| Message ID | 1492076108-117229-3-git-send-email-wei.w.wang@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Apr 13, 2017 at 05:35:05PM +0800, Wei Wang wrote: > Add a new feature, VIRTIO_BALLOON_F_BALLOON_CHUNKS, which enables Let's find a better name here. VIRTIO_BALLOON_F_PAGE_CHUNK > the transfer of the ballooned (i.e. inflated/deflated) pages in > chunks to the host. > > The implementation of the previous virtio-balloon is not very > efficient, because the ballooned pages are transferred to the > host one by one. Here is the breakdown of the time in percentage > spent on each step of the balloon inflating process (inflating > 7GB of an 8GB idle guest). > > 1) allocating pages (6.5%) > 2) sending PFNs to host (68.3%) > 3) address translation (6.1%) > 4) madvise (19%) > > It takes about 4126ms for the inflating process to complete. > The above profiling shows that the bottlenecks are stage 2) > and stage 4). > > This patch optimizes step 2) by transferring pages to the host in > chunks. A chunk consists of guest physically continuous pages, and > it is offered to the host via a base PFN (i.e. the start PFN of > those physically continuous pages) and the size (i.e. the total > number of the pages). A chunk is formated as below: formatted > -------------------------------------------------------- > | Base (52 bit) | Rsvd (12 bit) | > -------------------------------------------------------- > -------------------------------------------------------- > | Size (52 bit) | Rsvd (12 bit) | > -------------------------------------------------------- > > By doing so, step 4) can also be optimized by doing address > translation and madvise() in chunks rather than page by page. > > With this new feature, the above ballooning process takes ~590ms > resulting in an improvement of ~85%. > > TODO: optimize stage 1) by allocating/freeing a chunk of pages > instead of a single page each time. > > Signed-off-by: Wei Wang <wei.w.wang@intel.com> > Signed-off-by: Liang Li <liang.z.li@intel.com> > Suggested-by: Michael S. Tsirkin <mst@redhat.com> So we don't need the bitmap to talk to host, it is just a data structure we chose to maintain lists of pages, right? OK as far as it goes but you need much better isolation for it. Build a data structure with APIs such as _init, _cleanup, _add, _clear, _find_first, _find_next. Completely unrelated to pages, it just maintains bits. Then use it here. > --- > drivers/virtio/virtio_balloon.c | 384 +++++++++++++++++++++++++++++++++--- > include/uapi/linux/virtio_balloon.h | 13 ++ > 2 files changed, 374 insertions(+), 23 deletions(-) > > diff --git a/drivers/virtio/virtio_balloon.c b/drivers/virtio/virtio_balloon.c > index f59cb4f..5e2e7cc 100644 > --- a/drivers/virtio/virtio_balloon.c > +++ b/drivers/virtio/virtio_balloon.c > @@ -42,6 +42,10 @@ > #define OOM_VBALLOON_DEFAULT_PAGES 256 > #define VIRTBALLOON_OOM_NOTIFY_PRIORITY 80 > > +#define PAGE_BMAP_SIZE (8 * PAGE_SIZE) > +#define PFNS_PER_PAGE_BMAP (PAGE_BMAP_SIZE * BITS_PER_BYTE) > +#define PAGE_BMAP_COUNT_MAX 32 > + Please prefix with VIRTIO_BALLOON_ and add comments. > static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES; > module_param(oom_pages, int, S_IRUSR | S_IWUSR); > MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); > @@ -50,6 +54,10 @@ MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); > static struct vfsmount *balloon_mnt; > #endif > > +/* Types of pages to chunk */ > +#define PAGE_CHUNK_TYPE_BALLOON 0 > + Doesn't look like you are ever adding more types in this patchset. Pls keep code simple, generalize it later. > +#define MAX_PAGE_CHUNKS 4096 This is an order-4 allocation. I'd make it 4095 and then it's an order-3 one. > struct virtio_balloon { > struct virtio_device *vdev; > struct virtqueue *inflate_vq, *deflate_vq, *stats_vq; > @@ -78,6 +86,32 @@ struct virtio_balloon { > /* Synchronize access/update to this struct virtio_balloon elements */ > struct mutex balloon_lock; > > + /* > + * Buffer for PAGE_CHUNK_TYPE_BALLOON: > + * virtio_balloon_page_chunk_hdr + > + * virtio_balloon_page_chunk * MAX_PAGE_CHUNKS > + */ > + struct virtio_balloon_page_chunk_hdr *balloon_page_chunk_hdr; > + struct virtio_balloon_page_chunk *balloon_page_chunk; > + > + /* Bitmap used to record pages */ > + unsigned long *page_bmap[PAGE_BMAP_COUNT_MAX]; > + /* Number of the allocated page_bmap */ > + unsigned int page_bmaps; > + > + /* > + * The allocated page_bmap size may be smaller than the pfn range of > + * the ballooned pages. In this case, we need to use the page_bmap > + * multiple times to cover the entire pfn range. It's like using a > + * short ruler several times to finish measuring a long object. > + * The start location of the ruler in the next measurement is the end > + * location of the ruler in the previous measurement. > + * > + * pfn_max & pfn_min: forms the pfn range of the ballooned pages > + * pfn_start & pfn_stop: records the start and stop pfn in each cover cover? what does this mean? looks like you only use these to pass data to tell_host. so pass these as parameters and you won't need to keep them in this structure. And then you can move this comment to set_page_bmap where it belongs. > + */ > + unsigned long pfn_min, pfn_max, pfn_start, pfn_stop; > + > /* The array of pfns we tell the Host about. */ > unsigned int num_pfns; > __virtio32 pfns[VIRTIO_BALLOON_ARRAY_PFNS_MAX]; > @@ -110,20 +144,201 @@ static void balloon_ack(struct virtqueue *vq) > wake_up(&vb->acked); > } > > -static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq) > +static inline void init_page_bmap_range(struct virtio_balloon *vb) > +{ > + vb->pfn_min = ULONG_MAX; > + vb->pfn_max = 0; > +} > + > +static inline void update_page_bmap_range(struct virtio_balloon *vb, > + struct page *page) > +{ > + unsigned long balloon_pfn = page_to_balloon_pfn(page); > + > + vb->pfn_min = min(balloon_pfn, vb->pfn_min); > + vb->pfn_max = max(balloon_pfn, vb->pfn_max); > +} > + > +/* The page_bmap size is extended by adding more number of page_bmap */ did you mean Allocate more bitmaps to cover the given number of pfns and add them to page_bmap ? This isn't what this function does. It blindly assumes 1 bitmap is allocated and allocates more, up to PAGE_BMAP_COUNT_MAX. > +static void extend_page_bmap_size(struct virtio_balloon *vb, > + unsigned long pfns) > +{ > + int i, bmaps; > + unsigned long bmap_len; > + > + bmap_len = ALIGN(pfns, BITS_PER_LONG) / BITS_PER_BYTE; > + bmap_len = ALIGN(bmap_len, PAGE_BMAP_SIZE); Align? PAGE_BMAP_SIZE doesn't even have to be a power of 2 ... > + bmaps = min((int)(bmap_len / PAGE_BMAP_SIZE), > + PAGE_BMAP_COUNT_MAX); I got lost here. Please use things like ARRAY_SIZE instead of macros. > + > + for (i = 1; i < bmaps; i++) { > + vb->page_bmap[i] = kmalloc(PAGE_BMAP_SIZE, GFP_KERNEL); > + if (vb->page_bmap[i]) > + vb->page_bmaps++; > + else > + break; > + } > +} > + > +static void free_extended_page_bmap(struct virtio_balloon *vb) > +{ > + int i, bmaps = vb->page_bmaps; > + > + for (i = 1; i < bmaps; i++) { > + kfree(vb->page_bmap[i]); > + vb->page_bmap[i] = NULL; > + vb->page_bmaps--; > + } > +} > + What's the magic number 1 here? Maybe you want to document what is going on. Here's a guess: We keep a single bmap around at all times. If memory does not fit there, we allocate up to PAGE_BMAP_COUNT_MAX of chunks. > +static void free_page_bmap(struct virtio_balloon *vb) > +{ > + int i; > + > + for (i = 0; i < vb->page_bmaps; i++) > + kfree(vb->page_bmap[i]); > +} > + > +static void clear_page_bmap(struct virtio_balloon *vb) > +{ > + int i; > + > + for (i = 0; i < vb->page_bmaps; i++) > + memset(vb->page_bmap[i], 0, PAGE_BMAP_SIZE); > +} > + > +static void send_page_chunks(struct virtio_balloon *vb, struct virtqueue *vq, > + int type, bool busy_wait) busy_wait seems unused. pls drop. > { > struct scatterlist sg; > + struct virtio_balloon_page_chunk_hdr *hdr; > + void *buf; > unsigned int len; > > - sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); > + switch (type) { > + case PAGE_CHUNK_TYPE_BALLOON: > + hdr = vb->balloon_page_chunk_hdr; > + len = 0; > + break; > + default: > + dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", > + __func__, type); > + return; > + } > > - /* We should always be able to add one buffer to an empty queue. */ > - virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); > - virtqueue_kick(vq); > + buf = (void *)hdr - len; Moving back to before the header? How can this make sense? It works fine since len is 0, so just buf = hdr. > + len += sizeof(struct virtio_balloon_page_chunk_hdr); > + len += hdr->chunks * sizeof(struct virtio_balloon_page_chunk); > + sg_init_table(&sg, 1); > + sg_set_buf(&sg, buf, len); > + if (!virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL)) { > + virtqueue_kick(vq); > + if (busy_wait) > + while (!virtqueue_get_buf(vq, &len) && > + !virtqueue_is_broken(vq)) > + cpu_relax(); > + else > + wait_event(vb->acked, virtqueue_get_buf(vq, &len)); > + hdr->chunks = 0; Why zero it here after device used it? Better to zero before use. > + } > +} > + > +static void add_one_chunk(struct virtio_balloon *vb, struct virtqueue *vq, > + int type, u64 base, u64 size) what are the units here? Looks like it's in 4kbyte units? > +{ > + struct virtio_balloon_page_chunk_hdr *hdr; > + struct virtio_balloon_page_chunk *chunk; > + > + switch (type) { > + case PAGE_CHUNK_TYPE_BALLOON: > + hdr = vb->balloon_page_chunk_hdr; > + chunk = vb->balloon_page_chunk; > + break; > + default: > + dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", > + __func__, type); > + return; > + } > + chunk = chunk + hdr->chunks; > + chunk->base = cpu_to_le64(base << VIRTIO_BALLOON_CHUNK_BASE_SHIFT); > + chunk->size = cpu_to_le64(size << VIRTIO_BALLOON_CHUNK_SIZE_SHIFT); > + hdr->chunks++; Isn't this LE? You should keep it somewhere else. > + if (hdr->chunks == MAX_PAGE_CHUNKS) > + send_page_chunks(vb, vq, type, false); and zero chunks here? > +} > + > +static void chunking_pages_from_bmap(struct virtio_balloon *vb, Does this mean "convert_bmap_to_chunks"? > + struct virtqueue *vq, > + unsigned long pfn_start, > + unsigned long *bmap, > + unsigned long len) > +{ > + unsigned long pos = 0, end = len * BITS_PER_BYTE; > + > + while (pos < end) { > + unsigned long one = find_next_bit(bmap, end, pos); > + > + if (one < end) { > + unsigned long chunk_size, zero; > + > + zero = find_next_zero_bit(bmap, end, one + 1); zero and one are unhelpful names unless they equal 0 and 1. current/next? > + if (zero >= end) > + chunk_size = end - one; > + else > + chunk_size = zero - one; > + > + if (chunk_size) > + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, > + pfn_start + one, chunk_size); Still not so what does a bit refer to? page or 4kbytes? I think it should be a page. > + pos = one + chunk_size; > + } else > + break; > + } > +} > + > +static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq) > +{ > + if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS)) { > + int pfns, page_bmaps, i; > + unsigned long pfn_start, pfns_len; > + > + pfn_start = vb->pfn_start; > + pfns = vb->pfn_stop - pfn_start + 1; > + pfns = roundup(roundup(pfns, BITS_PER_LONG), > + PFNS_PER_PAGE_BMAP); > + page_bmaps = pfns / PFNS_PER_PAGE_BMAP; > + pfns_len = pfns / BITS_PER_BYTE; > + > + for (i = 0; i < page_bmaps; i++) { > + unsigned int bmap_len = PAGE_BMAP_SIZE; > + > + /* The last one takes the leftover only */ I don't understand what does this mean. > + if (i + 1 == page_bmaps) > + bmap_len = pfns_len - PAGE_BMAP_SIZE * i; > + > + chunking_pages_from_bmap(vb, vq, pfn_start + > + i * PFNS_PER_PAGE_BMAP, > + vb->page_bmap[i], bmap_len); > + } > + if (vb->balloon_page_chunk_hdr->chunks > 0) > + send_page_chunks(vb, vq, PAGE_CHUNK_TYPE_BALLOON, > + false); > + } else { > + struct scatterlist sg; > + unsigned int len; > > - /* When host has read buffer, this completes via balloon_ack */ > - wait_event(vb->acked, virtqueue_get_buf(vq, &len)); > + sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); > > + /* > + * We should always be able to add one buffer to an empty > + * queue. > + */ > + virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); > + virtqueue_kick(vq); > + > + /* When host has read buffer, this completes via balloon_ack */ > + wait_event(vb->acked, virtqueue_get_buf(vq, &len)); > + } > } > > static void set_page_pfns(struct virtio_balloon *vb, > @@ -131,20 +346,73 @@ static void set_page_pfns(struct virtio_balloon *vb, > { > unsigned int i; > > - /* Set balloon pfns pointing at this page. > - * Note that the first pfn points at start of the page. */ > + /* > + * Set balloon pfns pointing at this page. > + * Note that the first pfn points at start of the page. > + */ > for (i = 0; i < VIRTIO_BALLOON_PAGES_PER_PAGE; i++) > pfns[i] = cpu_to_virtio32(vb->vdev, > page_to_balloon_pfn(page) + i); > } > Nice cleanup but pls split this out. This patch is big enough as it is. > +static void set_page_bmap(struct virtio_balloon *vb, > + struct list_head *pages, struct virtqueue *vq) > +{ > + unsigned long pfn_start, pfn_stop; > + struct page *page; > + bool found; > + > + vb->pfn_min = rounddown(vb->pfn_min, BITS_PER_LONG); > + vb->pfn_max = roundup(vb->pfn_max, BITS_PER_LONG); > + > + extend_page_bmap_size(vb, vb->pfn_max - vb->pfn_min + 1); This might not do anything in particular might not cover the given pfn range. Do we care? Why not? > + pfn_start = vb->pfn_min; > + > + while (pfn_start < vb->pfn_max) { > + pfn_stop = pfn_start + PFNS_PER_PAGE_BMAP * vb->page_bmaps; > + pfn_stop = pfn_stop < vb->pfn_max ? pfn_stop : vb->pfn_max; > + > + vb->pfn_start = pfn_start; > + clear_page_bmap(vb); > + found = false; > + > + list_for_each_entry(page, pages, lru) { > + unsigned long bmap_idx, bmap_pos, balloon_pfn; > + > + balloon_pfn = page_to_balloon_pfn(page); > + if (balloon_pfn < pfn_start || balloon_pfn > pfn_stop) > + continue; > + bmap_idx = (balloon_pfn - pfn_start) / > + PFNS_PER_PAGE_BMAP; > + bmap_pos = (balloon_pfn - pfn_start) % > + PFNS_PER_PAGE_BMAP; > + set_bit(bmap_pos, vb->page_bmap[bmap_idx]); Looks like this will crash if bmap_idx is out of range or if page_bmap allocation failed. > + > + found = true; > + } > + if (found) { > + vb->pfn_stop = pfn_stop; > + tell_host(vb, vq); > + } > + pfn_start = pfn_stop; > + } > + free_extended_page_bmap(vb); > +} > + > static unsigned fill_balloon(struct virtio_balloon *vb, size_t num) > { > struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info; > unsigned num_allocated_pages; > + bool chunking = virtio_has_feature(vb->vdev, > + VIRTIO_BALLOON_F_BALLOON_CHUNKS); > > /* We can only do one array worth at a time. */ > - num = min(num, ARRAY_SIZE(vb->pfns)); > + if (chunking) { > + init_page_bmap_range(vb); > + } else { > + /* We can only do one array worth at a time. */ > + num = min(num, ARRAY_SIZE(vb->pfns)); > + } > > mutex_lock(&vb->balloon_lock); > for (vb->num_pfns = 0; vb->num_pfns < num; > @@ -159,7 +427,10 @@ static unsigned fill_balloon(struct virtio_balloon *vb, size_t num) > msleep(200); > break; > } > - set_page_pfns(vb, vb->pfns + vb->num_pfns, page); > + if (chunking) > + update_page_bmap_range(vb, page); > + else > + set_page_pfns(vb, vb->pfns + vb->num_pfns, page); > vb->num_pages += VIRTIO_BALLOON_PAGES_PER_PAGE; > if (!virtio_has_feature(vb->vdev, > VIRTIO_BALLOON_F_DEFLATE_ON_OOM)) > @@ -168,8 +439,13 @@ static unsigned fill_balloon(struct virtio_balloon *vb, size_t num) > > num_allocated_pages = vb->num_pfns; > /* Did we get any? */ > - if (vb->num_pfns != 0) > - tell_host(vb, vb->inflate_vq); > + if (vb->num_pfns != 0) { > + if (chunking) > + set_page_bmap(vb, &vb_dev_info->pages, > + vb->inflate_vq); > + else > + tell_host(vb, vb->inflate_vq); > + } > mutex_unlock(&vb->balloon_lock); > > return num_allocated_pages; > @@ -195,6 +471,13 @@ static unsigned leak_balloon(struct virtio_balloon *vb, size_t num) > struct page *page; > struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info; > LIST_HEAD(pages); > + bool chunking = virtio_has_feature(vb->vdev, > + VIRTIO_BALLOON_F_BALLOON_CHUNKS); > + if (chunking) > + init_page_bmap_range(vb); > + else > + /* We can only do one array worth at a time. */ > + num = min(num, ARRAY_SIZE(vb->pfns)); > > /* We can only do one array worth at a time. */ > num = min(num, ARRAY_SIZE(vb->pfns)); > @@ -208,6 +491,10 @@ static unsigned leak_balloon(struct virtio_balloon *vb, size_t num) > if (!page) > break; > set_page_pfns(vb, vb->pfns + vb->num_pfns, page); > + if (chunking) > + update_page_bmap_range(vb, page); > + else > + set_page_pfns(vb, vb->pfns + vb->num_pfns, page); > list_add(&page->lru, &pages); > vb->num_pages -= VIRTIO_BALLOON_PAGES_PER_PAGE; > } > @@ -218,8 +505,12 @@ static unsigned leak_balloon(struct virtio_balloon *vb, size_t num) > * virtio_has_feature(vdev, VIRTIO_BALLOON_F_MUST_TELL_HOST); > * is true, we *have* to do it in this order > */ > - if (vb->num_pfns != 0) > - tell_host(vb, vb->deflate_vq); > + if (vb->num_pfns != 0) { > + if (chunking) > + set_page_bmap(vb, &pages, vb->deflate_vq); > + else > + tell_host(vb, vb->deflate_vq); > + } > release_pages_balloon(vb, &pages); > mutex_unlock(&vb->balloon_lock); > return num_freed_pages; > @@ -431,6 +722,13 @@ static int init_vqs(struct virtio_balloon *vb) > } > > #ifdef CONFIG_BALLOON_COMPACTION > + > +static void tell_host_one_page(struct virtio_balloon *vb, > + struct virtqueue *vq, struct page *page) > +{ > + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, page_to_pfn(page), 1); This passes 4kbytes to host which seems wrong - I think you want a full page. > +} > + > /* > * virtballoon_migratepage - perform the balloon page migration on behalf of > * a compation thread. (called under page lock) > @@ -454,6 +752,8 @@ static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info, > { > struct virtio_balloon *vb = container_of(vb_dev_info, > struct virtio_balloon, vb_dev_info); > + bool chunking = virtio_has_feature(vb->vdev, > + VIRTIO_BALLOON_F_BALLOON_CHUNKS); > unsigned long flags; > > /* > @@ -475,16 +775,22 @@ static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info, > vb_dev_info->isolated_pages--; > __count_vm_event(BALLOON_MIGRATE); > spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags); > - vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; > - set_page_pfns(vb, vb->pfns, newpage); > - tell_host(vb, vb->inflate_vq); > - > + if (chunking) { > + tell_host_one_page(vb, vb->inflate_vq, newpage); > + } else { > + vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; > + set_page_pfns(vb, vb->pfns, newpage); > + tell_host(vb, vb->inflate_vq); > + } > /* balloon's page migration 2nd step -- deflate "page" */ > balloon_page_delete(page); > - vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; > - set_page_pfns(vb, vb->pfns, page); > - tell_host(vb, vb->deflate_vq); > - > + if (chunking) { > + tell_host_one_page(vb, vb->deflate_vq, page); > + } else { > + vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; > + set_page_pfns(vb, vb->pfns, page); > + tell_host(vb, vb->deflate_vq); > + } > mutex_unlock(&vb->balloon_lock); > > put_page(page); /* balloon reference */ > @@ -511,6 +817,32 @@ static struct file_system_type balloon_fs = { > > #endif /* CONFIG_BALLOON_COMPACTION */ > > +static void balloon_page_chunk_init(struct virtio_balloon *vb) > +{ > + void *buf; > + > + /* > + * By default, we allocate page_bmap[0] only. More page_bmap will be > + * allocated on demand. > + */ > + vb->page_bmap[0] = kmalloc(PAGE_BMAP_SIZE, GFP_KERNEL); > + buf = kmalloc(sizeof(struct virtio_balloon_page_chunk_hdr) + > + sizeof(struct virtio_balloon_page_chunk) * > + MAX_PAGE_CHUNKS, GFP_KERNEL); > + if (!vb->page_bmap[0] || !buf) { > + __virtio_clear_bit(vb->vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS); this doesn't work as expected as features has been OK'd by then. You want something like validate_features that I posted. See "virtio: allow drivers to validate features". > + kfree(vb->page_bmap[0]); Looks like this will double free. you want to zero them I think. > + kfree(vb->balloon_page_chunk_hdr); > + dev_warn(&vb->vdev->dev, "%s: failed\n", __func__); > + } else { > + vb->page_bmaps = 1; > + vb->balloon_page_chunk_hdr = buf; > + vb->balloon_page_chunk_hdr->chunks = 0; > + vb->balloon_page_chunk = buf + > + sizeof(struct virtio_balloon_page_chunk_hdr); > + } > +} > + > static int virtballoon_probe(struct virtio_device *vdev) > { > struct virtio_balloon *vb; > @@ -533,6 +865,10 @@ static int virtballoon_probe(struct virtio_device *vdev) > spin_lock_init(&vb->stop_update_lock); > vb->stop_update = false; > vb->num_pages = 0; > + > + if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS)) > + balloon_page_chunk_init(vb); > + > mutex_init(&vb->balloon_lock); > init_waitqueue_head(&vb->acked); > vb->vdev = vdev; > @@ -609,6 +945,7 @@ static void virtballoon_remove(struct virtio_device *vdev) > cancel_work_sync(&vb->update_balloon_stats_work); > > remove_common(vb); > + free_page_bmap(vb); > if (vb->vb_dev_info.inode) > iput(vb->vb_dev_info.inode); > kfree(vb); > @@ -649,6 +986,7 @@ static unsigned int features[] = { > VIRTIO_BALLOON_F_MUST_TELL_HOST, > VIRTIO_BALLOON_F_STATS_VQ, > VIRTIO_BALLOON_F_DEFLATE_ON_OOM, > + VIRTIO_BALLOON_F_BALLOON_CHUNKS, > }; > > static struct virtio_driver virtio_balloon_driver = { > diff --git a/include/uapi/linux/virtio_balloon.h b/include/uapi/linux/virtio_balloon.h > index 343d7dd..be317b7 100644 > --- a/include/uapi/linux/virtio_balloon.h > +++ b/include/uapi/linux/virtio_balloon.h > @@ -34,6 +34,7 @@ > #define VIRTIO_BALLOON_F_MUST_TELL_HOST 0 /* Tell before reclaiming pages */ > #define VIRTIO_BALLOON_F_STATS_VQ 1 /* Memory Stats virtqueue */ > #define VIRTIO_BALLOON_F_DEFLATE_ON_OOM 2 /* Deflate balloon on OOM */ > +#define VIRTIO_BALLOON_F_BALLOON_CHUNKS 3 /* Inflate/Deflate pages in chunks */ > > /* Size of a PFN in the balloon interface. */ > #define VIRTIO_BALLOON_PFN_SHIFT 12 > @@ -82,4 +83,16 @@ struct virtio_balloon_stat { > __virtio64 val; > } __attribute__((packed)); > > +struct virtio_balloon_page_chunk_hdr { > + /* Number of chunks in the payload */ > + __le32 chunks; You want to make this __le64 to align everything to 64 bit. > +}; > + > +#define VIRTIO_BALLOON_CHUNK_BASE_SHIFT 12 > +#define VIRTIO_BALLOON_CHUNK_SIZE_SHIFT 12 > +struct virtio_balloon_page_chunk { so rename this virtio_balloon_page_chunk_entry > + __le64 base; > + __le64 size; > +}; > + And then: struct virtio_balloon_page_chunk { struct virtio_balloon_page_chunk_hdr hdr; struct virtio_balloon_page_chunk_entry entries[]; }; > #endif /* _LINUX_VIRTIO_BALLOON_H */ > -- > 2.7.4
On Thu, Apr 13, 2017 at 07:34:19PM +0300, Michael S. Tsirkin wrote: > So we don't need the bitmap to talk to host, it is just > a data structure we chose to maintain lists of pages, right? > OK as far as it goes but you need much better isolation for it. > Build a data structure with APIs such as _init, _cleanup, _add, _clear, > _find_first, _find_next. > Completely unrelated to pages, it just maintains bits. > Then use it here. That sounds an awful lot like the xbitmap I wrote a few months ago ... http://git.infradead.org/users/willy/linux-dax.git/commit/727e401bee5ad7d37e0077291d90cc17475c6392
On 04/14/2017 12:34 AM, Michael S. Tsirkin wrote: > On Thu, Apr 13, 2017 at 05:35:05PM +0800, Wei Wang wrote: > > So we don't need the bitmap to talk to host, it is just > a data structure we chose to maintain lists of pages, right? Right. bitmap is the way to gather pages to chunk. It's only needed in the balloon page case. For the unused page case, we don't need it, since the free page blocks are already chunks. > OK as far as it goes but you need much better isolation for it. > Build a data structure with APIs such as _init, _cleanup, _add, _clear, > _find_first, _find_next. > Completely unrelated to pages, it just maintains bits. > Then use it here. > > >> static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES; >> module_param(oom_pages, int, S_IRUSR | S_IWUSR); >> MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); >> @@ -50,6 +54,10 @@ MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); >> static struct vfsmount *balloon_mnt; >> #endif >> >> +/* Types of pages to chunk */ >> +#define PAGE_CHUNK_TYPE_BALLOON 0 >> + > Doesn't look like you are ever adding more types in this > patchset. Pls keep code simple, generalize it later. > "#define PAGE_CHUNK_TYPE_UNUSED 1" is added in another patch. Types of page to chunk are treated differently. Different types of page chunks are sent to the host via different protocols. 1) PAGE_CHUNK_TYPE_BALLOON: Ballooned (i.e. inflated/deflated) pages to chunk. For the ballooned type, it uses the basic chunk msg format: virtio_balloon_page_chunk_hdr + virtio_balloon_page_chunk * MAX_PAGE_CHUNKS 2) PAGE_CHUNK_TYPE_UNUSED: unused pages to chunk. It uses this miscq msg format: miscq_hdr + virtio_balloon_page_chunk_hdr + virtio_balloon_page_chunk * MAX_PAGE_CHUNKS The chunk msg is actually the payload of the miscq msg. >> +#define MAX_PAGE_CHUNKS 4096 > This is an order-4 allocation. I'd make it 4095 and then it's > an order-3 one. Sounds good, thanks. I think it would be better to make it 4090. Leave some space for the hdr as well. > >> struct virtio_balloon { >> struct virtio_device *vdev; >> struct virtqueue *inflate_vq, *deflate_vq, *stats_vq; >> @@ -78,6 +86,32 @@ struct virtio_balloon { >> /* Synchronize access/update to this struct virtio_balloon elements */ >> struct mutex balloon_lock; >> >> + /* >> + * Buffer for PAGE_CHUNK_TYPE_BALLOON: >> + * virtio_balloon_page_chunk_hdr + >> + * virtio_balloon_page_chunk * MAX_PAGE_CHUNKS >> + */ >> + struct virtio_balloon_page_chunk_hdr *balloon_page_chunk_hdr; >> + struct virtio_balloon_page_chunk *balloon_page_chunk; >> + >> + /* Bitmap used to record pages */ >> + unsigned long *page_bmap[PAGE_BMAP_COUNT_MAX]; >> + /* Number of the allocated page_bmap */ >> + unsigned int page_bmaps; >> + >> + /* >> + * The allocated page_bmap size may be smaller than the pfn range of >> + * the ballooned pages. In this case, we need to use the page_bmap >> + * multiple times to cover the entire pfn range. It's like using a >> + * short ruler several times to finish measuring a long object. >> + * The start location of the ruler in the next measurement is the end >> + * location of the ruler in the previous measurement. >> + * >> + * pfn_max & pfn_min: forms the pfn range of the ballooned pages >> + * pfn_start & pfn_stop: records the start and stop pfn in each cover > cover? what does this mean? > > looks like you only use these to pass data to tell_host. > so pass these as parameters and you won't need to keep > them in this structure. > > And then you can move this comment to set_page_bmap where > it belongs. > >> + */ >> + unsigned long pfn_min, pfn_max, pfn_start, pfn_stop; >> + >> /* The array of pfns we tell the Host about. */ >> unsigned int num_pfns; >> __virtio32 pfns[VIRTIO_BALLOON_ARRAY_PFNS_MAX]; >> @@ -110,20 +144,201 @@ static void balloon_ack(struct virtqueue *vq) >> wake_up(&vb->acked); >> } >> >> -static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq) >> +static inline void init_page_bmap_range(struct virtio_balloon *vb) >> +{ >> + vb->pfn_min = ULONG_MAX; >> + vb->pfn_max = 0; >> +} >> + >> +static inline void update_page_bmap_range(struct virtio_balloon *vb, >> + struct page *page) >> +{ >> + unsigned long balloon_pfn = page_to_balloon_pfn(page); >> + >> + vb->pfn_min = min(balloon_pfn, vb->pfn_min); >> + vb->pfn_max = max(balloon_pfn, vb->pfn_max); >> +} >> + >> +/* The page_bmap size is extended by adding more number of page_bmap */ > did you mean > > Allocate more bitmaps to cover the given number of pfns > and add them to page_bmap > > ? > > This isn't what this function does. > It blindly assumes 1 bitmap is allocated > and allocates more, up to PAGE_BMAP_COUNT_MAX. > Please let me use a concrete analogy to explain this algorithm: We have a 2-meter long ruler (i.e. page_bmap[0]). Case 1: To measure a 1-meter long object (i.e. pfn_max=1, pfn_min=0), we can simply use the ruler once and get to know that the object is 1-meter long. Case 2: To measure a 11-meter long object (i.e. pfn_max=11, pfn_min=0). We will first see if we can extend the 2-meter long ruler, for example, to 12-meter by getting another five 2-meter rulers and combine them (i.e. extend_page_bmap_size() to allocate page_bmap[1], page_bmap[2]...page_bmap[5]). Case 2.1: If the length of the ruler is successfully extended to 12-meter, that is, we get a 12-meter long ruler, then we can simply use the ruler once and know the length of the object is 11-meter. Case 2.2: If the ruler failed to be extended. Then we need to use the 2-meter long ruler 6 times to measure the 11-meter long object: 1st time: pfn_start=0, pfn_stop=2; 2nd time: pfn_start=2, pfn_stop=4; .. 6th time: pfn_start=10, pfn_stop=11 Still, we covered the entire length of the long object with the short ruler that we have. But we used it 6 times (i.e. use page_bmap[0], 6 times). Based on the understanding of this analogy, I think the following questions would be easier to understand. >> +static void extend_page_bmap_size(struct virtio_balloon *vb, >> + unsigned long pfns) >> +{ >> + int i, bmaps; >> + unsigned long bmap_len; >> + >> + bmap_len = ALIGN(pfns, BITS_PER_LONG) / BITS_PER_BYTE; >> + bmap_len = ALIGN(bmap_len, PAGE_BMAP_SIZE); > Align? PAGE_BMAP_SIZE doesn't even have to be a power of 2 ... ThoughPAGE_BMAP_SIZE has been set to 32K in the implementation, would you prefer to use roundup() here? >> + bmaps = min((int)(bmap_len / PAGE_BMAP_SIZE), >> + PAGE_BMAP_COUNT_MAX); > I got lost here. > > Please use things like ARRAY_SIZE instead of macros. PAGE_BMAP_COUNT_MAX is the total amount of page_bmap[] that is allowed to be allocated on demand. It is 32 in the implementation. For example, if the calculation shows that it needs 100 page_bmap[], but we can only afford 32, so use 32 for bmaps, instead of 100. The the following implementation go through Case 2.2. >> + >> + for (i = 1; i < bmaps; i++) { >> + vb->page_bmap[i] = kmalloc(PAGE_BMAP_SIZE, GFP_KERNEL); >> + if (vb->page_bmap[i]) >> + vb->page_bmaps++; >> + else >> + break; >> + } >> +} >> + >> +static void free_extended_page_bmap(struct virtio_balloon *vb) >> +{ >> + int i, bmaps = vb->page_bmaps; >> + >> + for (i = 1; i < bmaps; i++) { >> + kfree(vb->page_bmap[i]); >> + vb->page_bmap[i] = NULL; >> + vb->page_bmaps--; >> + } >> +} >> + > What's the magic number 1 here? > Maybe you want to document what is going on. > Here's a guess: > > We keep a single bmap around at all times. > If memory does not fit there, we allocate up to > PAGE_BMAP_COUNT_MAX of chunks. > Right. By default, we have only 1 page_bmap[] allocated. >> +static void free_page_bmap(struct virtio_balloon *vb) >> +{ >> + int i; >> + >> + for (i = 0; i < vb->page_bmaps; i++) >> + kfree(vb->page_bmap[i]); >> +} >> + >> +static void clear_page_bmap(struct virtio_balloon *vb) >> +{ >> + int i; >> + >> + for (i = 0; i < vb->page_bmaps; i++) >> + memset(vb->page_bmap[i], 0, PAGE_BMAP_SIZE); >> +} >> + >> +static void send_page_chunks(struct virtio_balloon *vb, struct virtqueue *vq, >> + int type, bool busy_wait) > busy_wait seems unused. pls drop. It will be used in the other patch (the 5th patch) for sending unused pages. Probably I can add it from that patch. > >> { >> struct scatterlist sg; >> + struct virtio_balloon_page_chunk_hdr *hdr; >> + void *buf; >> unsigned int len; >> >> - sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); >> + switch (type) { >> + case PAGE_CHUNK_TYPE_BALLOON: >> + hdr = vb->balloon_page_chunk_hdr; >> + len = 0; >> + break; >> + default: >> + dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", >> + __func__, type); >> + return; >> + } >> >> - /* We should always be able to add one buffer to an empty queue. */ >> - virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); >> - virtqueue_kick(vq); >> + buf = (void *)hdr - len; > Moving back to before the header? How can this make sense? > It works fine since len is 0, so just buf = hdr. > For the unused page chunk case, it follows its own protocol: miscq_hdr + payload(chunk msg). "buf = (void *)hdr - len" moves the buf pointer to the miscq_hdr, to send the entire miscq msg. Please check the patch for implementing the unused page chunk, it will be clear. If necessary, I can put "buf = (void *)hdr - len" from that patch. >> + len += sizeof(struct virtio_balloon_page_chunk_hdr); >> + len += hdr->chunks * sizeof(struct virtio_balloon_page_chunk); >> + sg_init_table(&sg, 1); >> + sg_set_buf(&sg, buf, len); >> + if (!virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL)) { >> + virtqueue_kick(vq); >> + if (busy_wait) >> + while (!virtqueue_get_buf(vq, &len) && >> + !virtqueue_is_broken(vq)) >> + cpu_relax(); >> + else >> + wait_event(vb->acked, virtqueue_get_buf(vq, &len)); >> + hdr->chunks = 0; > Why zero it here after device used it? Better to zero before use. hdr->chunks tells the host how many chunks are there in the payload. After the device use it, it is ready to zero it. > >> + } >> +} >> + >> +static void add_one_chunk(struct virtio_balloon *vb, struct virtqueue *vq, >> + int type, u64 base, u64 size) > what are the units here? Looks like it's in 4kbyte units? what is the "unit" you referred to? This is the function to add one chunk, base pfn and size of the chunk are supplied to the function. > >> + if (hdr->chunks == MAX_PAGE_CHUNKS) >> + send_page_chunks(vb, vq, type, false); > and zero chunks here? >> +} >> + >> +static void chunking_pages_from_bmap(struct virtio_balloon *vb, > Does this mean "convert_bmap_to_chunks"? > Yes. >> + struct virtqueue *vq, >> + unsigned long pfn_start, >> + unsigned long *bmap, >> + unsigned long len) >> +{ >> + unsigned long pos = 0, end = len * BITS_PER_BYTE; >> + >> + while (pos < end) { >> + unsigned long one = find_next_bit(bmap, end, pos); >> + >> + if (one < end) { >> + unsigned long chunk_size, zero; >> + >> + zero = find_next_zero_bit(bmap, end, one + 1); > > zero and one are unhelpful names unless they equal 0 and 1. > current/next? > I think it is clear if we think about the bitmap, for example: 00001111000011110000 one = the position of the next "1" bit, zero= the position of the next "0" bit, starting from one. Then, it is clear, chunk_size= zero - one would it be better to use pos_0 and pos_1? >> + if (zero >= end) >> + chunk_size = end - one; >> + else >> + chunk_size = zero - one; >> + >> + if (chunk_size) >> + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, >> + pfn_start + one, chunk_size); > Still not so what does a bit refer to? page or 4kbytes? > I think it should be a page. A bit in the bitmap corresponds to a pfn of a balloon page(4KB). But I think it doesn't matter here, since it is pfn. Using the above example: 00001111000011110000 If the starting bit above corresponds to pfn-0x1000 (i.e. pfn_start) Then the chunk base = 0x1004 (one is the position of the "Set" bit, which is 4), so pfn_start +one=0x1004 >> +static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq) >> +{ >> + if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS)) { >> + int pfns, page_bmaps, i; >> + unsigned long pfn_start, pfns_len; >> + >> + pfn_start = vb->pfn_start; >> + pfns = vb->pfn_stop - pfn_start + 1; >> + pfns = roundup(roundup(pfns, BITS_PER_LONG), >> + PFNS_PER_PAGE_BMAP); >> + page_bmaps = pfns / PFNS_PER_PAGE_BMAP; >> + pfns_len = pfns / BITS_PER_BYTE; >> + >> + for (i = 0; i < page_bmaps; i++) { >> + unsigned int bmap_len = PAGE_BMAP_SIZE; >> + >> + /* The last one takes the leftover only */ > I don't understand what does this mean. Still use the ruler analogy here: the object is 11-meter long, and we have a 2-meter long ruler. The 5th time has covered 10 meters of the object, Then the last time, the leftover is 1 meter, which means we can use half of the ruler to cover the left 1 meter. Back to the implementation here, if there are only 10 pfns left in the last round, I think it's not necessary to search the entire page_bmap[] till the end. >> +static void set_page_bmap(struct virtio_balloon *vb, >> + struct list_head *pages, struct virtqueue *vq) >> +{ >> + unsigned long pfn_start, pfn_stop; >> + struct page *page; >> + bool found; >> + >> + vb->pfn_min = rounddown(vb->pfn_min, BITS_PER_LONG); >> + vb->pfn_max = roundup(vb->pfn_max, BITS_PER_LONG); >> + >> + extend_page_bmap_size(vb, vb->pfn_max - vb->pfn_min + 1); > This might not do anything in particular might not cover the > given pfn range. Do we care? Why not? We have allocated only 1 page_bmap[], which is able to cover 1GB memory. To inflate 2GB, it will try to extend by getting one more page_bmap, page_bmap[1]. >> + pfn_start = vb->pfn_min; >> + >> + while (pfn_start < vb->pfn_max) { >> + pfn_stop = pfn_start + PFNS_PER_PAGE_BMAP * vb->page_bmaps; >> + pfn_stop = pfn_stop < vb->pfn_max ? pfn_stop : vb->pfn_max; >> + >> + vb->pfn_start = pfn_start; >> + clear_page_bmap(vb); >> + found = false; >> + >> + list_for_each_entry(page, pages, lru) { >> + unsigned long bmap_idx, bmap_pos, balloon_pfn; >> + >> + balloon_pfn = page_to_balloon_pfn(page); >> + if (balloon_pfn < pfn_start || balloon_pfn > pfn_stop) >> + continue; >> + bmap_idx = (balloon_pfn - pfn_start) / >> + PFNS_PER_PAGE_BMAP; >> + bmap_pos = (balloon_pfn - pfn_start) % >> + PFNS_PER_PAGE_BMAP; >> + set_bit(bmap_pos, vb->page_bmap[bmap_idx]); > Looks like this will crash if bmap_idx is out of range or > if page_bmap allocation failed. No, it won't. Please think about the analogy Case 2.2: pfn_start is updated in each round. Like in the 2nd round, pfn_start is updated to 2, balloon_pfn will be a value between 2 and 4, so the result of "(balloon_pfn - pfn_start) / PFNS_PER_PAGE_BMAP" will always be 0 when we only have page_bmap[0]. > > #ifdef CONFIG_BALLOON_COMPACTION > + > +static void tell_host_one_page(struct virtio_balloon *vb, > + struct virtqueue *vq, struct page *page) > +{ > + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, page_to_pfn(page), 1); > This passes 4kbytes to host which seems wrong - I think you want a full page. OK. It should be add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, page_to_pfn(page), VIRTIO_BALLOON_PAGES_PER_PAGE) right? If Page=2*BalloonPage, it will pass 2*4K to the host. +static void balloon_page_chunk_init(struct virtio_balloon *vb) +{ + void *buf; + + /* + * By default, we allocate page_bmap[0] only. More page_bmap will be + * allocated on demand. + */ + vb->page_bmap[0] = kmalloc(PAGE_BMAP_SIZE, GFP_KERNEL); + buf = kmalloc(sizeof(struct virtio_balloon_page_chunk_hdr) + + sizeof(struct virtio_balloon_page_chunk) * + MAX_PAGE_CHUNKS, GFP_KERNEL); + if (!vb->page_bmap[0] || !buf) { + __virtio_clear_bit(vb->vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS); > this doesn't work as expected as features has been OK'd by then. > You want something like > validate_features that I posted. See > "virtio: allow drivers to validate features". OK. I will change it after that patch is merged. > >> + kfree(vb->page_bmap[0]); > Looks like this will double free. you want to zero them I think. > OK. I'll NULL the pointers after kfree(). Best, Wei
On Fri, Apr 14, 2017 at 04:37:52PM +0800, Wei Wang wrote: > On 04/14/2017 12:34 AM, Michael S. Tsirkin wrote: > > On Thu, Apr 13, 2017 at 05:35:05PM +0800, Wei Wang wrote: > > > > So we don't need the bitmap to talk to host, it is just > > a data structure we chose to maintain lists of pages, right? > Right. bitmap is the way to gather pages to chunk. > It's only needed in the balloon page case. > For the unused page case, we don't need it, since the free > page blocks are already chunks. > > > OK as far as it goes but you need much better isolation for it. > > Build a data structure with APIs such as _init, _cleanup, _add, _clear, > > _find_first, _find_next. > > Completely unrelated to pages, it just maintains bits. > > Then use it here. > > > > > > > static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES; > > > module_param(oom_pages, int, S_IRUSR | S_IWUSR); > > > MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); > > > @@ -50,6 +54,10 @@ MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); > > > static struct vfsmount *balloon_mnt; > > > #endif > > > +/* Types of pages to chunk */ > > > +#define PAGE_CHUNK_TYPE_BALLOON 0 > > > + > > Doesn't look like you are ever adding more types in this > > patchset. Pls keep code simple, generalize it later. > > > "#define PAGE_CHUNK_TYPE_UNUSED 1" is added in another patch. I would say add the extra code there too. Or maybe we can avoid adding it altogether. > Types of page to chunk are treated differently. Different types of page > chunks are sent to the host via different protocols. > > 1) PAGE_CHUNK_TYPE_BALLOON: Ballooned (i.e. inflated/deflated) pages > to chunk. For the ballooned type, it uses the basic chunk msg format: > > virtio_balloon_page_chunk_hdr + > virtio_balloon_page_chunk * MAX_PAGE_CHUNKS > > 2) PAGE_CHUNK_TYPE_UNUSED: unused pages to chunk. It uses this miscq msg > format: > miscq_hdr + > virtio_balloon_page_chunk_hdr + > virtio_balloon_page_chunk * MAX_PAGE_CHUNKS > > The chunk msg is actually the payload of the miscq msg. > > So just combine the two message formats and then it'll all be easier? > > > +#define MAX_PAGE_CHUNKS 4096 > > This is an order-4 allocation. I'd make it 4095 and then it's > > an order-3 one. > > Sounds good, thanks. > I think it would be better to make it 4090. Leave some space for the hdr > as well. And miscq hdr. In fact just let compiler do the math - something like: (8 * PAGE_SIZE - sizeof(hdr)) / sizeof(chunk) I skimmed explanation of algorithms below but please make sure code speaks for itself and add comments inline to document it. Whenever you answered me inline this is where you want to try to make code clearer and add comments. Also, pls find ways to abstract the data structure so we don't need to deal with its internals all over the code. .... > > > > > { > > > struct scatterlist sg; > > > + struct virtio_balloon_page_chunk_hdr *hdr; > > > + void *buf; > > > unsigned int len; > > > - sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); > > > + switch (type) { > > > + case PAGE_CHUNK_TYPE_BALLOON: > > > + hdr = vb->balloon_page_chunk_hdr; > > > + len = 0; > > > + break; > > > + default: > > > + dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", > > > + __func__, type); > > > + return; > > > + } > > > - /* We should always be able to add one buffer to an empty queue. */ > > > - virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); > > > - virtqueue_kick(vq); > > > + buf = (void *)hdr - len; > > Moving back to before the header? How can this make sense? > > It works fine since len is 0, so just buf = hdr. > > > For the unused page chunk case, it follows its own protocol: > miscq_hdr + payload(chunk msg). > "buf = (void *)hdr - len" moves the buf pointer to the miscq_hdr, to send > the entire miscq msg. Well just pass the correct pointer in. > Please check the patch for implementing the unused page chunk, > it will be clear. If necessary, I can put "buf = (void *)hdr - len" from > that patch. Exactly. And all this pointer math is very messy. Please look for ways to clean it. It's generally easy to fill structures: struct foo *foo = kmalloc(..., sizeof(*foo) + n * sizeof(foo->a[0])); for (i = 0; i < n; ++i) foo->a[i] = b; this is the kind of code that's easy to understand and it's obvious there are no overflows and no info leaks here. > > > > + len += sizeof(struct virtio_balloon_page_chunk_hdr); > > > + len += hdr->chunks * sizeof(struct virtio_balloon_page_chunk); > > > + sg_init_table(&sg, 1); > > > + sg_set_buf(&sg, buf, len); > > > + if (!virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL)) { > > > + virtqueue_kick(vq); > > > + if (busy_wait) > > > + while (!virtqueue_get_buf(vq, &len) && > > > + !virtqueue_is_broken(vq)) > > > + cpu_relax(); > > > + else > > > + wait_event(vb->acked, virtqueue_get_buf(vq, &len)); > > > + hdr->chunks = 0; > > Why zero it here after device used it? Better to zero before use. > > hdr->chunks tells the host how many chunks are there in the payload. > After the device use it, it is ready to zero it. It's rather confusing. Try to pass # of chunks around in some other way. > > > > > + } > > > +} > > > + > > > +static void add_one_chunk(struct virtio_balloon *vb, struct virtqueue *vq, > > > + int type, u64 base, u64 size) > > what are the units here? Looks like it's in 4kbyte units? > > what is the "unit" you referred to? > This is the function to add one chunk, base pfn and size of the chunk are > supplied to the function. > Are both size and base in bytes then? But you do not send them to host as is, you shift them for some reason before sending them to host. > > > > > > + if (hdr->chunks == MAX_PAGE_CHUNKS) > > > + send_page_chunks(vb, vq, type, false); > > and zero chunks here? > > > +} > > > + > > > +static void chunking_pages_from_bmap(struct virtio_balloon *vb, > > Does this mean "convert_bmap_to_chunks"? > > > > Yes. > Pls name it accordingly then. > > > + struct virtqueue *vq, > > > + unsigned long pfn_start, > > > + unsigned long *bmap, > > > + unsigned long len) > > > +{ > > > + unsigned long pos = 0, end = len * BITS_PER_BYTE; > > > + > > > + while (pos < end) { > > > + unsigned long one = find_next_bit(bmap, end, pos); > > > + > > > + if (one < end) { > > > + unsigned long chunk_size, zero; > > > + > > > + zero = find_next_zero_bit(bmap, end, one + 1); > > > > zero and one are unhelpful names unless they equal 0 and 1. > > current/next? > > > > I think it is clear if we think about the bitmap, for example: > 00001111000011110000 > one = the position of the next "1" bit, > zero= the position of the next "0" bit, starting from one. > > Then, it is clear, chunk_size= zero - one > > would it be better to use pos_0 and pos_1? Oh, so it's next_zero_bit and next_bit. > > > + if (zero >= end) > > > + chunk_size = end - one; > > > + else > > > + chunk_size = zero - one; > > > + > > > + if (chunk_size) > > > + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, > > > + pfn_start + one, chunk_size); > > Still not so what does a bit refer to? page or 4kbytes? > > I think it should be a page. > A bit in the bitmap corresponds to a pfn of a balloon page(4KB). That's a waste on systems with large page sizes, and it does not look like you handle that case correctly. > But I think it doesn't matter here, since it is pfn. > Using the above example: > 00001111000011110000 > > If the starting bit above corresponds to pfn-0x1000 (i.e. pfn_start) > Then the chunk base = 0x1004 > (one is the position of the "Set" bit, which is 4), so pfn_start +one=0x1004 > > > > +static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq) > > > +{ > > > + if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS)) { > > > + int pfns, page_bmaps, i; > > > + unsigned long pfn_start, pfns_len; > > > + > > > + pfn_start = vb->pfn_start; > > > + pfns = vb->pfn_stop - pfn_start + 1; > > > + pfns = roundup(roundup(pfns, BITS_PER_LONG), > > > + PFNS_PER_PAGE_BMAP); > > > + page_bmaps = pfns / PFNS_PER_PAGE_BMAP; > > > + pfns_len = pfns / BITS_PER_BYTE; > > > + > > > + for (i = 0; i < page_bmaps; i++) { > > > + unsigned int bmap_len = PAGE_BMAP_SIZE; > > > + > > > + /* The last one takes the leftover only */ > > I don't understand what does this mean. > Still use the ruler analogy here: the object is 11-meter long, and we have > a 2-meter long ruler. The 5th time has covered 10 meters of the object, Then > the last time, the leftover is 1 meter, which means we can use half of the > ruler > to cover the left 1 meter. > > Back to the implementation here, if there are only 10 pfns left in the last > round, > I think it's not necessary to search the entire page_bmap[] till the end. Pls reword the comment to make it a whole sentence. > > > +static void set_page_bmap(struct virtio_balloon *vb, > > > + struct list_head *pages, struct virtqueue *vq) > > > +{ > > > + unsigned long pfn_start, pfn_stop; > > > + struct page *page; > > > + bool found; > > > + > > > + vb->pfn_min = rounddown(vb->pfn_min, BITS_PER_LONG); > > > + vb->pfn_max = roundup(vb->pfn_max, BITS_PER_LONG); > > > + > > > + extend_page_bmap_size(vb, vb->pfn_max - vb->pfn_min + 1); > > This might not do anything in particular might not cover the > > given pfn range. Do we care? Why not? > > We have allocated only 1 page_bmap[], which is able to cover 1GB memory. > To inflate 2GB, it will try to extend by getting one more page_bmap, > page_bmap[1]. > > > > + pfn_start = vb->pfn_min; > > > + > > > + while (pfn_start < vb->pfn_max) { > > > + pfn_stop = pfn_start + PFNS_PER_PAGE_BMAP * vb->page_bmaps; > > > + pfn_stop = pfn_stop < vb->pfn_max ? pfn_stop : vb->pfn_max; > > > + > > > + vb->pfn_start = pfn_start; > > > + clear_page_bmap(vb); > > > + found = false; > > > + > > > + list_for_each_entry(page, pages, lru) { > > > + unsigned long bmap_idx, bmap_pos, balloon_pfn; > > > + > > > + balloon_pfn = page_to_balloon_pfn(page); > > > + if (balloon_pfn < pfn_start || balloon_pfn > pfn_stop) > > > + continue; > > > + bmap_idx = (balloon_pfn - pfn_start) / > > > + PFNS_PER_PAGE_BMAP; > > > + bmap_pos = (balloon_pfn - pfn_start) % > > > + PFNS_PER_PAGE_BMAP; > > > + set_bit(bmap_pos, vb->page_bmap[bmap_idx]); > > Looks like this will crash if bmap_idx is out of range or > > if page_bmap allocation failed. > > No, it won't. Please think about the analogy Case 2.2: pfn_start is updated > in each round. Like in the 2nd round, pfn_start is updated to 2, balloon_pfn > will be a value between 2 and 4, so the result of > "(balloon_pfn - pfn_start) / PFNS_PER_PAGE_BMAP" will always be 0 when > we only have page_bmap[0]. All these cases confuse too much. Pls abstract away the underlying data structure (or better find an appropriate existing one). Things should become clearer then. > > #ifdef CONFIG_BALLOON_COMPACTION > > + > > +static void tell_host_one_page(struct virtio_balloon *vb, > > + struct virtqueue *vq, struct page *page) > > +{ > > + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, page_to_pfn(page), 1); > > This passes 4kbytes to host which seems wrong - I think you want a full page. > > OK. It should be > add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, > page_to_pfn(page), VIRTIO_BALLOON_PAGES_PER_PAGE) > > right? > > If Page=2*BalloonPage, it will pass 2*4K to the host. I guess, or better use whole page units. > +static void balloon_page_chunk_init(struct virtio_balloon *vb) > +{ > + void *buf; > + > + /* > + * By default, we allocate page_bmap[0] only. More page_bmap will be > + * allocated on demand. > + */ > + vb->page_bmap[0] = kmalloc(PAGE_BMAP_SIZE, GFP_KERNEL); > + buf = kmalloc(sizeof(struct virtio_balloon_page_chunk_hdr) + > + sizeof(struct virtio_balloon_page_chunk) * > + MAX_PAGE_CHUNKS, GFP_KERNEL); > + if (!vb->page_bmap[0] || !buf) { > + __virtio_clear_bit(vb->vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS); > > > this doesn't work as expected as features has been OK'd by then. > > You want something like > > validate_features that I posted. See > > "virtio: allow drivers to validate features". > > OK. I will change it after that patch is merged. It's upstream now. > > > > > + kfree(vb->page_bmap[0]); > > Looks like this will double free. you want to zero them I think. > > > > OK. I'll NULL the pointers after kfree(). > > > > Best, > Wei > >
On 04/15/2017 05:38 AM, Michael S. Tsirkin wrote: > On Fri, Apr 14, 2017 at 04:37:52PM +0800, Wei Wang wrote: >> On 04/14/2017 12:34 AM, Michael S. Tsirkin wrote: >>> On Thu, Apr 13, 2017 at 05:35:05PM +0800, Wei Wang wrote: >>> >>> So we don't need the bitmap to talk to host, it is just >>> a data structure we chose to maintain lists of pages, right? >> Right. bitmap is the way to gather pages to chunk. >> It's only needed in the balloon page case. >> For the unused page case, we don't need it, since the free >> page blocks are already chunks. >> >>> OK as far as it goes but you need much better isolation for it. >>> Build a data structure with APIs such as _init, _cleanup, _add, _clear, >>> _find_first, _find_next. >>> Completely unrelated to pages, it just maintains bits. >>> Then use it here. >>> >>> >>>> static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES; >>>> module_param(oom_pages, int, S_IRUSR | S_IWUSR); >>>> MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); >>>> @@ -50,6 +54,10 @@ MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); >>>> static struct vfsmount *balloon_mnt; >>>> #endif >>>> +/* Types of pages to chunk */ >>>> +#define PAGE_CHUNK_TYPE_BALLOON 0 >>>> + >>> Doesn't look like you are ever adding more types in this >>> patchset. Pls keep code simple, generalize it later. >>> >> "#define PAGE_CHUNK_TYPE_UNUSED 1" is added in another patch. > I would say add the extra code there too. Or maybe we can avoid > adding it altogether. I'm trying to have the two features( i.e. "balloon pages" and "unused pages") decoupled while trying to use common functions to deal with the commonalities. That's the reason to define the above macro. Without the macro, we will need to have separate functions, for example, instead of one "add_one_chunk()", we need to have add_one_balloon_page_chunk() and add_one_unused_page_chunk(), and some of the implementations will be kind of duplicate in the two functions. Probably we can add it when the second feature comes to the code. > >> Types of page to chunk are treated differently. Different types of page >> chunks are sent to the host via different protocols. >> >> 1) PAGE_CHUNK_TYPE_BALLOON: Ballooned (i.e. inflated/deflated) pages >> to chunk. For the ballooned type, it uses the basic chunk msg format: >> >> virtio_balloon_page_chunk_hdr + >> virtio_balloon_page_chunk * MAX_PAGE_CHUNKS >> >> 2) PAGE_CHUNK_TYPE_UNUSED: unused pages to chunk. It uses this miscq msg >> format: >> miscq_hdr + >> virtio_balloon_page_chunk_hdr + >> virtio_balloon_page_chunk * MAX_PAGE_CHUNKS >> >> The chunk msg is actually the payload of the miscq msg. >> >> > So just combine the two message formats and then it'll all be easier? > Yes, it'll be simple with only one msg format. But the problem I see here is that miscq hdr is something necessary for the "unused page" usage, but not needed by the "balloon page" usage. To be more precise, struct virtio_balloon_miscq_hdr { __le16 cmd; __le16 flags; }; 'cmd' specifies the command from the miscq (I envision that miscq will be further used to handle other possible miscellaneous requests either from the host or to the host), so 'cmd' is necessary for the miscq. But the inflateq is exclusively used for inflating pages, so adding a command to it would be redundant and look a little bewildered there. 'flags': We currently use bit 0 of flags to indicate the completion ofa command, this is also useful in the "unused page" usage, and not needed by the "balloon page" usage. >>>> +#define MAX_PAGE_CHUNKS 4096 >>> This is an order-4 allocation. I'd make it 4095 and then it's >>> an order-3 one. >> Sounds good, thanks. >> I think it would be better to make it 4090. Leave some space for the hdr >> as well. > And miscq hdr. In fact just let compiler do the math - something like: > (8 * PAGE_SIZE - sizeof(hdr)) / sizeof(chunk) Agree, thanks. > > I skimmed explanation of algorithms below but please make sure > code speaks for itself and add comments inline to document it. > Whenever you answered me inline this is where you want to > try to make code clearer and add comments. > > Also, pls find ways to abstract the data structure so we don't > need to deal with its internals all over the code. > > > .... > >>>> { >>>> struct scatterlist sg; >>>> + struct virtio_balloon_page_chunk_hdr *hdr; >>>> + void *buf; >>>> unsigned int len; >>>> - sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); >>>> + switch (type) { >>>> + case PAGE_CHUNK_TYPE_BALLOON: >>>> + hdr = vb->balloon_page_chunk_hdr; >>>> + len = 0; >>>> + break; >>>> + default: >>>> + dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", >>>> + __func__, type); >>>> + return; >>>> + } >>>> - /* We should always be able to add one buffer to an empty queue. */ >>>> - virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); >>>> - virtqueue_kick(vq); >>>> + buf = (void *)hdr - len; >>> Moving back to before the header? How can this make sense? >>> It works fine since len is 0, so just buf = hdr. >>> >> For the unused page chunk case, it follows its own protocol: >> miscq_hdr + payload(chunk msg). >> "buf = (void *)hdr - len" moves the buf pointer to the miscq_hdr, to send >> the entire miscq msg. > Well just pass the correct pointer in. > OK. The miscq msg is { miscq_hdr; chunk_msg; } We can probably change the code like this: #define CHUNK_TO_MISCQ_MSG(chunk) (chunk - sizeof(struct virtio_balloon_miscq_hdr)) switch (type) { case PAGE_CHUNK_TYPE_BALLOON: msg_buf = vb->balloon_page_chunk_hdr; msg_len = sizeof(struct virtio_balloon_page_chunk_hdr) + nr_chunks * sizeof(struct virtio_balloon_page_chunk_entry); break; case PAGE_CHUNK_TYPE_UNUSED: msg_buf = CHUNK_TO_MISCQ_MSG(vb->unused_page_chunk_hdr); msg_len = sizeof(struct virtio_balloon_miscq_hdr) + sizeof(struct virtio_balloon_page_chunk_hdr) + nr_chunks * sizeof(struct virtio_balloon_page_chunk_entry); break; default: dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", __func__, type); return; } >> Please check the patch for implementing the unused page chunk, >> it will be clear. If necessary, I can put "buf = (void *)hdr - len" from >> that patch. > Exactly. And all this pointer math is very messy. Please look for ways > to clean it. It's generally easy to fill structures: > > struct foo *foo = kmalloc(..., sizeof(*foo) + n * sizeof(foo->a[0])); > for (i = 0; i < n; ++i) > foo->a[i] = b; > > this is the kind of code that's easy to understand and it's > obvious there are no overflows and no info leaks here. > OK, will take your suggestion: struct virtio_balloon_page_chunk { struct virtio_balloon_page_chunk_hdr hdr; struct virtio_balloon_page_chunk_entry entries[]; }; >>>> + len += sizeof(struct virtio_balloon_page_chunk_hdr); >>>> + len += hdr->chunks * sizeof(struct virtio_balloon_page_chunk); >>>> + sg_init_table(&sg, 1); >>>> + sg_set_buf(&sg, buf, len); >>>> + if (!virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL)) { >>>> + virtqueue_kick(vq); >>>> + if (busy_wait) >>>> + while (!virtqueue_get_buf(vq, &len) && >>>> + !virtqueue_is_broken(vq)) >>>> + cpu_relax(); >>>> + else >>>> + wait_event(vb->acked, virtqueue_get_buf(vq, &len)); >>>> + hdr->chunks = 0; >>> Why zero it here after device used it? Better to zero before use. >> hdr->chunks tells the host how many chunks are there in the payload. >> After the device use it, it is ready to zero it. > It's rather confusing. Try to pass # of chunks around > in some other way. Not sure if this was explained clearly - we just let the chunk msg hdr indicates the # of chunks in the payload. I think this should be a pretty normal usage, like the network UDP hdr, which uses a length field to indicate the packet length. >>>> + } >>>> +} >>>> + >>>> +static void add_one_chunk(struct virtio_balloon *vb, struct virtqueue *vq, >>>> + int type, u64 base, u64 size) >>> what are the units here? Looks like it's in 4kbyte units? >> what is the "unit" you referred to? >> This is the function to add one chunk, base pfn and size of the chunk are >> supplied to the function. >> > Are both size and base in bytes then? > But you do not send them to host as is, you shift them for some reason > before sending them to host. > Not in bytes actually. base is a base pfn, which is the starting address of the continuous pfns. Size is the chunk size, which is the number of continuous pfns. They are shifted based on the chunk format we agreed before: -------------------------------------------------------- | Base (52 bit) | Rsvd (12 bit) | -------------------------------------------------------- -------------------------------------------------------- | Size (52 bit) | Rsvd (12 bit) | -------------------------------------------------------- Here, the pfn will be the balloon page pfn (4KB).In this way, the host doesn't need to know PAGE_SIZE of the guest. >>>> + if (zero >= end) >>>> + chunk_size = end - one; >>>> + else >>>> + chunk_size = zero - one; >>>> + >>>> + if (chunk_size) >>>> + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, >>>> + pfn_start + one, chunk_size); >>> Still not so what does a bit refer to? page or 4kbytes? >>> I think it should be a page. >> A bit in the bitmap corresponds to a pfn of a balloon page(4KB). > That's a waste on systems with large page sizes, and it does not > look like you handle that case correctly. OK, I will change the bitmap to be PAGE_SIZE based here, instead of BALLOON_PAGE_SIZE based. When convert them into chunks, making it based on BALLOON_PAGE_SIZE. Best, Wei
Hi Michael, could you please give some feedback? On Monday, April 17, 2017 11:35 AM, Wei Wang wrote: > On 04/15/2017 05:38 AM, Michael S. Tsirkin wrote: > > On Fri, Apr 14, 2017 at 04:37:52PM +0800, Wei Wang wrote: > >> On 04/14/2017 12:34 AM, Michael S. Tsirkin wrote: > >>> On Thu, Apr 13, 2017 at 05:35:05PM +0800, Wei Wang wrote: > >>> > >>> So we don't need the bitmap to talk to host, it is just a data > >>> structure we chose to maintain lists of pages, right? > >> Right. bitmap is the way to gather pages to chunk. > >> It's only needed in the balloon page case. > >> For the unused page case, we don't need it, since the free page > >> blocks are already chunks. > >> > >>> OK as far as it goes but you need much better isolation for it. > >>> Build a data structure with APIs such as _init, _cleanup, _add, > >>> _clear, _find_first, _find_next. > >>> Completely unrelated to pages, it just maintains bits. > >>> Then use it here. > >>> > >>> > >>>> static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES; > >>>> module_param(oom_pages, int, S_IRUSR | S_IWUSR); > >>>> MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); @@ -50,6 > >>>> +54,10 @@ MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); > >>>> static struct vfsmount *balloon_mnt; > >>>> #endif > >>>> +/* Types of pages to chunk */ > >>>> +#define PAGE_CHUNK_TYPE_BALLOON 0 > >>>> + > >>> Doesn't look like you are ever adding more types in this patchset. > >>> Pls keep code simple, generalize it later. > >>> > >> "#define PAGE_CHUNK_TYPE_UNUSED 1" is added in another patch. > > I would say add the extra code there too. Or maybe we can avoid adding > > it altogether. > > I'm trying to have the two features( i.e. "balloon pages" and "unused pages") > decoupled while trying to use common functions to deal with the commonalities. > That's the reason to define the above macro. > Without the macro, we will need to have separate functions, for example, > instead of one "add_one_chunk()", we need to have > add_one_balloon_page_chunk() and add_one_unused_page_chunk(), and some > of the implementations will be kind of duplicate in the two functions. > Probably we can add it when the second feature comes to the code. > > > > >> Types of page to chunk are treated differently. Different types of > >> page chunks are sent to the host via different protocols. > >> > >> 1) PAGE_CHUNK_TYPE_BALLOON: Ballooned (i.e. inflated/deflated) pages > >> to chunk. For the ballooned type, it uses the basic chunk msg format: > >> > >> virtio_balloon_page_chunk_hdr + > >> virtio_balloon_page_chunk * MAX_PAGE_CHUNKS > >> > >> 2) PAGE_CHUNK_TYPE_UNUSED: unused pages to chunk. It uses this miscq > >> msg > >> format: > >> miscq_hdr + > >> virtio_balloon_page_chunk_hdr + > >> virtio_balloon_page_chunk * MAX_PAGE_CHUNKS > >> > >> The chunk msg is actually the payload of the miscq msg. > >> > >> > > So just combine the two message formats and then it'll all be easier? > > > > Yes, it'll be simple with only one msg format. But the problem I see here is that > miscq hdr is something necessary for the "unused page" > usage, but not needed by the "balloon page" usage. To be more precise, struct > virtio_balloon_miscq_hdr { > __le16 cmd; > __le16 flags; > }; > 'cmd' specifies the command from the miscq (I envision that miscq will be > further used to handle other possible miscellaneous requests either from the > host or to the host), so 'cmd' is necessary for the miscq. But the inflateq is > exclusively used for inflating pages, so adding a command to it would be > redundant and look a little bewildered there. > 'flags': We currently use bit 0 of flags to indicate the completion ofa command, > this is also useful in the "unused page" usage, and not needed by the "balloon > page" usage. > >>>> +#define MAX_PAGE_CHUNKS 4096 > >>> This is an order-4 allocation. I'd make it 4095 and then it's an > >>> order-3 one. > >> Sounds good, thanks. > >> I think it would be better to make it 4090. Leave some space for the > >> hdr as well. > > And miscq hdr. In fact just let compiler do the math - something like: > > (8 * PAGE_SIZE - sizeof(hdr)) / sizeof(chunk) > Agree, thanks. > > > > > I skimmed explanation of algorithms below but please make sure code > > speaks for itself and add comments inline to document it. > > Whenever you answered me inline this is where you want to try to make > > code clearer and add comments. > > > > Also, pls find ways to abstract the data structure so we don't need to > > deal with its internals all over the code. > > > > > > .... > > > >>>> { > >>>> struct scatterlist sg; > >>>> + struct virtio_balloon_page_chunk_hdr *hdr; > >>>> + void *buf; > >>>> unsigned int len; > >>>> - sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); > >>>> + switch (type) { > >>>> + case PAGE_CHUNK_TYPE_BALLOON: > >>>> + hdr = vb->balloon_page_chunk_hdr; > >>>> + len = 0; > >>>> + break; > >>>> + default: > >>>> + dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown > pages\n", > >>>> + __func__, type); > >>>> + return; > >>>> + } > >>>> - /* We should always be able to add one buffer to an empty queue. */ > >>>> - virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); > >>>> - virtqueue_kick(vq); > >>>> + buf = (void *)hdr - len; > >>> Moving back to before the header? How can this make sense? > >>> It works fine since len is 0, so just buf = hdr. > >>> > >> For the unused page chunk case, it follows its own protocol: > >> miscq_hdr + payload(chunk msg). > >> "buf = (void *)hdr - len" moves the buf pointer to the miscq_hdr, > >> to send the entire miscq msg. > > Well just pass the correct pointer in. > > > OK. The miscq msg is > { > miscq_hdr; > chunk_msg; > } > > We can probably change the code like this: > > #define CHUNK_TO_MISCQ_MSG(chunk) (chunk - sizeof(struct > virtio_balloon_miscq_hdr)) > > switch (type) { > case PAGE_CHUNK_TYPE_BALLOON: > msg_buf = vb->balloon_page_chunk_hdr; > msg_len = sizeof(struct virtio_balloon_page_chunk_hdr) + > nr_chunks * sizeof(struct virtio_balloon_page_chunk_entry); > break; > case PAGE_CHUNK_TYPE_UNUSED: > msg_buf = CHUNK_TO_MISCQ_MSG(vb->unused_page_chunk_hdr); > msg_len = sizeof(struct virtio_balloon_miscq_hdr) + sizeof(struct > virtio_balloon_page_chunk_hdr) + > nr_chunks * sizeof(struct virtio_balloon_page_chunk_entry); > break; > default: > dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", > __func__, type); > return; > } > > > > >> Please check the patch for implementing the unused page chunk, it > >> will be clear. If necessary, I can put "buf = (void *)hdr - len" from > >> that patch. > > Exactly. And all this pointer math is very messy. Please look for ways > > to clean it. It's generally easy to fill structures: > > > > struct foo *foo = kmalloc(..., sizeof(*foo) + n * sizeof(foo->a[0])); > > for (i = 0; i < n; ++i) > > foo->a[i] = b; > > > > this is the kind of code that's easy to understand and it's obvious > > there are no overflows and no info leaks here. > > > OK, will take your suggestion: > > struct virtio_balloon_page_chunk { > struct virtio_balloon_page_chunk_hdr hdr; > struct virtio_balloon_page_chunk_entry entries[]; }; > > > >>>> + len += sizeof(struct virtio_balloon_page_chunk_hdr); > >>>> + len += hdr->chunks * sizeof(struct virtio_balloon_page_chunk); > >>>> + sg_init_table(&sg, 1); > >>>> + sg_set_buf(&sg, buf, len); > >>>> + if (!virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL)) { > >>>> + virtqueue_kick(vq); > >>>> + if (busy_wait) > >>>> + while (!virtqueue_get_buf(vq, &len) && > >>>> + !virtqueue_is_broken(vq)) > >>>> + cpu_relax(); > >>>> + else > >>>> + wait_event(vb->acked, virtqueue_get_buf(vq, &len)); > >>>> + hdr->chunks = 0; > >>> Why zero it here after device used it? Better to zero before use. > >> hdr->chunks tells the host how many chunks are there in the payload. > >> After the device use it, it is ready to zero it. > > It's rather confusing. Try to pass # of chunks around in some other > > way. > > Not sure if this was explained clearly - we just let the chunk msg hdr indicates > the # of chunks in the payload. I think this should be a pretty normal usage, like > the network UDP hdr, which uses a length field to indicate the packet length. > > >>>> + } > >>>> +} > >>>> + > >>>> +static void add_one_chunk(struct virtio_balloon *vb, struct virtqueue *vq, > >>>> + int type, u64 base, u64 size) > >>> what are the units here? Looks like it's in 4kbyte units? > >> what is the "unit" you referred to? > >> This is the function to add one chunk, base pfn and size of the chunk > >> are supplied to the function. > >> > > Are both size and base in bytes then? > > But you do not send them to host as is, you shift them for some reason > > before sending them to host. > > > Not in bytes actually. base is a base pfn, which is the starting address of the > continuous pfns. Size is the chunk size, which is the number of continuous pfns. > > They are shifted based on the chunk format we agreed before: > > -------------------------------------------------------- > | Base (52 bit) | Rsvd (12 bit) | > -------------------------------------------------------- > -------------------------------------------------------- > | Size (52 bit) | Rsvd (12 bit) | > -------------------------------------------------------- > > > Here, the pfn will be the balloon page pfn (4KB).In this way, the host doesn't > need to know PAGE_SIZE of the guest. > > > > >>>> + if (zero >= end) > >>>> + chunk_size = end - one; > >>>> + else > >>>> + chunk_size = zero - one; > >>>> + > >>>> + if (chunk_size) > >>>> + add_one_chunk(vb, vq, > PAGE_CHUNK_TYPE_BALLOON, > >>>> + pfn_start + one, chunk_size); > >>> Still not so what does a bit refer to? page or 4kbytes? > >>> I think it should be a page. > >> A bit in the bitmap corresponds to a pfn of a balloon page(4KB). > > That's a waste on systems with large page sizes, and it does not look > > like you handle that case correctly. > > OK, I will change the bitmap to be PAGE_SIZE based here, instead of > BALLOON_PAGE_SIZE based. When convert them into chunks, making it based > on BALLOON_PAGE_SIZE. > > > Best, > Wei > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: virtio-dev-unsubscribe@lists.oasis-open.org > For additional commands, e-mail: virtio-dev-help@lists.oasis-open.org
On Wed, Apr 26, 2017 at 11:03:34AM +0000, Wang, Wei W wrote:
> Hi Michael, could you please give some feedback?
I'm sorry, I'm not sure feedback on what you are requesting.
The interface looks reasonable now, even though there's
a way to make it even simpler if we can limit chunk size
to 2G (in fact 4G - 1). Do you think we can live with this
limitation?
But the code still needs some cleanup.
On 04/27/2017 07:20 AM, Michael S. Tsirkin wrote: > On Wed, Apr 26, 2017 at 11:03:34AM +0000, Wang, Wei W wrote: >> Hi Michael, could you please give some feedback? > I'm sorry, I'm not sure feedback on what you are requesting. Oh, just some trivial things (e.g. use a field in the header, hdr->chunks to indicate the number of chunks in the payload) that wasn't confirmed. I will prepare the new version with fixing the agreed issues, and we can continue to discuss those parts if you still find them improper. > > The interface looks reasonable now, even though there's > a way to make it even simpler if we can limit chunk size > to 2G (in fact 4G - 1). Do you think we can live with this > limitation? Yes, I think we can. So, is it good to change to use the previous 64-bit chunk format (52-bit base + 12-bit size)? > > But the code still needs some cleanup. > OK. We'll also still to discuss your comments in the patch 05. Best, Wei
On Thu, Apr 27, 2017 at 02:31:49PM +0800, Wei Wang wrote: > On 04/27/2017 07:20 AM, Michael S. Tsirkin wrote: > > On Wed, Apr 26, 2017 at 11:03:34AM +0000, Wang, Wei W wrote: > > > Hi Michael, could you please give some feedback? > > I'm sorry, I'm not sure feedback on what you are requesting. > Oh, just some trivial things (e.g. use a field in the > header, hdr->chunks to indicate the number of chunks > in the payload) that wasn't confirmed. > > I will prepare the new version with fixing the agreed issues, > and we can continue to discuss those parts if you still find > them improper. > > > > > > The interface looks reasonable now, even though there's > > a way to make it even simpler if we can limit chunk size > > to 2G (in fact 4G - 1). Do you think we can live with this > > limitation? > Yes, I think we can. So, is it good to change to use the > previous 64-bit chunk format (52-bit base + 12-bit size)? This isn't what I meant. virtio ring has descriptors with a 64 bit address and 32 bit size. If size < 4g is not a significant limitation, why not just use that to pass address/size in a standard s/g list, possibly using INDIRECT? > > > > > But the code still needs some cleanup. > > > > OK. We'll also still to discuss your comments in the patch 05. > > Best, > Wei
On 05/06/2017 06:26 AM, Michael S. Tsirkin wrote: > On Thu, Apr 27, 2017 at 02:31:49PM +0800, Wei Wang wrote: > > On 04/27/2017 07:20 AM, Michael S. Tsirkin wrote: > > > On Wed, Apr 26, 2017 at 11:03:34AM +0000, Wang, Wei W wrote: > > > > Hi Michael, could you please give some feedback? > > > I'm sorry, I'm not sure feedback on what you are requesting. > > Oh, just some trivial things (e.g. use a field in the header, > > hdr->chunks to indicate the number of chunks in the payload) that > > wasn't confirmed. > > > > I will prepare the new version with fixing the agreed issues, and we > > can continue to discuss those parts if you still find them improper. > > > > > > > > > > The interface looks reasonable now, even though there's a way to > > > make it even simpler if we can limit chunk size to 2G (in fact 4G - > > > 1). Do you think we can live with this limitation? > > Yes, I think we can. So, is it good to change to use the previous > > 64-bit chunk format (52-bit base + 12-bit size)? > > This isn't what I meant. virtio ring has descriptors with a 64 bit address and 32 bit > size. > > If size < 4g is not a significant limitation, why not just use that to pass > address/size in a standard s/g list, possibly using INDIRECT? OK, I see your point, thanks. Post the two options here for an analysis: Option1 (what we have now): struct virtio_balloon_page_chunk { __le64 chunk_num; struct virtio_balloon_page_chunk_entry entry[]; }; Option2: struct virtio_balloon_page_chunk { __le64 chunk_num; struct scatterlist entry[]; }; I don't have an issue to change it to Option2, but I would prefer Option1, because I think there is no be obvious difference between the two options, while Option1 appears to have little advantages here: 1) "struct virtio_balloon_page_chunk_entry" has smaller size than "struct scatterlist", so the same size of allocated page chunk buffer can hold more entry[] using Option1; 2) INDIRECT needs on demand kmalloc(); 3) no 4G size limit; What do you think? Best, Wei
On Sun, May 07, 2017 at 04:19:28AM +0000, Wang, Wei W wrote: > On 05/06/2017 06:26 AM, Michael S. Tsirkin wrote: > > On Thu, Apr 27, 2017 at 02:31:49PM +0800, Wei Wang wrote: > > > On 04/27/2017 07:20 AM, Michael S. Tsirkin wrote: > > > > On Wed, Apr 26, 2017 at 11:03:34AM +0000, Wang, Wei W wrote: > > > > > Hi Michael, could you please give some feedback? > > > > I'm sorry, I'm not sure feedback on what you are requesting. > > > Oh, just some trivial things (e.g. use a field in the header, > > > hdr->chunks to indicate the number of chunks in the payload) that > > > wasn't confirmed. > > > > > > I will prepare the new version with fixing the agreed issues, and we > > > can continue to discuss those parts if you still find them improper. > > > > > > > > > > > > > > The interface looks reasonable now, even though there's a way to > > > > make it even simpler if we can limit chunk size to 2G (in fact 4G - > > > > 1). Do you think we can live with this limitation? > > > Yes, I think we can. So, is it good to change to use the previous > > > 64-bit chunk format (52-bit base + 12-bit size)? > > > > This isn't what I meant. virtio ring has descriptors with a 64 bit address and 32 bit > > size. > > > > If size < 4g is not a significant limitation, why not just use that to pass > > address/size in a standard s/g list, possibly using INDIRECT? > > OK, I see your point, thanks. Post the two options here for an analysis: > Option1 (what we have now): > struct virtio_balloon_page_chunk { > __le64 chunk_num; > struct virtio_balloon_page_chunk_entry entry[]; > }; > Option2: > struct virtio_balloon_page_chunk { > __le64 chunk_num; > struct scatterlist entry[]; > }; This isn't what I meant really :) I meant vring_desc. > I don't have an issue to change it to Option2, but I would prefer Option1, > because I think there is no be obvious difference between the two options, > while Option1 appears to have little advantages here: > 1) "struct virtio_balloon_page_chunk_entry" has smaller size than > "struct scatterlist", so the same size of allocated page chunk buffer > can hold more entry[] using Option1; > 2) INDIRECT needs on demand kmalloc(); Within alloc_indirect? We can fix that with a separate patch. > 3) no 4G size limit; Do you see lots of >=4g chunks in practice? > What do you think? > > Best, > Wei > > OTOH using existing vring APIs handles things like DMA transparently.
On 05/09/2017 01:40 AM, Michael S. Tsirkin wrote: > On Sun, May 07, 2017 at 04:19:28AM +0000, Wang, Wei W wrote: >> On 05/06/2017 06:26 AM, Michael S. Tsirkin wrote: >>> On Thu, Apr 27, 2017 at 02:31:49PM +0800, Wei Wang wrote: >>>> On 04/27/2017 07:20 AM, Michael S. Tsirkin wrote: >>>>> On Wed, Apr 26, 2017 at 11:03:34AM +0000, Wang, Wei W wrote: >>>>>> Hi Michael, could you please give some feedback? >>>>> I'm sorry, I'm not sure feedback on what you are requesting. >>>> Oh, just some trivial things (e.g. use a field in the header, >>>> hdr->chunks to indicate the number of chunks in the payload) that >>>> wasn't confirmed. >>>> >>>> I will prepare the new version with fixing the agreed issues, and we >>>> can continue to discuss those parts if you still find them improper. >>>> >>>> >>>>> The interface looks reasonable now, even though there's a way to >>>>> make it even simpler if we can limit chunk size to 2G (in fact 4G - >>>>> 1). Do you think we can live with this limitation? >>>> Yes, I think we can. So, is it good to change to use the previous >>>> 64-bit chunk format (52-bit base + 12-bit size)? >>> This isn't what I meant. virtio ring has descriptors with a 64 bit address and 32 bit >>> size. >>> >>> If size < 4g is not a significant limitation, why not just use that to pass >>> address/size in a standard s/g list, possibly using INDIRECT? >> OK, I see your point, thanks. Post the two options here for an analysis: >> Option1 (what we have now): >> struct virtio_balloon_page_chunk { >> __le64 chunk_num; >> struct virtio_balloon_page_chunk_entry entry[]; >> }; >> Option2: >> struct virtio_balloon_page_chunk { >> __le64 chunk_num; >> struct scatterlist entry[]; >> }; > This isn't what I meant really :) I meant vring_desc. OK. Repost the code change: Option2: struct virtio_balloon_page_chunk { __le64 chunk_num; struct ving_desc entry[]; }; We pre-allocate a table of desc, and each desc is used to hold a chunk. In that case, the virtqueue_add() function, which deals with sg, is not usable for us. We may need to add a new one, virtqueue_add_indirect_desc(), to add a pre-allocated indirect descriptor table to vring. > >> I don't have an issue to change it to Option2, but I would prefer Option1, >> because I think there is no be obvious difference between the two options, >> while Option1 appears to have little advantages here: >> 1) "struct virtio_balloon_page_chunk_entry" has smaller size than >> "struct scatterlist", so the same size of allocated page chunk buffer >> can hold more entry[] using Option1; >> 2) INDIRECT needs on demand kmalloc(); > Within alloc_indirect? We can fix that with a separate patch. > > >> 3) no 4G size limit; > Do you see lots of >=4g chunks in practice? It wouldn't be much in practice, but we still need the extra code to handle the case - break larger chunks into less-than 4g ones. Best, Wei
diff --git a/drivers/virtio/virtio_balloon.c b/drivers/virtio/virtio_balloon.c index f59cb4f..5e2e7cc 100644 --- a/drivers/virtio/virtio_balloon.c +++ b/drivers/virtio/virtio_balloon.c @@ -42,6 +42,10 @@ #define OOM_VBALLOON_DEFAULT_PAGES 256 #define VIRTBALLOON_OOM_NOTIFY_PRIORITY 80 +#define PAGE_BMAP_SIZE (8 * PAGE_SIZE) +#define PFNS_PER_PAGE_BMAP (PAGE_BMAP_SIZE * BITS_PER_BYTE) +#define PAGE_BMAP_COUNT_MAX 32 + static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES; module_param(oom_pages, int, S_IRUSR | S_IWUSR); MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); @@ -50,6 +54,10 @@ MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); static struct vfsmount *balloon_mnt; #endif +/* Types of pages to chunk */ +#define PAGE_CHUNK_TYPE_BALLOON 0 + +#define MAX_PAGE_CHUNKS 4096 struct virtio_balloon { struct virtio_device *vdev; struct virtqueue *inflate_vq, *deflate_vq, *stats_vq; @@ -78,6 +86,32 @@ struct virtio_balloon { /* Synchronize access/update to this struct virtio_balloon elements */ struct mutex balloon_lock; + /* + * Buffer for PAGE_CHUNK_TYPE_BALLOON: + * virtio_balloon_page_chunk_hdr + + * virtio_balloon_page_chunk * MAX_PAGE_CHUNKS + */ + struct virtio_balloon_page_chunk_hdr *balloon_page_chunk_hdr; + struct virtio_balloon_page_chunk *balloon_page_chunk; + + /* Bitmap used to record pages */ + unsigned long *page_bmap[PAGE_BMAP_COUNT_MAX]; + /* Number of the allocated page_bmap */ + unsigned int page_bmaps; + + /* + * The allocated page_bmap size may be smaller than the pfn range of + * the ballooned pages. In this case, we need to use the page_bmap + * multiple times to cover the entire pfn range. It's like using a + * short ruler several times to finish measuring a long object. + * The start location of the ruler in the next measurement is the end + * location of the ruler in the previous measurement. + * + * pfn_max & pfn_min: forms the pfn range of the ballooned pages + * pfn_start & pfn_stop: records the start and stop pfn in each cover + */ + unsigned long pfn_min, pfn_max, pfn_start, pfn_stop; + /* The array of pfns we tell the Host about. */ unsigned int num_pfns; __virtio32 pfns[VIRTIO_BALLOON_ARRAY_PFNS_MAX]; @@ -110,20 +144,201 @@ static void balloon_ack(struct virtqueue *vq) wake_up(&vb->acked); } -static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq) +static inline void init_page_bmap_range(struct virtio_balloon *vb) +{ + vb->pfn_min = ULONG_MAX; + vb->pfn_max = 0; +} + +static inline void update_page_bmap_range(struct virtio_balloon *vb, + struct page *page) +{ + unsigned long balloon_pfn = page_to_balloon_pfn(page); + + vb->pfn_min = min(balloon_pfn, vb->pfn_min); + vb->pfn_max = max(balloon_pfn, vb->pfn_max); +} + +/* The page_bmap size is extended by adding more number of page_bmap */ +static void extend_page_bmap_size(struct virtio_balloon *vb, + unsigned long pfns) +{ + int i, bmaps; + unsigned long bmap_len; + + bmap_len = ALIGN(pfns, BITS_PER_LONG) / BITS_PER_BYTE; + bmap_len = ALIGN(bmap_len, PAGE_BMAP_SIZE); + bmaps = min((int)(bmap_len / PAGE_BMAP_SIZE), + PAGE_BMAP_COUNT_MAX); + + for (i = 1; i < bmaps; i++) { + vb->page_bmap[i] = kmalloc(PAGE_BMAP_SIZE, GFP_KERNEL); + if (vb->page_bmap[i]) + vb->page_bmaps++; + else + break; + } +} + +static void free_extended_page_bmap(struct virtio_balloon *vb) +{ + int i, bmaps = vb->page_bmaps; + + for (i = 1; i < bmaps; i++) { + kfree(vb->page_bmap[i]); + vb->page_bmap[i] = NULL; + vb->page_bmaps--; + } +} + +static void free_page_bmap(struct virtio_balloon *vb) +{ + int i; + + for (i = 0; i < vb->page_bmaps; i++) + kfree(vb->page_bmap[i]); +} + +static void clear_page_bmap(struct virtio_balloon *vb) +{ + int i; + + for (i = 0; i < vb->page_bmaps; i++) + memset(vb->page_bmap[i], 0, PAGE_BMAP_SIZE); +} + +static void send_page_chunks(struct virtio_balloon *vb, struct virtqueue *vq, + int type, bool busy_wait) { struct scatterlist sg; + struct virtio_balloon_page_chunk_hdr *hdr; + void *buf; unsigned int len; - sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); + switch (type) { + case PAGE_CHUNK_TYPE_BALLOON: + hdr = vb->balloon_page_chunk_hdr; + len = 0; + break; + default: + dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", + __func__, type); + return; + } - /* We should always be able to add one buffer to an empty queue. */ - virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); - virtqueue_kick(vq); + buf = (void *)hdr - len; + len += sizeof(struct virtio_balloon_page_chunk_hdr); + len += hdr->chunks * sizeof(struct virtio_balloon_page_chunk); + sg_init_table(&sg, 1); + sg_set_buf(&sg, buf, len); + if (!virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL)) { + virtqueue_kick(vq); + if (busy_wait) + while (!virtqueue_get_buf(vq, &len) && + !virtqueue_is_broken(vq)) + cpu_relax(); + else + wait_event(vb->acked, virtqueue_get_buf(vq, &len)); + hdr->chunks = 0; + } +} + +static void add_one_chunk(struct virtio_balloon *vb, struct virtqueue *vq, + int type, u64 base, u64 size) +{ + struct virtio_balloon_page_chunk_hdr *hdr; + struct virtio_balloon_page_chunk *chunk; + + switch (type) { + case PAGE_CHUNK_TYPE_BALLOON: + hdr = vb->balloon_page_chunk_hdr; + chunk = vb->balloon_page_chunk; + break; + default: + dev_warn(&vb->vdev->dev, "%s: chunk %d of unknown pages\n", + __func__, type); + return; + } + chunk = chunk + hdr->chunks; + chunk->base = cpu_to_le64(base << VIRTIO_BALLOON_CHUNK_BASE_SHIFT); + chunk->size = cpu_to_le64(size << VIRTIO_BALLOON_CHUNK_SIZE_SHIFT); + hdr->chunks++; + if (hdr->chunks == MAX_PAGE_CHUNKS) + send_page_chunks(vb, vq, type, false); +} + +static void chunking_pages_from_bmap(struct virtio_balloon *vb, + struct virtqueue *vq, + unsigned long pfn_start, + unsigned long *bmap, + unsigned long len) +{ + unsigned long pos = 0, end = len * BITS_PER_BYTE; + + while (pos < end) { + unsigned long one = find_next_bit(bmap, end, pos); + + if (one < end) { + unsigned long chunk_size, zero; + + zero = find_next_zero_bit(bmap, end, one + 1); + if (zero >= end) + chunk_size = end - one; + else + chunk_size = zero - one; + + if (chunk_size) + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, + pfn_start + one, chunk_size); + pos = one + chunk_size; + } else + break; + } +} + +static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq) +{ + if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS)) { + int pfns, page_bmaps, i; + unsigned long pfn_start, pfns_len; + + pfn_start = vb->pfn_start; + pfns = vb->pfn_stop - pfn_start + 1; + pfns = roundup(roundup(pfns, BITS_PER_LONG), + PFNS_PER_PAGE_BMAP); + page_bmaps = pfns / PFNS_PER_PAGE_BMAP; + pfns_len = pfns / BITS_PER_BYTE; + + for (i = 0; i < page_bmaps; i++) { + unsigned int bmap_len = PAGE_BMAP_SIZE; + + /* The last one takes the leftover only */ + if (i + 1 == page_bmaps) + bmap_len = pfns_len - PAGE_BMAP_SIZE * i; + + chunking_pages_from_bmap(vb, vq, pfn_start + + i * PFNS_PER_PAGE_BMAP, + vb->page_bmap[i], bmap_len); + } + if (vb->balloon_page_chunk_hdr->chunks > 0) + send_page_chunks(vb, vq, PAGE_CHUNK_TYPE_BALLOON, + false); + } else { + struct scatterlist sg; + unsigned int len; - /* When host has read buffer, this completes via balloon_ack */ - wait_event(vb->acked, virtqueue_get_buf(vq, &len)); + sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); + /* + * We should always be able to add one buffer to an empty + * queue. + */ + virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); + virtqueue_kick(vq); + + /* When host has read buffer, this completes via balloon_ack */ + wait_event(vb->acked, virtqueue_get_buf(vq, &len)); + } } static void set_page_pfns(struct virtio_balloon *vb, @@ -131,20 +346,73 @@ static void set_page_pfns(struct virtio_balloon *vb, { unsigned int i; - /* Set balloon pfns pointing at this page. - * Note that the first pfn points at start of the page. */ + /* + * Set balloon pfns pointing at this page. + * Note that the first pfn points at start of the page. + */ for (i = 0; i < VIRTIO_BALLOON_PAGES_PER_PAGE; i++) pfns[i] = cpu_to_virtio32(vb->vdev, page_to_balloon_pfn(page) + i); } +static void set_page_bmap(struct virtio_balloon *vb, + struct list_head *pages, struct virtqueue *vq) +{ + unsigned long pfn_start, pfn_stop; + struct page *page; + bool found; + + vb->pfn_min = rounddown(vb->pfn_min, BITS_PER_LONG); + vb->pfn_max = roundup(vb->pfn_max, BITS_PER_LONG); + + extend_page_bmap_size(vb, vb->pfn_max - vb->pfn_min + 1); + pfn_start = vb->pfn_min; + + while (pfn_start < vb->pfn_max) { + pfn_stop = pfn_start + PFNS_PER_PAGE_BMAP * vb->page_bmaps; + pfn_stop = pfn_stop < vb->pfn_max ? pfn_stop : vb->pfn_max; + + vb->pfn_start = pfn_start; + clear_page_bmap(vb); + found = false; + + list_for_each_entry(page, pages, lru) { + unsigned long bmap_idx, bmap_pos, balloon_pfn; + + balloon_pfn = page_to_balloon_pfn(page); + if (balloon_pfn < pfn_start || balloon_pfn > pfn_stop) + continue; + bmap_idx = (balloon_pfn - pfn_start) / + PFNS_PER_PAGE_BMAP; + bmap_pos = (balloon_pfn - pfn_start) % + PFNS_PER_PAGE_BMAP; + set_bit(bmap_pos, vb->page_bmap[bmap_idx]); + + found = true; + } + if (found) { + vb->pfn_stop = pfn_stop; + tell_host(vb, vq); + } + pfn_start = pfn_stop; + } + free_extended_page_bmap(vb); +} + static unsigned fill_balloon(struct virtio_balloon *vb, size_t num) { struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info; unsigned num_allocated_pages; + bool chunking = virtio_has_feature(vb->vdev, + VIRTIO_BALLOON_F_BALLOON_CHUNKS); /* We can only do one array worth at a time. */ - num = min(num, ARRAY_SIZE(vb->pfns)); + if (chunking) { + init_page_bmap_range(vb); + } else { + /* We can only do one array worth at a time. */ + num = min(num, ARRAY_SIZE(vb->pfns)); + } mutex_lock(&vb->balloon_lock); for (vb->num_pfns = 0; vb->num_pfns < num; @@ -159,7 +427,10 @@ static unsigned fill_balloon(struct virtio_balloon *vb, size_t num) msleep(200); break; } - set_page_pfns(vb, vb->pfns + vb->num_pfns, page); + if (chunking) + update_page_bmap_range(vb, page); + else + set_page_pfns(vb, vb->pfns + vb->num_pfns, page); vb->num_pages += VIRTIO_BALLOON_PAGES_PER_PAGE; if (!virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM)) @@ -168,8 +439,13 @@ static unsigned fill_balloon(struct virtio_balloon *vb, size_t num) num_allocated_pages = vb->num_pfns; /* Did we get any? */ - if (vb->num_pfns != 0) - tell_host(vb, vb->inflate_vq); + if (vb->num_pfns != 0) { + if (chunking) + set_page_bmap(vb, &vb_dev_info->pages, + vb->inflate_vq); + else + tell_host(vb, vb->inflate_vq); + } mutex_unlock(&vb->balloon_lock); return num_allocated_pages; @@ -195,6 +471,13 @@ static unsigned leak_balloon(struct virtio_balloon *vb, size_t num) struct page *page; struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info; LIST_HEAD(pages); + bool chunking = virtio_has_feature(vb->vdev, + VIRTIO_BALLOON_F_BALLOON_CHUNKS); + if (chunking) + init_page_bmap_range(vb); + else + /* We can only do one array worth at a time. */ + num = min(num, ARRAY_SIZE(vb->pfns)); /* We can only do one array worth at a time. */ num = min(num, ARRAY_SIZE(vb->pfns)); @@ -208,6 +491,10 @@ static unsigned leak_balloon(struct virtio_balloon *vb, size_t num) if (!page) break; set_page_pfns(vb, vb->pfns + vb->num_pfns, page); + if (chunking) + update_page_bmap_range(vb, page); + else + set_page_pfns(vb, vb->pfns + vb->num_pfns, page); list_add(&page->lru, &pages); vb->num_pages -= VIRTIO_BALLOON_PAGES_PER_PAGE; } @@ -218,8 +505,12 @@ static unsigned leak_balloon(struct virtio_balloon *vb, size_t num) * virtio_has_feature(vdev, VIRTIO_BALLOON_F_MUST_TELL_HOST); * is true, we *have* to do it in this order */ - if (vb->num_pfns != 0) - tell_host(vb, vb->deflate_vq); + if (vb->num_pfns != 0) { + if (chunking) + set_page_bmap(vb, &pages, vb->deflate_vq); + else + tell_host(vb, vb->deflate_vq); + } release_pages_balloon(vb, &pages); mutex_unlock(&vb->balloon_lock); return num_freed_pages; @@ -431,6 +722,13 @@ static int init_vqs(struct virtio_balloon *vb) } #ifdef CONFIG_BALLOON_COMPACTION + +static void tell_host_one_page(struct virtio_balloon *vb, + struct virtqueue *vq, struct page *page) +{ + add_one_chunk(vb, vq, PAGE_CHUNK_TYPE_BALLOON, page_to_pfn(page), 1); +} + /* * virtballoon_migratepage - perform the balloon page migration on behalf of * a compation thread. (called under page lock) @@ -454,6 +752,8 @@ static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info, { struct virtio_balloon *vb = container_of(vb_dev_info, struct virtio_balloon, vb_dev_info); + bool chunking = virtio_has_feature(vb->vdev, + VIRTIO_BALLOON_F_BALLOON_CHUNKS); unsigned long flags; /* @@ -475,16 +775,22 @@ static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info, vb_dev_info->isolated_pages--; __count_vm_event(BALLOON_MIGRATE); spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags); - vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; - set_page_pfns(vb, vb->pfns, newpage); - tell_host(vb, vb->inflate_vq); - + if (chunking) { + tell_host_one_page(vb, vb->inflate_vq, newpage); + } else { + vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; + set_page_pfns(vb, vb->pfns, newpage); + tell_host(vb, vb->inflate_vq); + } /* balloon's page migration 2nd step -- deflate "page" */ balloon_page_delete(page); - vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; - set_page_pfns(vb, vb->pfns, page); - tell_host(vb, vb->deflate_vq); - + if (chunking) { + tell_host_one_page(vb, vb->deflate_vq, page); + } else { + vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; + set_page_pfns(vb, vb->pfns, page); + tell_host(vb, vb->deflate_vq); + } mutex_unlock(&vb->balloon_lock); put_page(page); /* balloon reference */ @@ -511,6 +817,32 @@ static struct file_system_type balloon_fs = { #endif /* CONFIG_BALLOON_COMPACTION */ +static void balloon_page_chunk_init(struct virtio_balloon *vb) +{ + void *buf; + + /* + * By default, we allocate page_bmap[0] only. More page_bmap will be + * allocated on demand. + */ + vb->page_bmap[0] = kmalloc(PAGE_BMAP_SIZE, GFP_KERNEL); + buf = kmalloc(sizeof(struct virtio_balloon_page_chunk_hdr) + + sizeof(struct virtio_balloon_page_chunk) * + MAX_PAGE_CHUNKS, GFP_KERNEL); + if (!vb->page_bmap[0] || !buf) { + __virtio_clear_bit(vb->vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS); + kfree(vb->page_bmap[0]); + kfree(vb->balloon_page_chunk_hdr); + dev_warn(&vb->vdev->dev, "%s: failed\n", __func__); + } else { + vb->page_bmaps = 1; + vb->balloon_page_chunk_hdr = buf; + vb->balloon_page_chunk_hdr->chunks = 0; + vb->balloon_page_chunk = buf + + sizeof(struct virtio_balloon_page_chunk_hdr); + } +} + static int virtballoon_probe(struct virtio_device *vdev) { struct virtio_balloon *vb; @@ -533,6 +865,10 @@ static int virtballoon_probe(struct virtio_device *vdev) spin_lock_init(&vb->stop_update_lock); vb->stop_update = false; vb->num_pages = 0; + + if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_BALLOON_CHUNKS)) + balloon_page_chunk_init(vb); + mutex_init(&vb->balloon_lock); init_waitqueue_head(&vb->acked); vb->vdev = vdev; @@ -609,6 +945,7 @@ static void virtballoon_remove(struct virtio_device *vdev) cancel_work_sync(&vb->update_balloon_stats_work); remove_common(vb); + free_page_bmap(vb); if (vb->vb_dev_info.inode) iput(vb->vb_dev_info.inode); kfree(vb); @@ -649,6 +986,7 @@ static unsigned int features[] = { VIRTIO_BALLOON_F_MUST_TELL_HOST, VIRTIO_BALLOON_F_STATS_VQ, VIRTIO_BALLOON_F_DEFLATE_ON_OOM, + VIRTIO_BALLOON_F_BALLOON_CHUNKS, }; static struct virtio_driver virtio_balloon_driver = { diff --git a/include/uapi/linux/virtio_balloon.h b/include/uapi/linux/virtio_balloon.h index 343d7dd..be317b7 100644 --- a/include/uapi/linux/virtio_balloon.h +++ b/include/uapi/linux/virtio_balloon.h @@ -34,6 +34,7 @@ #define VIRTIO_BALLOON_F_MUST_TELL_HOST 0 /* Tell before reclaiming pages */ #define VIRTIO_BALLOON_F_STATS_VQ 1 /* Memory Stats virtqueue */ #define VIRTIO_BALLOON_F_DEFLATE_ON_OOM 2 /* Deflate balloon on OOM */ +#define VIRTIO_BALLOON_F_BALLOON_CHUNKS 3 /* Inflate/Deflate pages in chunks */ /* Size of a PFN in the balloon interface. */ #define VIRTIO_BALLOON_PFN_SHIFT 12 @@ -82,4 +83,16 @@ struct virtio_balloon_stat { __virtio64 val; } __attribute__((packed)); +struct virtio_balloon_page_chunk_hdr { + /* Number of chunks in the payload */ + __le32 chunks; +}; + +#define VIRTIO_BALLOON_CHUNK_BASE_SHIFT 12 +#define VIRTIO_BALLOON_CHUNK_SIZE_SHIFT 12 +struct virtio_balloon_page_chunk { + __le64 base; + __le64 size; +}; + #endif /* _LINUX_VIRTIO_BALLOON_H */