Drivers: hv: balloon: Implement hot-add functionality
authorK. Y. Srinivasan <kys@microsoft.com>
Fri, 15 Mar 2013 19:25:43 +0000 (12:25 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 15 Mar 2013 19:11:51 +0000 (12:11 -0700)
Implement the memory hot-add functionality. With this, Linux guests can fully
participate in the Dynamic Memory protocol implemented in the Windows hosts.

In this version of the patch, based Olaf Herring's feedback, I have gotten
rid of the module level dependency on MEMORY_HOTPLUG. Instead the code within
the driver that depends on MEMORY_HOTPLUG has the appropriate compilation
switches. This would allow this driver to support pure ballooning in cases
where the kernel does not support memory hotplug.

Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/hv/hv_balloon.c

index 4743db9e5f349d0b7356706f0152391696393cb6..2cf7d4e964bd8a3afbde9eb6d01d0a3d6ba7b47e 100644 (file)
@@ -412,6 +412,27 @@ struct dm_info_msg {
  * End protocol definitions.
  */
 
+/*
+ * State to manage hot adding memory into the guest.
+ * The range start_pfn : end_pfn specifies the range
+ * that the host has asked us to hot add. The range
+ * start_pfn : ha_end_pfn specifies the range that we have
+ * currently hot added. We hot add in multiples of 128M
+ * chunks; it is possible that we may not be able to bring
+ * online all the pages in the region. The range
+ * covered_start_pfn : covered_end_pfn defines the pages that can
+ * be brough online.
+ */
+
+struct hv_hotadd_state {
+       struct list_head list;
+       unsigned long start_pfn;
+       unsigned long covered_start_pfn;
+       unsigned long covered_end_pfn;
+       unsigned long ha_end_pfn;
+       unsigned long end_pfn;
+};
+
 struct balloon_state {
        __u32 num_pages;
        struct work_struct wrk;
@@ -419,16 +440,17 @@ struct balloon_state {
 
 struct hot_add_wrk {
        union dm_mem_page_range ha_page_range;
+       union dm_mem_page_range ha_region_range;
        struct work_struct wrk;
 };
 
-static bool hot_add;
+static bool hot_add = true;
 static bool do_hot_add;
 /*
  * Delay reporting memory pressure by
  * the specified number of seconds.
  */
-static uint pressure_report_delay = 30;
+static uint pressure_report_delay = 45;
 
 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
@@ -456,6 +478,7 @@ enum hv_dm_state {
 static __u8 recv_buffer[PAGE_SIZE];
 static __u8 *send_buffer;
 #define PAGES_IN_2M    512
+#define HA_CHUNK (32 * 1024)
 
 struct hv_dynmem_device {
        struct hv_device *dev;
@@ -478,6 +501,17 @@ struct hv_dynmem_device {
         */
        struct hot_add_wrk ha_wrk;
 
+       /*
+        * This state tracks if the host has specified a hot-add
+        * region.
+        */
+       bool host_specified_ha_region;
+
+       /*
+        * State to synchronize hot-add.
+        */
+       struct completion  ol_waitevent;
+       bool ha_waiting;
        /*
         * This thread handles hot-add
         * requests from the host as well as notifying
@@ -486,6 +520,11 @@ struct hv_dynmem_device {
         */
        struct task_struct *thread;
 
+       /*
+        * A list of hot-add regions.
+        */
+       struct list_head ha_region_list;
+
        /*
         * We start with the highest version we can support
         * and downgrade based on the host; we save here the
@@ -496,35 +535,329 @@ struct hv_dynmem_device {
 
 static struct hv_dynmem_device dm_device;
 
-static void hot_add_req(struct work_struct *dummy)
+#ifdef CONFIG_MEMORY_HOTPLUG
+
+void hv_bring_pgs_online(unsigned long start_pfn, unsigned long size)
 {
+       int i;
 
-       struct dm_hot_add_response resp;
+       for (i = 0; i < size; i++) {
+               struct page *pg;
+               pg = pfn_to_page(start_pfn + i);
+               __online_page_set_limits(pg);
+               __online_page_increment_counters(pg);
+               __online_page_free(pg);
+       }
+}
+
+static void hv_mem_hot_add(unsigned long start, unsigned long size,
+                               unsigned long pfn_count,
+                               struct hv_hotadd_state *has)
+{
+       int ret = 0;
+       int i, nid, t;
+       unsigned long start_pfn;
+       unsigned long processed_pfn;
+       unsigned long total_pfn = pfn_count;
+
+       for (i = 0; i < (size/HA_CHUNK); i++) {
+               start_pfn = start + (i * HA_CHUNK);
+               has->ha_end_pfn +=  HA_CHUNK;
+
+               if (total_pfn > HA_CHUNK) {
+                       processed_pfn = HA_CHUNK;
+                       total_pfn -= HA_CHUNK;
+               } else {
+                       processed_pfn = total_pfn;
+                       total_pfn = 0;
+               }
+
+               has->covered_end_pfn +=  processed_pfn;
 
-       if (do_hot_add) {
+               init_completion(&dm_device.ol_waitevent);
+               dm_device.ha_waiting = true;
 
-               pr_info("Memory hot add not supported\n");
+               nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
+               ret = add_memory(nid, PFN_PHYS((start_pfn)),
+                               (HA_CHUNK << PAGE_SHIFT));
+
+               if (ret) {
+                       pr_info("hot_add memory failed error is %d\n", ret);
+                       has->ha_end_pfn -= HA_CHUNK;
+                       has->covered_end_pfn -=  processed_pfn;
+                       break;
+               }
 
                /*
-                * Currently we do not support hot add.
-                * Just fail the request.
+                * Wait for the memory block to be onlined.
                 */
+               t = wait_for_completion_timeout(&dm_device.ol_waitevent, 5*HZ);
+               if (t == 0) {
+                       pr_info("hot_add memory timedout\n");
+                       has->ha_end_pfn -= HA_CHUNK;
+                       has->covered_end_pfn -=  processed_pfn;
+                       break;
+               }
+
        }
 
+       return;
+}
+
+static void hv_online_page(struct page *pg)
+{
+       struct list_head *cur;
+       struct hv_hotadd_state *has;
+       unsigned long cur_start_pgp;
+       unsigned long cur_end_pgp;
+
+       if (dm_device.ha_waiting) {
+               dm_device.ha_waiting = false;
+               complete(&dm_device.ol_waitevent);
+       }
+
+       list_for_each(cur, &dm_device.ha_region_list) {
+               has = list_entry(cur, struct hv_hotadd_state, list);
+               cur_start_pgp = (unsigned long)
+                               pfn_to_page(has->covered_start_pfn);
+               cur_end_pgp = (unsigned long)pfn_to_page(has->covered_end_pfn);
+
+               if (((unsigned long)pg >= cur_start_pgp) &&
+                       ((unsigned long)pg < cur_end_pgp)) {
+                       /*
+                        * This frame is currently backed; online the
+                        * page.
+                        */
+                       __online_page_set_limits(pg);
+                       __online_page_increment_counters(pg);
+                       __online_page_free(pg);
+                       has->covered_start_pfn++;
+               }
+       }
+}
+
+static bool pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
+{
+       struct list_head *cur;
+       struct hv_hotadd_state *has;
+       unsigned long residual, new_inc;
+
+       if (list_empty(&dm_device.ha_region_list))
+               return false;
+
+       list_for_each(cur, &dm_device.ha_region_list) {
+               has = list_entry(cur, struct hv_hotadd_state, list);
+
+               /*
+                * If the pfn range we are dealing with is not in the current
+                * "hot add block", move on.
+                */
+               if ((start_pfn >= has->end_pfn))
+                       continue;
+               /*
+                * If the current hot add-request extends beyond
+                * our current limit; extend it.
+                */
+               if ((start_pfn + pfn_cnt) > has->end_pfn) {
+                       residual = (start_pfn + pfn_cnt - has->end_pfn);
+                       /*
+                        * Extend the region by multiples of HA_CHUNK.
+                        */
+                       new_inc = (residual / HA_CHUNK) * HA_CHUNK;
+                       if (residual % HA_CHUNK)
+                               new_inc += HA_CHUNK;
+
+                       has->end_pfn += new_inc;
+               }
+
+               /*
+                * If the current start pfn is not where the covered_end
+                * is, update it.
+                */
+
+               if (has->covered_end_pfn != start_pfn) {
+                       has->covered_end_pfn = start_pfn;
+                       has->covered_start_pfn = start_pfn;
+               }
+               return true;
+
+       }
+
+       return false;
+}
+
+static unsigned long handle_pg_range(unsigned long pg_start,
+                                       unsigned long pg_count)
+{
+       unsigned long start_pfn = pg_start;
+       unsigned long pfn_cnt = pg_count;
+       unsigned long size;
+       struct list_head *cur;
+       struct hv_hotadd_state *has;
+       unsigned long pgs_ol = 0;
+       unsigned long old_covered_state;
+
+       if (list_empty(&dm_device.ha_region_list))
+               return 0;
+
+       list_for_each(cur, &dm_device.ha_region_list) {
+               has = list_entry(cur, struct hv_hotadd_state, list);
+
+               /*
+                * If the pfn range we are dealing with is not in the current
+                * "hot add block", move on.
+                */
+               if ((start_pfn >= has->end_pfn))
+                       continue;
+
+               old_covered_state = has->covered_end_pfn;
+
+               if (start_pfn < has->ha_end_pfn) {
+                       /*
+                        * This is the case where we are backing pages
+                        * in an already hot added region. Bring
+                        * these pages online first.
+                        */
+                       pgs_ol = has->ha_end_pfn - start_pfn;
+                       if (pgs_ol > pfn_cnt)
+                               pgs_ol = pfn_cnt;
+                       hv_bring_pgs_online(start_pfn, pgs_ol);
+                       has->covered_end_pfn +=  pgs_ol;
+                       has->covered_start_pfn +=  pgs_ol;
+                       pfn_cnt -= pgs_ol;
+               }
+
+               if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
+                       /*
+                        * We have some residual hot add range
+                        * that needs to be hot added; hot add
+                        * it now. Hot add a multiple of
+                        * of HA_CHUNK that fully covers the pages
+                        * we have.
+                        */
+                       size = (has->end_pfn - has->ha_end_pfn);
+                       if (pfn_cnt <= size) {
+                               size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
+                               if (pfn_cnt % HA_CHUNK)
+                                       size += HA_CHUNK;
+                       } else {
+                               pfn_cnt = size;
+                       }
+                       hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
+               }
+               /*
+                * If we managed to online any pages that were given to us,
+                * we declare success.
+                */
+               return has->covered_end_pfn - old_covered_state;
+
+       }
+
+       return 0;
+}
+
+static unsigned long process_hot_add(unsigned long pg_start,
+                                       unsigned long pfn_cnt,
+                                       unsigned long rg_start,
+                                       unsigned long rg_size)
+{
+       struct hv_hotadd_state *ha_region = NULL;
+
+       if (pfn_cnt == 0)
+               return 0;
+
+       if (!dm_device.host_specified_ha_region)
+               if (pfn_covered(pg_start, pfn_cnt))
+                       goto do_pg_range;
+
+       /*
+        * If the host has specified a hot-add range; deal with it first.
+        */
+
+       if ((rg_size != 0) && (!dm_device.host_specified_ha_region)) {
+               ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
+               if (!ha_region)
+                       return 0;
+
+               INIT_LIST_HEAD(&ha_region->list);
+
+               list_add_tail(&ha_region->list, &dm_device.ha_region_list);
+               ha_region->start_pfn = rg_start;
+               ha_region->ha_end_pfn = rg_start;
+               ha_region->covered_start_pfn = pg_start;
+               ha_region->covered_end_pfn = pg_start;
+               ha_region->end_pfn = rg_start + rg_size;
+       }
+
+do_pg_range:
+       /*
+        * Process the page range specified; bringing them
+        * online if possible.
+        */
+       return handle_pg_range(pg_start, pfn_cnt);
+}
+
+#endif
+
+static void hot_add_req(struct work_struct *dummy)
+{
+       struct dm_hot_add_response resp;
+#ifdef CONFIG_MEMORY_HOTPLUG
+       unsigned long pg_start, pfn_cnt;
+       unsigned long rg_start, rg_sz;
+#endif
+       struct hv_dynmem_device *dm = &dm_device;
+
        memset(&resp, 0, sizeof(struct dm_hot_add_response));
        resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
        resp.hdr.size = sizeof(struct dm_hot_add_response);
        resp.hdr.trans_id = atomic_inc_return(&trans_id);
 
-       resp.page_count = 0;
-       resp.result = 0;
+#ifdef CONFIG_MEMORY_HOTPLUG
+       pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
+       pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
 
-       dm_device.state = DM_INITIALIZED;
-       vmbus_sendpacket(dm_device.dev->channel, &resp,
+       rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
+       rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
+
+       if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
+               unsigned long region_size;
+               unsigned long region_start;
+
+               /*
+                * The host has not specified the hot-add region.
+                * Based on the hot-add page range being specified,
+                * compute a hot-add region that can cover the pages
+                * that need to be hot-added while ensuring the alignment
+                * and size requirements of Linux as it relates to hot-add.
+                */
+               region_start = pg_start;
+               region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
+               if (pfn_cnt % HA_CHUNK)
+                       region_size += HA_CHUNK;
+
+               region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
+
+               rg_start = region_start;
+               rg_sz = region_size;
+       }
+
+       resp.page_count = process_hot_add(pg_start, pfn_cnt,
+                                       rg_start, rg_sz);
+#endif
+       if (resp.page_count > 0)
+               resp.result = 1;
+       else
+               resp.result = 0;
+
+       if (!do_hot_add || (resp.page_count == 0))
+               pr_info("Memory hot add failed\n");
+
+       dm->state = DM_INITIALIZED;
+       vmbus_sendpacket(dm->dev->channel, &resp,
                        sizeof(struct dm_hot_add_response),
                        (unsigned long)NULL,
                        VM_PKT_DATA_INBAND, 0);
-
 }
 
 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
@@ -867,6 +1200,7 @@ static void balloon_onchannelcallback(void *context)
        struct dm_balloon *bal_msg;
        struct dm_hot_add *ha_msg;
        union dm_mem_page_range *ha_pg_range;
+       union dm_mem_page_range *ha_region;
 
        memset(recv_buffer, 0, sizeof(recv_buffer));
        vmbus_recvpacket(dev->channel, recv_buffer,
@@ -907,8 +1241,26 @@ static void balloon_onchannelcallback(void *context)
                                pr_warn("Currently hot-adding\n");
                        dm->state = DM_HOT_ADD;
                        ha_msg = (struct dm_hot_add *)recv_buffer;
-                       ha_pg_range = &ha_msg->range;
-                       dm_device.ha_wrk.ha_page_range = *ha_pg_range;
+                       if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
+                               /*
+                                * This is a normal hot-add request specifying
+                                * hot-add memory.
+                                */
+                               ha_pg_range = &ha_msg->range;
+                               dm->ha_wrk.ha_page_range = *ha_pg_range;
+                               dm->ha_wrk.ha_region_range.page_range = 0;
+                       } else {
+                               /*
+                                * Host is specifying that we first hot-add
+                                * a region and then partially populate this
+                                * region.
+                                */
+                               dm->host_specified_ha_region = true;
+                               ha_pg_range = &ha_msg->range;
+                               ha_region = &ha_pg_range[1];
+                               dm->ha_wrk.ha_page_range = *ha_pg_range;
+                               dm->ha_wrk.ha_region_range = *ha_region;
+                       }
                        schedule_work(&dm_device.ha_wrk.wrk);
                        break;
 
@@ -952,8 +1304,10 @@ static int balloon_probe(struct hv_device *dev,
        dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
        init_completion(&dm_device.host_event);
        init_completion(&dm_device.config_event);
+       INIT_LIST_HEAD(&dm_device.ha_region_list);
        INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
        INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
+       dm_device.host_specified_ha_region = false;
 
        dm_device.thread =
                 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
@@ -962,6 +1316,10 @@ static int balloon_probe(struct hv_device *dev,
                goto probe_error1;
        }
 
+#ifdef CONFIG_MEMORY_HOTPLUG
+       set_online_page_callback(&hv_online_page);
+#endif
+
        hv_set_drvdata(dev, &dm_device);
        /*
         * Initiate the hand shake with the host and negotiate
@@ -1006,12 +1364,6 @@ static int balloon_probe(struct hv_device *dev,
        cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
 
        cap_msg.caps.cap_bits.balloon = 1;
-       /*
-        * While we currently don't support hot-add,
-        * we still advertise this capability since the
-        * host requires that guests partcipating in the
-        * dynamic memory protocol support hot add.
-        */
        cap_msg.caps.cap_bits.hot_add = 1;
 
        /*
@@ -1049,6 +1401,9 @@ static int balloon_probe(struct hv_device *dev,
        return 0;
 
 probe_error2:
+#ifdef CONFIG_MEMORY_HOTPLUG
+       restore_online_page_callback(&hv_online_page);
+#endif
        kthread_stop(dm_device.thread);
 
 probe_error1:
@@ -1061,15 +1416,26 @@ probe_error0:
 static int balloon_remove(struct hv_device *dev)
 {
        struct hv_dynmem_device *dm = hv_get_drvdata(dev);
+       struct list_head *cur, *tmp;
+       struct hv_hotadd_state *has;
 
        if (dm->num_pages_ballooned != 0)
                pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
 
        cancel_work_sync(&dm->balloon_wrk.wrk);
        cancel_work_sync(&dm->ha_wrk.wrk);
+
        vmbus_close(dev->channel);
        kthread_stop(dm->thread);
        kfree(send_buffer);
+#ifdef CONFIG_MEMORY_HOTPLUG
+       restore_online_page_callback(&hv_online_page);
+#endif
+       list_for_each_safe(cur, tmp, &dm->ha_region_list) {
+               has = list_entry(cur, struct hv_hotadd_state, list);
+               list_del(&has->list);
+               kfree(has);
+       }
 
        return 0;
 }