Message ID | 20230117220914.2062125-16-peterx@redhat.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | migration: Support hugetlb doublemaps | expand |
Peter Xu <peterx@redhat.com> wrote: > When a ramblock is backed by hugetlbfs and the user specified using > double-map feature, we trap the faults on these regions using minor mode. > Teach QEMU about that. > > Add some sanity check on the fault flags when receiving a uffd message. > For minor fault trapped ranges, we should always see the MINOR flag set, > while when using generic missing faults we should never see it. > > Signed-off-by: Peter Xu <peterx@redhat.com> Reviewed-by: Juan Quintela <quintela@redhat.com> > - if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { Does qemu have a macro to do this bitmap handling? > { > MigrationIncomingState *mis = opaque; > struct uffd_msg msg; > + uint64_t address; > int ret; > size_t index; > RAMBlock *rb = NULL; > @@ -945,6 +980,7 @@ static void *postcopy_ram_fault_thread(void *opaque) > } > > while (true) { > + bool use_minor_fault, minor_flag; I think that something on the lines of: bool src_minor_fault, dst_minor_fault; will make things simpler. Reviewing, I have to go back to definition place to know which is which. > ram_addr_t rb_offset; > int poll_result; > > @@ -1022,22 +1058,37 @@ static void *postcopy_ram_fault_thread(void *opaque) > break; > } > > - rb_offset = ROUND_DOWN(rb_offset, migration_ram_pagesize(rb)); > - trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address, > - qemu_ram_get_idstr(rb), > - rb_offset, > - msg.arg.pagefault.feat.ptid); > - mark_postcopy_blocktime_begin( > - (uintptr_t)(msg.arg.pagefault.address), > - msg.arg.pagefault.feat.ptid, rb); > + address = ROUND_DOWN(msg.arg.pagefault.address, > + migration_ram_pagesize(rb)); > + use_minor_fault = postcopy_use_minor_fault(rb); > + minor_flag = !!(msg.arg.pagefault.flags & > + UFFD_PAGEFAULT_FLAG_MINOR); > > + /* > + * Do sanity check on the message flags to make sure this is > + * the one we expect to receive. When using minor fault on > + * this ramblock, it should _always_ be set; when not using > + * minor fault, it should _never_ be set. > + */ > + if (use_minor_fault ^ minor_flag) { > + error_report("%s: Unexpected page fault flags (0x%"PRIx64") " > + "for address 0x%"PRIx64" (mode=%s)", __func__, > + (uint64_t)msg.arg.pagefault.flags, > + (uint64_t)msg.arg.pagefault.address, > + use_minor_fault ? "MINOR" : "MISSING"); > + } > + > + trace_postcopy_ram_fault_thread_request( > + address, qemu_ram_get_idstr(rb), rb_offset, > + msg.arg.pagefault.feat.ptid); > + mark_postcopy_blocktime_begin( > + (uintptr_t)(address), msg.arg.pagefault.feat.ptid, rb); > retry: > /* > * Send the request to the source - we want to request one > * of our host page sizes (which is >= TPS) > */ > - ret = postcopy_request_page(mis, rb, rb_offset, > - msg.arg.pagefault.address); > + ret = postcopy_request_page(mis, rb, rb_offset, address); This is the only change that I find 'problematic'. On old code, rb_offset has been ROUND_DOWN, on new code it is not. On old code we pass msg.arg.pagefault.address, now we use ROUND_DOW(msg.arg.pagefault.address, mighration_ram_pagesize(rb)). > if (ret) { > /* May be network failure, try to wait for recovery */ > postcopy_pause_fault_thread(mis); > @@ -1694,3 +1745,13 @@ void *postcopy_preempt_thread(void *opaque) > > return NULL; > } > + > +/* > + * Whether we should use MINOR fault to trap page faults? It will be used > + * when doublemap is enabled on hugetlbfs. The default value will be > + * false, which means we'll keep using the legacy MISSING faults. > + */ > +bool postcopy_use_minor_fault(RAMBlock *rb) > +{ > + return migrate_hugetlb_doublemap() && qemu_ram_is_hugetlb(rb); > +} Are you planing using this function outside postocpy-ram.c? Otherwise if you move up its definition you can make it static and drop the header change. Later, Juan.
On Mon, Jan 30, 2023 at 06:45:20AM +0100, Juan Quintela wrote: > Peter Xu <peterx@redhat.com> wrote: > > When a ramblock is backed by hugetlbfs and the user specified using > > double-map feature, we trap the faults on these regions using minor mode. > > Teach QEMU about that. > > > > Add some sanity check on the fault flags when receiving a uffd message. > > For minor fault trapped ranges, we should always see the MINOR flag set, > > while when using generic missing faults we should never see it. > > > > Signed-off-by: Peter Xu <peterx@redhat.com> > > Reviewed-by: Juan Quintela <quintela@redhat.com> > > > > > - if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { > > Does qemu have a macro to do this bitmap handling? Not yet that's suitable. It's open coded like this in many places of postcopy. One thing close enough is bitmap_test_and_clear() but too heavy. > > > { > > MigrationIncomingState *mis = opaque; > > struct uffd_msg msg; > > + uint64_t address; > > int ret; > > size_t index; > > RAMBlock *rb = NULL; > > @@ -945,6 +980,7 @@ static void *postcopy_ram_fault_thread(void *opaque) > > } > > > > while (true) { > > + bool use_minor_fault, minor_flag; > > I think that something on the lines of: > bool src_minor_fault, dst_minor_fault; > > will make things simpler. Reviewing, I have to go back to definition > place to know which is which. These two values represents "what we expect" and "what we got from the message", so the only thing is I'm not sure whether src/dst matches the best here. How about "expect_minor_fault" and "has_minor_fault" instead? > > > ram_addr_t rb_offset; > > int poll_result; > > > > @@ -1022,22 +1058,37 @@ static void *postcopy_ram_fault_thread(void *opaque) > > break; > > } > > > > - rb_offset = ROUND_DOWN(rb_offset, migration_ram_pagesize(rb)); > > - trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address, > > - qemu_ram_get_idstr(rb), > > - rb_offset, > > - msg.arg.pagefault.feat.ptid); > > - mark_postcopy_blocktime_begin( > > - (uintptr_t)(msg.arg.pagefault.address), > > - msg.arg.pagefault.feat.ptid, rb); > > + address = ROUND_DOWN(msg.arg.pagefault.address, > > + migration_ram_pagesize(rb)); > > + use_minor_fault = postcopy_use_minor_fault(rb); > > + minor_flag = !!(msg.arg.pagefault.flags & > > + UFFD_PAGEFAULT_FLAG_MINOR); > > > > + /* > > + * Do sanity check on the message flags to make sure this is > > + * the one we expect to receive. When using minor fault on > > + * this ramblock, it should _always_ be set; when not using > > + * minor fault, it should _never_ be set. > > + */ > > + if (use_minor_fault ^ minor_flag) { > > + error_report("%s: Unexpected page fault flags (0x%"PRIx64") " > > + "for address 0x%"PRIx64" (mode=%s)", __func__, > > + (uint64_t)msg.arg.pagefault.flags, > > + (uint64_t)msg.arg.pagefault.address, > > + use_minor_fault ? "MINOR" : "MISSING"); > > + } > > + > > + trace_postcopy_ram_fault_thread_request( > > + address, qemu_ram_get_idstr(rb), rb_offset, > > + msg.arg.pagefault.feat.ptid); > > + mark_postcopy_blocktime_begin( > > + (uintptr_t)(address), msg.arg.pagefault.feat.ptid, rb); > > retry: > > /* > > * Send the request to the source - we want to request one > > * of our host page sizes (which is >= TPS) > > */ > > - ret = postcopy_request_page(mis, rb, rb_offset, > > - msg.arg.pagefault.address); > > + ret = postcopy_request_page(mis, rb, rb_offset, address); > > This is the only change that I find 'problematic'. > On old code, rb_offset has been ROUND_DOWN, on new code it is not. > On old code we pass msg.arg.pagefault.address, now we use > ROUND_DOW(msg.arg.pagefault.address, mighration_ram_pagesize(rb)). Thanks for spotting such a detail even for a RFC series. :) It's actually rounded down to target psize, here since we're in postcopy we should require target psize equals to host psize (or I bet it won't really work at all). So the relevant rounddown was actually done here: rb = qemu_ram_block_from_host( (void *)(uintptr_t)msg.arg.pagefault.address, true, &rb_offset); In which there's: *offset = (host - block->host); if (round_offset) { *offset &= TARGET_PAGE_MASK; } So when I rework that chunk of code I directly dropped the ROUND_DOWN() because I find it duplicated. > > > if (ret) { > > /* May be network failure, try to wait for recovery */ > > postcopy_pause_fault_thread(mis); > > @@ -1694,3 +1745,13 @@ void *postcopy_preempt_thread(void *opaque) > > > > return NULL; > > } > > + > > +/* > > + * Whether we should use MINOR fault to trap page faults? It will be used > > + * when doublemap is enabled on hugetlbfs. The default value will be > > + * false, which means we'll keep using the legacy MISSING faults. > > + */ > > +bool postcopy_use_minor_fault(RAMBlock *rb) > > +{ > > + return migrate_hugetlb_doublemap() && qemu_ram_is_hugetlb(rb); > > +} > > Are you planing using this function outside postocpy-ram.c? Otherwise > if you move up its definition you can make it static and drop the header > change. Yes, it'll be further used in ram.c later in the patch "migration: Rework ram discard logic for hugetlb double-map" right below. Thanks,
Peter Xu <peterx@redhat.com> wrote: > On Mon, Jan 30, 2023 at 06:45:20AM +0100, Juan Quintela wrote: >> Peter Xu <peterx@redhat.com> wrote: >> > When a ramblock is backed by hugetlbfs and the user specified using >> > double-map feature, we trap the faults on these regions using minor mode. >> > Teach QEMU about that. >> > >> > Add some sanity check on the fault flags when receiving a uffd message. >> > For minor fault trapped ranges, we should always see the MINOR flag set, >> > while when using generic missing faults we should never see it. >> > >> > Signed-off-by: Peter Xu <peterx@redhat.com> >> >> Reviewed-by: Juan Quintela <quintela@redhat.com> >> >> >> >> > - if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { >> >> Does qemu have a macro to do this bitmap handling? > > Not yet that's suitable. It's open coded like this in many places of > postcopy. One thing close enough is bitmap_test_and_clear() but too heavy. > >> >> > { >> > MigrationIncomingState *mis = opaque; >> > struct uffd_msg msg; >> > + uint64_t address; >> > int ret; >> > size_t index; >> > RAMBlock *rb = NULL; >> > @@ -945,6 +980,7 @@ static void *postcopy_ram_fault_thread(void *opaque) >> > } >> > >> > while (true) { >> > + bool use_minor_fault, minor_flag; >> >> I think that something on the lines of: >> bool src_minor_fault, dst_minor_fault; >> >> will make things simpler. Reviewing, I have to go back to definition >> place to know which is which. > > These two values represents "what we expect" and "what we got from the > message", so the only thing is I'm not sure whether src/dst matches the > best here. > > How about "expect_minor_fault" and "has_minor_fault" instead? Perfect with me. >> > /* >> > * Send the request to the source - we want to request one >> > * of our host page sizes (which is >= TPS) >> > */ >> > - ret = postcopy_request_page(mis, rb, rb_offset, >> > - msg.arg.pagefault.address); >> > + ret = postcopy_request_page(mis, rb, rb_offset, address); >> >> This is the only change that I find 'problematic'. >> On old code, rb_offset has been ROUND_DOWN, on new code it is not. >> On old code we pass msg.arg.pagefault.address, now we use >> ROUND_DOW(msg.arg.pagefault.address, mighration_ram_pagesize(rb)). > > Thanks for spotting such a detail even for a RFC series. :) > > It's actually rounded down to target psize, here since we're in postcopy we > should require target psize equals to host psize (or I bet it won't really > work at all). So the relevant rounddown was actually done here: > > rb = qemu_ram_block_from_host( > (void *)(uintptr_t)msg.arg.pagefault.address, > true, &rb_offset); > > In which there's: > > *offset = (host - block->host); > if (round_offset) { > *offset &= TARGET_PAGE_MASK; > } > > So when I rework that chunk of code I directly dropped the ROUND_DOWN() > because I find it duplicated. Ok. > >> >> > if (ret) { >> > /* May be network failure, try to wait for recovery */ >> > postcopy_pause_fault_thread(mis); >> > @@ -1694,3 +1745,13 @@ void *postcopy_preempt_thread(void *opaque) >> > >> > return NULL; >> > } >> > + >> > +/* >> > + * Whether we should use MINOR fault to trap page faults? It will be used >> > + * when doublemap is enabled on hugetlbfs. The default value will be >> > + * false, which means we'll keep using the legacy MISSING faults. >> > + */ >> > +bool postcopy_use_minor_fault(RAMBlock *rb) >> > +{ >> > + return migrate_hugetlb_doublemap() && qemu_ram_is_hugetlb(rb); >> > +} >> >> Are you planing using this function outside postocpy-ram.c? Otherwise >> if you move up its definition you can make it static and drop the header >> change. > > Yes, it'll be further used in ram.c later in the patch "migration: Rework > ram discard logic for hugetlb double-map" right below. Aha. Thanks.
diff --git a/migration/postcopy-ram.c b/migration/postcopy-ram.c index acae1dc6ae..86ff73c2c0 100644 --- a/migration/postcopy-ram.c +++ b/migration/postcopy-ram.c @@ -325,12 +325,25 @@ static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis) if (qemu_real_host_page_size() != ram_pagesize_summary()) { bool have_hp = false; - /* We've got a huge page */ + + /* + * If we're using doublemap, we need MINOR fault, otherwise we need + * MISSING fault (which is the default). + */ + if (migrate_hugetlb_doublemap()) { +#ifdef UFFD_FEATURE_MINOR_HUGETLBFS + have_hp = supported_features & UFFD_FEATURE_MINOR_HUGETLBFS; +#endif + } else { #ifdef UFFD_FEATURE_MISSING_HUGETLBFS - have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS; + have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS; #endif + } + if (!have_hp) { - error_report("Userfault on this host does not support huge pages"); + error_report("Userfault on this host does not support huge pages " + "with %s fault traps", migrate_hugetlb_doublemap() ? + "MINOR" : "MISSING"); return false; } } @@ -669,22 +682,43 @@ static int ram_block_enable_notify(RAMBlock *rb, void *opaque) { MigrationIncomingState *mis = opaque; struct uffdio_register reg_struct; + bool minor_fault = postcopy_use_minor_fault(rb); reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb); reg_struct.range.len = rb->postcopy_length; + + /* + * For hugetlbfs with double-map enabled, we trap pages using minor + * mode, otherwise we use missing mode. Note: we also register missing + * mode for doublemap, but we should never hit it. + */ reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; + if (minor_fault) { + reg_struct.mode |= UFFDIO_REGISTER_MODE_MINOR; + } /* Now tell our userfault_fd that it's responsible for this area */ if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) { error_report("%s userfault register: %s", __func__, strerror(errno)); return -1; } - if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { - error_report("%s userfault: Region doesn't support COPY", __func__); - return -1; - } - if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) { - qemu_ram_set_uf_zeroable(rb); + + if (minor_fault) { + /* Using minor faults for this ramblock */ + if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_CONTINUE))) { + error_report("%s userfault: Region doesn't support CONTINUE", + __func__); + return -1; + } + } else { + /* Using missing faults for this ramblock */ + if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { + error_report("%s userfault: Region doesn't support COPY", __func__); + return -1; + } + if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) { + qemu_ram_set_uf_zeroable(rb); + } } return 0; @@ -916,6 +950,7 @@ static void *postcopy_ram_fault_thread(void *opaque) { MigrationIncomingState *mis = opaque; struct uffd_msg msg; + uint64_t address; int ret; size_t index; RAMBlock *rb = NULL; @@ -945,6 +980,7 @@ static void *postcopy_ram_fault_thread(void *opaque) } while (true) { + bool use_minor_fault, minor_flag; ram_addr_t rb_offset; int poll_result; @@ -1022,22 +1058,37 @@ static void *postcopy_ram_fault_thread(void *opaque) break; } - rb_offset = ROUND_DOWN(rb_offset, migration_ram_pagesize(rb)); - trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address, - qemu_ram_get_idstr(rb), - rb_offset, - msg.arg.pagefault.feat.ptid); - mark_postcopy_blocktime_begin( - (uintptr_t)(msg.arg.pagefault.address), - msg.arg.pagefault.feat.ptid, rb); + address = ROUND_DOWN(msg.arg.pagefault.address, + migration_ram_pagesize(rb)); + use_minor_fault = postcopy_use_minor_fault(rb); + minor_flag = !!(msg.arg.pagefault.flags & + UFFD_PAGEFAULT_FLAG_MINOR); + /* + * Do sanity check on the message flags to make sure this is + * the one we expect to receive. When using minor fault on + * this ramblock, it should _always_ be set; when not using + * minor fault, it should _never_ be set. + */ + if (use_minor_fault ^ minor_flag) { + error_report("%s: Unexpected page fault flags (0x%"PRIx64") " + "for address 0x%"PRIx64" (mode=%s)", __func__, + (uint64_t)msg.arg.pagefault.flags, + (uint64_t)msg.arg.pagefault.address, + use_minor_fault ? "MINOR" : "MISSING"); + } + + trace_postcopy_ram_fault_thread_request( + address, qemu_ram_get_idstr(rb), rb_offset, + msg.arg.pagefault.feat.ptid); + mark_postcopy_blocktime_begin( + (uintptr_t)(address), msg.arg.pagefault.feat.ptid, rb); retry: /* * Send the request to the source - we want to request one * of our host page sizes (which is >= TPS) */ - ret = postcopy_request_page(mis, rb, rb_offset, - msg.arg.pagefault.address); + ret = postcopy_request_page(mis, rb, rb_offset, address); if (ret) { /* May be network failure, try to wait for recovery */ postcopy_pause_fault_thread(mis); @@ -1694,3 +1745,13 @@ void *postcopy_preempt_thread(void *opaque) return NULL; } + +/* + * Whether we should use MINOR fault to trap page faults? It will be used + * when doublemap is enabled on hugetlbfs. The default value will be + * false, which means we'll keep using the legacy MISSING faults. + */ +bool postcopy_use_minor_fault(RAMBlock *rb) +{ + return migrate_hugetlb_doublemap() && qemu_ram_is_hugetlb(rb); +} diff --git a/migration/postcopy-ram.h b/migration/postcopy-ram.h index b4867a32d5..32734d2340 100644 --- a/migration/postcopy-ram.h +++ b/migration/postcopy-ram.h @@ -193,5 +193,6 @@ enum PostcopyChannels { void postcopy_preempt_new_channel(MigrationIncomingState *mis, QEMUFile *file); void postcopy_preempt_setup(MigrationState *s); int postcopy_preempt_establish_channel(MigrationState *s); +bool postcopy_use_minor_fault(RAMBlock *rb); #endif
When a ramblock is backed by hugetlbfs and the user specified using double-map feature, we trap the faults on these regions using minor mode. Teach QEMU about that. Add some sanity check on the fault flags when receiving a uffd message. For minor fault trapped ranges, we should always see the MINOR flag set, while when using generic missing faults we should never see it. Signed-off-by: Peter Xu <peterx@redhat.com> --- migration/postcopy-ram.c | 99 ++++++++++++++++++++++++++++++++-------- migration/postcopy-ram.h | 1 + 2 files changed, 81 insertions(+), 19 deletions(-)