| Message ID | 20180228072558.7434-6-haozhong.zhang@intel.com |
|---|---|
| State | New |
| Headers | show |
| Series | nvdimm: guarantee persistence of QEMU writes to persistent memory | expand |
* Haozhong Zhang (haozhong.zhang@intel.com) wrote: > When loading a zero page, check whether it will be loaded to > persistent memory If yes, load it by libpmem function > pmem_memset_nodrain(). Combined with a call to pmem_drain() at the > end of RAM loading, we can guarantee all those zero pages are > persistently loaded. > > Depending on the host HW/SW configurations, pmem_drain() can be > "sfence". Therefore, we do not call pmem_drain() after each > pmem_memset_nodrain(), or use pmem_memset_persist() (equally > pmem_memset_nodrain() + pmem_drain()), in order to avoid unnecessary > overhead. > > Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> I'm still thinking this is way too invasive; especially the next patch that touches qemu_file. One thing that would help a little, but not really enough, would be to define a : struct MemOps { void (*copy)(like a memcpy); void (*set)(like a memset); } then you could have: struct MemOps normalops = {memcpy, memset}; struct MemOps pmem_nodrain_ops = { pmem_memcopy_nodrain, pmem_memset_nodrain }; then things like ram_handle_compressed would be: void ram_handle_compressed(void *host, uint8_t ch, uint64_t size, const struct MemOps *mem) { if (ch != 0 || !is_zero_range(host, size)) { mem->set(host, ch,size); } } which means the change is pretty tiny to each function. > diff --git a/migration/rdma.c b/migration/rdma.c > index da474fc19f..573bcd2cb0 100644 > --- a/migration/rdma.c > +++ b/migration/rdma.c > @@ -3229,7 +3229,7 @@ static int qemu_rdma_registration_handle(QEMUFile *f, void *opaque) > host_addr = block->local_host_addr + > (comp->offset - block->offset); > > - ram_handle_compressed(host_addr, comp->value, comp->length); > + ram_handle_compressed(host_addr, comp->value, comp->length, false); Is that right? Is RDMA not allowed to work on PMEM? (and anyway this call is a normal clear rather than an actual RDMA op). Dave > break; > > case RDMA_CONTROL_REGISTER_FINISHED: > diff --git a/stubs/pmem.c b/stubs/pmem.c > index 03d990e571..a65b3bfc6b 100644 > --- a/stubs/pmem.c > +++ b/stubs/pmem.c > @@ -17,3 +17,12 @@ void *pmem_memcpy_persist(void *pmemdest, const void *src, size_t len) > { > return memcpy(pmemdest, src, len); > } > + > +void *pmem_memset_nodrain(void *pmemdest, int c, size_t len) > +{ > + return memset(pmemdest, c, len); > +} > + > +void pmem_drain(void) > +{ > +} > -- > 2.14.1 > -- Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
On 03/29/18 19:59 +0100, Dr. David Alan Gilbert wrote: > * Haozhong Zhang (haozhong.zhang@intel.com) wrote: > > When loading a zero page, check whether it will be loaded to > > persistent memory If yes, load it by libpmem function > > pmem_memset_nodrain(). Combined with a call to pmem_drain() at the > > end of RAM loading, we can guarantee all those zero pages are > > persistently loaded. > > > > Depending on the host HW/SW configurations, pmem_drain() can be > > "sfence". Therefore, we do not call pmem_drain() after each > > pmem_memset_nodrain(), or use pmem_memset_persist() (equally > > pmem_memset_nodrain() + pmem_drain()), in order to avoid unnecessary > > overhead. > > > > Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> > > I'm still thinking this is way too invasive; especially the next patch > that touches qemu_file. > > One thing that would help a little, but not really enough, would be > to define a : > > struct MemOps { > void (*copy)(like a memcpy); > void (*set)(like a memset); > } > > then you could have: > > struct MemOps normalops = {memcpy, memset}; > struct MemOps pmem_nodrain_ops = { pmem_memcopy_nodrain, pmem_memset_nodrain }; > > then things like ram_handle_compressed would be: > > void ram_handle_compressed(void *host, uint8_t ch, uint64_t size, const struct MemOps *mem) > { > if (ch != 0 || !is_zero_range(host, size)) { > mem->set(host, ch,size); > } > } > > which means the change is pretty tiny to each function. This looks much better than mine. I'm also considering Stefan's comment that flushing at the end of all memory migration rather than invasively change every types of copy in migration stream. We are going to perform some microbenchmarks on the real hardware, and then decide which way to take. > > > diff --git a/migration/rdma.c b/migration/rdma.c > > index da474fc19f..573bcd2cb0 100644 > > --- a/migration/rdma.c > > +++ b/migration/rdma.c > > @@ -3229,7 +3229,7 @@ static int qemu_rdma_registration_handle(QEMUFile *f, void *opaque) > > host_addr = block->local_host_addr + > > (comp->offset - block->offset); > > > > - ram_handle_compressed(host_addr, comp->value, comp->length); > > + ram_handle_compressed(host_addr, comp->value, comp->length, false); > > Is that right? Is RDMA not allowed to work on PMEM? > (and anyway this call is a normal clear rather than an actual RDMA op). > Well, this patch exclude RMDA case intentionally. Once it's clear how to guarantee the persistence of remote PMEM write in RDMA, we will propose additional patch to add support in QEMU. Thanks, Haozhong > Dave > > > break; > > > > case RDMA_CONTROL_REGISTER_FINISHED: > > diff --git a/stubs/pmem.c b/stubs/pmem.c > > index 03d990e571..a65b3bfc6b 100644 > > --- a/stubs/pmem.c > > +++ b/stubs/pmem.c > > @@ -17,3 +17,12 @@ void *pmem_memcpy_persist(void *pmemdest, const void *src, size_t len) > > { > > return memcpy(pmemdest, src, len); > > } > > + > > +void *pmem_memset_nodrain(void *pmemdest, int c, size_t len) > > +{ > > + return memset(pmemdest, c, len); > > +} > > + > > +void pmem_drain(void) > > +{ > > +} > > -- > > 2.14.1 > > > -- > Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
diff --git a/include/qemu/pmem.h b/include/qemu/pmem.h index 16f5b2653a..ce96379f3c 100644 --- a/include/qemu/pmem.h +++ b/include/qemu/pmem.h @@ -17,6 +17,8 @@ #else /* !CONFIG_LIBPMEM */ void *pmem_memcpy_persist(void *pmemdest, const void *src, size_t len); +void *pmem_memset_nodrain(void *pmemdest, int c, size_t len); +void pmem_drain(void); #endif /* CONFIG_LIBPMEM */ diff --git a/migration/ram.c b/migration/ram.c index 5e33e5cc79..3904ceee79 100644 --- a/migration/ram.c +++ b/migration/ram.c @@ -51,6 +51,7 @@ #include "qemu/rcu_queue.h" #include "migration/colo.h" #include "migration/block.h" +#include "qemu/pmem.h" /***********************************************************/ /* ram save/restore */ @@ -2479,11 +2480,16 @@ static inline void *host_from_ram_block_offset(RAMBlock *block, * @host: host address for the zero page * @ch: what the page is filled from. We only support zero * @size: size of the zero page + * @is_pmem: whether @host is in the persistent memory */ -void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) +void ram_handle_compressed(void *host, uint8_t ch, uint64_t size, bool is_pmem) { if (ch != 0 || !is_zero_range(host, size)) { - memset(host, ch, size); + if (!is_pmem) { + memset(host, ch, size); + } else { + pmem_memset_nodrain(host, ch, size); + } } } @@ -2839,6 +2845,7 @@ static int ram_load(QEMUFile *f, void *opaque, int version_id) bool postcopy_running = postcopy_is_running(); /* ADVISE is earlier, it shows the source has the postcopy capability on */ bool postcopy_advised = postcopy_is_advised(); + bool need_pmem_drain = false; seq_iter++; @@ -2864,6 +2871,8 @@ static int ram_load(QEMUFile *f, void *opaque, int version_id) ram_addr_t addr, total_ram_bytes; void *host = NULL; uint8_t ch; + RAMBlock *block = NULL; + bool is_pmem = false; addr = qemu_get_be64(f); flags = addr & ~TARGET_PAGE_MASK; @@ -2880,7 +2889,7 @@ static int ram_load(QEMUFile *f, void *opaque, int version_id) if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { - RAMBlock *block = ram_block_from_stream(f, flags); + block = ram_block_from_stream(f, flags); host = host_from_ram_block_offset(block, addr); if (!host) { @@ -2890,6 +2899,9 @@ static int ram_load(QEMUFile *f, void *opaque, int version_id) } ramblock_recv_bitmap_set(block, host); trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); + + is_pmem = ramblock_is_pmem(block); + need_pmem_drain = need_pmem_drain || is_pmem; } switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { @@ -2943,7 +2955,7 @@ static int ram_load(QEMUFile *f, void *opaque, int version_id) case RAM_SAVE_FLAG_ZERO: ch = qemu_get_byte(f); - ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); + ram_handle_compressed(host, ch, TARGET_PAGE_SIZE, is_pmem); break; case RAM_SAVE_FLAG_PAGE: @@ -2986,6 +2998,11 @@ static int ram_load(QEMUFile *f, void *opaque, int version_id) } wait_for_decompress_done(); + + if (need_pmem_drain) { + pmem_drain(); + } + rcu_read_unlock(); trace_ram_load_complete(ret, seq_iter); return ret; diff --git a/migration/ram.h b/migration/ram.h index f3a227b4fc..18934ae9e4 100644 --- a/migration/ram.h +++ b/migration/ram.h @@ -57,7 +57,7 @@ int ram_postcopy_send_discard_bitmap(MigrationState *ms); int ram_discard_range(const char *block_name, uint64_t start, size_t length); int ram_postcopy_incoming_init(MigrationIncomingState *mis); -void ram_handle_compressed(void *host, uint8_t ch, uint64_t size); +void ram_handle_compressed(void *host, uint8_t ch, uint64_t size, bool is_pmem); int ramblock_recv_bitmap_test(RAMBlock *rb, void *host_addr); void ramblock_recv_bitmap_set(RAMBlock *rb, void *host_addr); diff --git a/migration/rdma.c b/migration/rdma.c index da474fc19f..573bcd2cb0 100644 --- a/migration/rdma.c +++ b/migration/rdma.c @@ -3229,7 +3229,7 @@ static int qemu_rdma_registration_handle(QEMUFile *f, void *opaque) host_addr = block->local_host_addr + (comp->offset - block->offset); - ram_handle_compressed(host_addr, comp->value, comp->length); + ram_handle_compressed(host_addr, comp->value, comp->length, false); break; case RDMA_CONTROL_REGISTER_FINISHED: diff --git a/stubs/pmem.c b/stubs/pmem.c index 03d990e571..a65b3bfc6b 100644 --- a/stubs/pmem.c +++ b/stubs/pmem.c @@ -17,3 +17,12 @@ void *pmem_memcpy_persist(void *pmemdest, const void *src, size_t len) { return memcpy(pmemdest, src, len); } + +void *pmem_memset_nodrain(void *pmemdest, int c, size_t len) +{ + return memset(pmemdest, c, len); +} + +void pmem_drain(void) +{ +}
When loading a zero page, check whether it will be loaded to persistent memory If yes, load it by libpmem function pmem_memset_nodrain(). Combined with a call to pmem_drain() at the end of RAM loading, we can guarantee all those zero pages are persistently loaded. Depending on the host HW/SW configurations, pmem_drain() can be "sfence". Therefore, we do not call pmem_drain() after each pmem_memset_nodrain(), or use pmem_memset_persist() (equally pmem_memset_nodrain() + pmem_drain()), in order to avoid unnecessary overhead. Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> --- include/qemu/pmem.h | 2 ++ migration/ram.c | 25 +++++++++++++++++++++---- migration/ram.h | 2 +- migration/rdma.c | 2 +- stubs/pmem.c | 9 +++++++++ 5 files changed, 34 insertions(+), 6 deletions(-)