@samitouri / QOSamiQemu / commits / f629431c09

accel/mshv: enable dirty page tracking

This change introduces the functions required to perform dirty page tracking to speed up migrations. We are using the sync, global_start, and global_stop hooks. The sync is implemented in batches. Before we can disable the dirty page tracking we have to set all dirty bits. Signed-off-by: Magnus Kulke <magnuskulke@linux.microsoft.com> Link: https://lore.kernel.org/r/20260417105618.3621-35-magnuskulke@linux.microsoft.com Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Magnus Kulke committed Apr 17, 2026 at 12:56 UTC f629431c09736cf1de900c0a1e08be157663f96d
3 files changed +218
accel/mshv/mem.c
+211
@@ -12,10 +12,13 @@
12
13 #include "qemu/osdep.h"
14 #include "qemu/error-report.h"
15 +#include "qapi/error.h"
16 #include "linux/mshv.h"
17 #include "system/address-spaces.h"
18 #include "system/mshv.h"
19 #include "system/mshv_int.h"
20 +#include "hw/hyperv/hvhdk_mini.h"
21 +#include "system/physmem.h"
22 #include "exec/memattrs.h"
23 #include <sys/ioctl.h>
24 #include "trace.h"
@@ -211,3 +214,211 @@ void mshv_set_phys_mem(MshvMemoryListener *mml, MemoryRegionSection *section,
214 abort();
215 }
216 }
217 +
218 +static int enable_dirty_page_tracking(int vm_fd)
219 +{
220 + int ret;
221 + struct hv_input_set_partition_property in = {0};
222 + struct mshv_root_hvcall args = {0};
223 +
224 + in.property_code = HV_PARTITION_PROPERTY_GPA_PAGE_ACCESS_TRACKING;
225 + in.property_value = 1;
226 +
227 + args.code = HVCALL_SET_PARTITION_PROPERTY;
228 + args.in_sz = sizeof(in);
229 + args.in_ptr = (uint64_t)&in;
230 +
231 + ret = mshv_hvcall(vm_fd, &args);
232 + if (ret < 0) {
233 + error_report("Failed to enable dirty page tracking: %s",
234 + strerror(errno));
235 + return -1;
236 + }
237 +
238 + return 0;
239 +}
240 +
241 +/*
242 + * Retrieve dirty page bitmap for a GPA range, clearing the dirty bits
243 + * atomically. Large ranges are handled in batches.
244 + */
245 +static int get_dirty_log(int vm_fd, uint64_t base_pfn, uint64_t page_count,
246 + unsigned long *bitmap, size_t bitmap_size)
247 +{
248 + uint64_t batch, bitmap_offset, completed = 0;
249 + struct mshv_gpap_access_bitmap args = {0};
250 + int ret;
251 +
252 + QEMU_BUILD_BUG_ON(MSHV_DIRTY_PAGES_BATCH_SIZE % BITS_PER_LONG != 0);
253 + assert(bitmap_size >= ROUND_UP(page_count, BITS_PER_LONG) / 8);
254 +
255 + while (completed < page_count) {
256 + batch = MIN(MSHV_DIRTY_PAGES_BATCH_SIZE, page_count - completed);
257 + bitmap_offset = completed / BITS_PER_LONG;
258 +
259 + args.access_type = MSHV_GPAP_ACCESS_TYPE_DIRTY;
260 + args.access_op = MSHV_GPAP_ACCESS_OP_CLEAR;
261 + args.page_count = batch;
262 + args.gpap_base = base_pfn + completed;
263 + args.bitmap_ptr = (uint64_t)(bitmap + bitmap_offset);
264 +
265 + ret = ioctl(vm_fd, MSHV_GET_GPAP_ACCESS_BITMAP, &args);
266 + if (ret < 0) {
267 + error_report("Failed to get dirty log (base_pfn=0x%" PRIx64
268 + " batch=%" PRIu64 "): %s",
269 + base_pfn + completed, batch, strerror(errno));
270 + return -1;
271 + }
272 + completed += batch;
273 + }
274 +
275 + return 0;
276 +}
277 +
278 +bool mshv_log_global_start(MemoryListener *listener, Error **errp)
279 +{
280 + int ret;
281 +
282 + ret = enable_dirty_page_tracking(mshv_state->vm);
283 + if (ret < 0) {
284 + error_setg_errno(errp, -ret, "Failed to enable dirty page tracking");
285 + return false;
286 + }
287 + return true;
288 +}
289 +
290 +static int disable_dirty_page_tracking(int vm_fd)
291 +{
292 + int ret;
293 + struct hv_input_set_partition_property in = {0};
294 + struct mshv_root_hvcall args = {0};
295 +
296 + in.property_code = HV_PARTITION_PROPERTY_GPA_PAGE_ACCESS_TRACKING;
297 + in.property_value = 0;
298 +
299 + args.code = HVCALL_SET_PARTITION_PROPERTY;
300 + args.in_sz = sizeof(in);
301 + args.in_ptr = (uint64_t)&in;
302 +
303 + ret = mshv_hvcall(vm_fd, &args);
304 + if (ret < 0) {
305 + error_report("Failed to disable dirty page tracking: %s",
306 + strerror(errno));
307 + return -1;
308 + }
309 +
310 + return 0;
311 +}
312 +
313 +static int set_dirty_pages(int vm_fd, uint64_t base_pfn, uint64_t page_count)
314 +{
315 + uint64_t batch, completed = 0;
316 + unsigned long bitmap[MSHV_DIRTY_PAGES_BATCH_SIZE / BITS_PER_LONG];
317 + struct mshv_gpap_access_bitmap args = {0};
318 + int ret;
319 +
320 + while (completed < page_count) {
321 + batch = MIN(MSHV_DIRTY_PAGES_BATCH_SIZE, page_count - completed);
322 +
323 + args.access_type = MSHV_GPAP_ACCESS_TYPE_DIRTY;
324 + args.access_op = MSHV_GPAP_ACCESS_OP_SET;
325 + args.page_count = batch;
326 + args.gpap_base = base_pfn + completed;
327 + args.bitmap_ptr = (uint64_t)bitmap;
328 +
329 + ret = ioctl(vm_fd, MSHV_GET_GPAP_ACCESS_BITMAP, &args);
330 + if (ret < 0) {
331 + error_report("Failed to set dirty pages (base_pfn=0x%" PRIx64
332 + " batch=%" PRIu64 "): %s",
333 + base_pfn + completed, batch, strerror(errno));
334 + return -1;
335 + }
336 + completed += batch;
337 + }
338 +
339 + return 0;
340 +}
341 +
342 +static bool set_dirty_bits_cb(Int128 start, Int128 len, const MemoryRegion *mr,
343 + hwaddr offset_in_region, void *opaque)
344 +{
345 + int ret, *errp = opaque;
346 + hwaddr gpa, size;
347 + uint64_t page_count, base_pfn;
348 +
349 + gpa = int128_get64(start);
350 + size = int128_get64(len);
351 + page_count = size >> MSHV_PAGE_SHIFT;
352 + base_pfn = gpa >> MSHV_PAGE_SHIFT;
353 +
354 + if (!mr->ram || mr->readonly) {
355 + return false;
356 + }
357 +
358 + if (page_count == 0) {
359 + return false;
360 + }
361 +
362 + ret = set_dirty_pages(mshv_state->vm, base_pfn, page_count);
363 +
364 + /* true aborts the iteration, which is what we want if there's an error */
365 + if (ret < 0) {
366 + *errp = ret;
367 + return true;
368 + }
369 +
370 + return false;
371 +}
372 +
373 +void mshv_log_global_stop(MemoryListener *listener)
374 +{
375 + int err = 0;
376 + /* MSHV requires all dirty bits to be set before disabling tracking. */
377 + FlatView *fv = address_space_to_flatview(&address_space_memory);
378 + flatview_for_each_range(fv, set_dirty_bits_cb, &err);
379 +
380 + if (err < 0) {
381 + error_report("Failed to set dirty bits before disabling tracking");
382 + }
383 +
384 + disable_dirty_page_tracking(mshv_state->vm);
385 +}
386 +
387 +void mshv_log_sync(MemoryListener *listener, MemoryRegionSection *section)
388 +{
389 + hwaddr size, start_addr, mr_offset;
390 + uint64_t page_count, base_pfn;
391 + size_t bitmap_size;
392 + unsigned long *bitmap;
393 + ram_addr_t ram_addr;
394 + int ret;
395 + MemoryRegion *mr = section->mr;
396 +
397 + if (!memory_region_is_ram(mr) || memory_region_is_rom(mr)) {
398 + return;
399 + }
400 +
401 + size = align_section(section, &start_addr);
402 + if (!size) {
403 + return;
404 + }
405 +
406 + page_count = size >> MSHV_PAGE_SHIFT;
407 + base_pfn = start_addr >> MSHV_PAGE_SHIFT;
408 + bitmap_size = ROUND_UP(page_count, BITS_PER_LONG) / 8;
409 + bitmap = g_malloc0(bitmap_size);
410 +
411 + ret = get_dirty_log(mshv_state->vm, base_pfn, page_count, bitmap,
412 + bitmap_size);
413 + if (ret < 0) {
414 + g_free(bitmap);
415 + return;
416 + }
417 +
418 + mr_offset = section->offset_within_region + start_addr -
419 + section->offset_within_address_space;
420 + ram_addr = memory_region_get_ram_addr(mr) + mr_offset;
421 +
422 + physical_memory_set_dirty_lebitmap(bitmap, ram_addr, page_count);
423 + g_free(bitmap);
424 +}
accel/mshv/mshv-all.c
+3
@@ -477,6 +477,9 @@ static MemoryListener mshv_memory_listener = {
477 .region_del = mem_region_del,
478 .eventfd_add = mem_ioeventfd_add,
479 .eventfd_del = mem_ioeventfd_del,
480 + .log_sync = mshv_log_sync,
481 + .log_global_start = mshv_log_global_start,
482 + .log_global_stop = mshv_log_global_stop,
483 };
484
485 static MemoryListener mshv_io_listener = {
include/system/mshv_int.h
+4
@@ -17,6 +17,7 @@
17 #include "hw/hyperv/hvhdk.h"
18
19 #define MSHV_MSR_ENTRIES_COUNT 64
20 +#define MSHV_DIRTY_PAGES_BATCH_SIZE 0x10000
21
22 typedef struct hyperv_message hv_message;
23
@@ -109,6 +110,9 @@ int mshv_guest_mem_write(uint64_t gpa, const uint8_t *data, uintptr_t size,
110 bool is_secure_mode);
111 void mshv_set_phys_mem(MshvMemoryListener *mml, MemoryRegionSection *section,
112 bool add);
113 +void mshv_log_sync(MemoryListener *listener, MemoryRegionSection *section);
114 +bool mshv_log_global_start(MemoryListener *listener, Error **errp);
115 +void mshv_log_global_stop(MemoryListener *listener);
116
117 /* msr */
118 typedef struct MshvMsrEntry {