty rate (900MB/s, as mentioned). Then I migrate during the IOs.
Here's the total migration time of such VM (for each scenario, I run the migration 5 times then I get an average migration total time used):
|--------------+---------------------+-------------| | scenario | migration times (s) | average (s) | |--------------+---------------------+-------------| | no CLEAR_LOG | 55, 54, 56, 74, 54 | 58 | | 1G chunk | 40, 39, 41, 39, 40 | 40 | | 128M chunk | 38, 40, 37, 40, 38 | 38 | | 16M chunk | 42, 40, 38, 41, 38 | 39 | | 1M chunk | 37, 40, 36, 40, 39 | 38 | |--------------+---------------------+-------------|
The first "no CLEAR_LOG" means the master branch which still uses the GET_DIRTY only. The latter four scenarios are all with the new CLEAR_LOG interface, aka, this series. The test result shows that 128M chunk size seems to be a good default value instead of 1G (which this series used). I'll adjust that accordingly when I post the next version.
[1] https://github.com/xzpeter/clibs/blob/master/bsd/mig_mon/mig_mon.c
Regards,
Signature · 1 line
Peter Xu