ore: RAID5 Write
[linux-3.10.git] / fs / exofs / ore_raid.c
index 8d4b93a..29c47e5 100644 (file)
  */
 
 #include <linux/gfp.h>
+#include <linux/async_tx.h>
 
 #include "ore_raid.h"
 
+#undef ORE_DBGMSG2
+#define ORE_DBGMSG2 ORE_DBGMSG
+
 struct page *_raid_page_alloc(void)
 {
        return alloc_page(GFP_KERNEL);
@@ -27,6 +31,236 @@ void _raid_page_free(struct page *p)
        __free_page(p);
 }
 
+/* This struct is forward declare in ore_io_state, but is private to here.
+ * It is put on ios->sp2d for RAID5/6 writes only. See _gen_xor_unit.
+ *
+ * __stripe_pages_2d is a 2d array of pages, and it is also a corner turn.
+ * Ascending page index access is sp2d(p-minor, c-major). But storage is
+ * sp2d[p-minor][c-major], so it can be properlly presented to the async-xor
+ * API.
+ */
+struct __stripe_pages_2d {
+       /* Cache some hot path repeated calculations */
+       unsigned parity;
+       unsigned data_devs;
+       unsigned pages_in_unit;
+
+       bool needed ;
+
+       /* Array size is pages_in_unit (layout->stripe_unit / PAGE_SIZE) */
+       struct __1_page_stripe {
+               bool alloc;
+               unsigned write_count;
+               struct async_submit_ctl submit;
+               struct dma_async_tx_descriptor *tx;
+
+               /* The size of this array is data_devs + parity */
+               struct page **pages;
+               struct page **scribble;
+               /* bool array, size of this array is data_devs */
+               char *page_is_read;
+       } _1p_stripes[];
+};
+
+/* This can get bigger then a page. So support multiple page allocations
+ * _sp2d_free should be called even if _sp2d_alloc fails (by returning
+ * none-zero).
+ */
+static int _sp2d_alloc(unsigned pages_in_unit, unsigned group_width,
+                      unsigned parity, struct __stripe_pages_2d **psp2d)
+{
+       struct __stripe_pages_2d *sp2d;
+       unsigned data_devs = group_width - parity;
+       struct _alloc_all_bytes {
+               struct __alloc_stripe_pages_2d {
+                       struct __stripe_pages_2d sp2d;
+                       struct __1_page_stripe _1p_stripes[pages_in_unit];
+               } __asp2d;
+               struct __alloc_1p_arrays {
+                       struct page *pages[group_width];
+                       struct page *scribble[group_width];
+                       char page_is_read[data_devs];
+               } __a1pa[pages_in_unit];
+       } *_aab;
+       struct __alloc_1p_arrays *__a1pa;
+       struct __alloc_1p_arrays *__a1pa_end;
+       const unsigned sizeof__a1pa = sizeof(_aab->__a1pa[0]);
+       unsigned num_a1pa, alloc_size, i;
+
+       /* FIXME: check these numbers in ore_verify_layout */
+       BUG_ON(sizeof(_aab->__asp2d) > PAGE_SIZE);
+       BUG_ON(sizeof__a1pa > PAGE_SIZE);
+
+       if (sizeof(*_aab) > PAGE_SIZE) {
+               num_a1pa = (PAGE_SIZE - sizeof(_aab->__asp2d)) / sizeof__a1pa;
+               alloc_size = sizeof(_aab->__asp2d) + sizeof__a1pa * num_a1pa;
+       } else {
+               num_a1pa = pages_in_unit;
+               alloc_size = sizeof(*_aab);
+       }
+
+       _aab = kzalloc(alloc_size, GFP_KERNEL);
+       if (unlikely(!_aab)) {
+               ORE_DBGMSG("!! Failed to alloc sp2d size=%d\n", alloc_size);
+               return -ENOMEM;
+       }
+
+       sp2d = &_aab->__asp2d.sp2d;
+       *psp2d = sp2d; /* From here Just call _sp2d_free */
+
+       __a1pa = _aab->__a1pa;
+       __a1pa_end = __a1pa + num_a1pa;
+
+       for (i = 0; i < pages_in_unit; ++i) {
+               if (unlikely(__a1pa >= __a1pa_end)) {
+                       num_a1pa = min_t(unsigned, PAGE_SIZE / sizeof__a1pa,
+                                                       pages_in_unit - i);
+
+                       __a1pa = kzalloc(num_a1pa * sizeof__a1pa, GFP_KERNEL);
+                       if (unlikely(!__a1pa)) {
+                               ORE_DBGMSG("!! Failed to _alloc_1p_arrays=%d\n",
+                                          num_a1pa);
+                               return -ENOMEM;
+                       }
+                       __a1pa_end = __a1pa + num_a1pa;
+                       /* First *pages is marked for kfree of the buffer */
+                       sp2d->_1p_stripes[i].alloc = true;
+               }
+
+               sp2d->_1p_stripes[i].pages = __a1pa->pages;
+               sp2d->_1p_stripes[i].scribble = __a1pa->scribble ;
+               sp2d->_1p_stripes[i].page_is_read = __a1pa->page_is_read;
+               ++__a1pa;
+       }
+
+       sp2d->parity = parity;
+       sp2d->data_devs = data_devs;
+       sp2d->pages_in_unit = pages_in_unit;
+       return 0;
+}
+
+static void _sp2d_reset(struct __stripe_pages_2d *sp2d,
+                       const struct _ore_r4w_op *r4w, void *priv)
+{
+       unsigned data_devs = sp2d->data_devs;
+       unsigned group_width = data_devs + sp2d->parity;
+       unsigned p;
+
+       if (!sp2d->needed)
+               return;
+
+       for (p = 0; p < sp2d->pages_in_unit; p++) {
+               struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
+
+               if (_1ps->write_count < group_width) {
+                       unsigned c;
+
+                       for (c = 0; c < data_devs; c++)
+                               if (_1ps->page_is_read[c]) {
+                                       struct page *page = _1ps->pages[c];
+
+                                       r4w->put_page(priv, page);
+                                       _1ps->page_is_read[c] = false;
+                               }
+               }
+
+               memset(_1ps->pages, 0, group_width * sizeof(*_1ps->pages));
+               _1ps->write_count = 0;
+               _1ps->tx = NULL;
+       }
+
+       sp2d->needed = false;
+}
+
+static void _sp2d_free(struct __stripe_pages_2d *sp2d)
+{
+       unsigned i;
+
+       if (!sp2d)
+               return;
+
+       for (i = 0; i < sp2d->pages_in_unit; ++i) {
+               if (sp2d->_1p_stripes[i].alloc)
+                       kfree(sp2d->_1p_stripes[i].pages);
+       }
+
+       kfree(sp2d);
+}
+
+static unsigned _sp2d_min_pg(struct __stripe_pages_2d *sp2d)
+{
+       unsigned p;
+
+       for (p = 0; p < sp2d->pages_in_unit; p++) {
+               struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
+
+               if (_1ps->write_count)
+                       return p;
+       }
+
+       return ~0;
+}
+
+static unsigned _sp2d_max_pg(struct __stripe_pages_2d *sp2d)
+{
+       unsigned p;
+
+       for (p = sp2d->pages_in_unit - 1; p >= 0; --p) {
+               struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
+
+               if (_1ps->write_count)
+                       return p;
+       }
+
+       return ~0;
+}
+
+static void _gen_xor_unit(struct __stripe_pages_2d *sp2d)
+{
+       unsigned p;
+       for (p = 0; p < sp2d->pages_in_unit; p++) {
+               struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
+
+               if (!_1ps->write_count)
+                       continue;
+
+               init_async_submit(&_1ps->submit,
+                       ASYNC_TX_XOR_ZERO_DST | ASYNC_TX_ACK,
+                       NULL,
+                       NULL, NULL,
+                       (addr_conv_t *)_1ps->scribble);
+
+               /* TODO: raid6 */
+               _1ps->tx = async_xor(_1ps->pages[sp2d->data_devs], _1ps->pages,
+                                    0, sp2d->data_devs, PAGE_SIZE,
+                                    &_1ps->submit);
+       }
+
+       for (p = 0; p < sp2d->pages_in_unit; p++) {
+               struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
+               /* NOTE: We wait for HW synchronously (I don't have such HW
+                * to test with.) Is parallelism needed with today's multi
+                * cores?
+                */
+               async_tx_issue_pending(_1ps->tx);
+       }
+}
+
+void _ore_add_stripe_page(struct __stripe_pages_2d *sp2d,
+                      struct ore_striping_info *si, struct page *page)
+{
+       struct __1_page_stripe *_1ps;
+
+       sp2d->needed = true;
+
+       _1ps = &sp2d->_1p_stripes[si->cur_pg];
+       _1ps->pages[si->cur_comp] = page;
+       ++_1ps->write_count;
+
+       si->cur_pg = (si->cur_pg + 1) % sp2d->pages_in_unit;
+       /* si->cur_comp is advanced outside at main loop */
+}
+
 void _ore_add_sg_seg(struct ore_per_dev_state *per_dev, unsigned cur_len,
                     bool not_last)
 {
@@ -76,6 +310,240 @@ void _ore_add_sg_seg(struct ore_per_dev_state *per_dev, unsigned cur_len,
        }
 }
 
+static int _alloc_read_4_write(struct ore_io_state *ios)
+{
+       struct ore_layout *layout = ios->layout;
+       int ret;
+       /* We want to only read those pages not in cache so worst case
+        * is a stripe populated with every other page
+        */
+       unsigned sgs_per_dev = ios->sp2d->pages_in_unit + 2;
+
+       ret = _ore_get_io_state(layout, ios->oc,
+                               layout->group_width * layout->mirrors_p1,
+                               sgs_per_dev, 0, &ios->ios_read_4_write);
+       return ret;
+}
+
+/* @si contains info of the to-be-inserted page. Update of @si should be
+ * maintained by caller. Specificaly si->dev, si->obj_offset, ...
+ */
+static int _add_to_read_4_write(struct ore_io_state *ios,
+                               struct ore_striping_info *si, struct page *page)
+{
+       struct request_queue *q;
+       struct ore_per_dev_state *per_dev;
+       struct ore_io_state *read_ios;
+       unsigned first_dev = si->dev - (si->dev %
+                         (ios->layout->group_width * ios->layout->mirrors_p1));
+       unsigned comp = si->dev - first_dev;
+       unsigned added_len;
+
+       if (!ios->ios_read_4_write) {
+               int ret = _alloc_read_4_write(ios);
+
+               if (unlikely(ret))
+                       return ret;
+       }
+
+       read_ios = ios->ios_read_4_write;
+       read_ios->numdevs = ios->layout->group_width * ios->layout->mirrors_p1;
+
+       per_dev = &read_ios->per_dev[comp];
+       if (!per_dev->length) {
+               per_dev->bio = bio_kmalloc(GFP_KERNEL,
+                                          ios->sp2d->pages_in_unit);
+               if (unlikely(!per_dev->bio)) {
+                       ORE_DBGMSG("Failed to allocate BIO size=%u\n",
+                                    ios->sp2d->pages_in_unit);
+                       return -ENOMEM;
+               }
+               per_dev->offset = si->obj_offset;
+               per_dev->dev = si->dev;
+       } else if (si->obj_offset != (per_dev->offset + per_dev->length)) {
+               u64 gap = si->obj_offset - (per_dev->offset + per_dev->length);
+
+               _ore_add_sg_seg(per_dev, gap, true);
+       }
+       q = osd_request_queue(ore_comp_dev(read_ios->oc, per_dev->dev));
+       added_len = bio_add_pc_page(q, per_dev->bio, page, PAGE_SIZE, 0);
+       if (unlikely(added_len != PAGE_SIZE)) {
+               ORE_DBGMSG("Failed to bio_add_pc_page bi_vcnt=%d\n",
+                             per_dev->bio->bi_vcnt);
+               return -ENOMEM;
+       }
+
+       per_dev->length += PAGE_SIZE;
+       return 0;
+}
+
+static void _mark_read4write_pages_uptodate(struct ore_io_state *ios, int ret)
+{
+       struct bio_vec *bv;
+       unsigned i, d;
+
+       /* loop on all devices all pages */
+       for (d = 0; d < ios->numdevs; d++) {
+               struct bio *bio = ios->per_dev[d].bio;
+
+               if (!bio)
+                       continue;
+
+               __bio_for_each_segment(bv, bio, i, 0) {
+                       struct page *page = bv->bv_page;
+
+                       SetPageUptodate(page);
+                       if (PageError(page))
+                               ClearPageError(page);
+               }
+       }
+}
+
+/* read_4_write is hacked to read the start of the first stripe and/or
+ * the end of the last stripe. If needed, with an sg-gap at each device/page.
+ * It is assumed to be called after the to_be_written pages of the first stripe
+ * are populating ios->sp2d[][]
+ *
+ * NOTE: We call ios->r4w->lock_fn for all pages needed for parity calculations
+ * These pages are held at sp2d[p].pages[c] but with
+ * sp2d[p].page_is_read[c] = true. At _sp2d_reset these pages are
+ * ios->r4w->lock_fn(). The ios->r4w->lock_fn might signal that the page is
+ * @uptodate=true, so we don't need to read it, only unlock, after IO.
+ *
+ * TODO: The read_4_write should calc a need_to_read_pages_count, if bigger then
+ * to-be-written count, we should consider the xor-in-place mode.
+ * need_to_read_pages_count is the actual number of pages not present in cache.
+ * maybe "devs_in_group - ios->sp2d[p].write_count" is a good enough
+ * approximation? In this mode the read pages are put in the empty places of
+ * ios->sp2d[p][*], xor is calculated the same way. These pages are
+ * allocated/freed and don't go through cache
+ */
+static int _read_4_write(struct ore_io_state *ios)
+{
+       struct ore_io_state *ios_read;
+       struct ore_striping_info read_si;
+       struct __stripe_pages_2d *sp2d = ios->sp2d;
+       u64 offset = ios->si.first_stripe_start;
+       u64 last_stripe_end;
+       unsigned bytes_in_stripe = ios->si.bytes_in_stripe;
+       unsigned i, c, p, min_p = sp2d->pages_in_unit, max_p = -1;
+       int ret;
+
+       if (offset == ios->offset) /* Go to start collect $200 */
+               goto read_last_stripe;
+
+       min_p = _sp2d_min_pg(sp2d);
+       max_p = _sp2d_max_pg(sp2d);
+
+       for (c = 0; ; c++) {
+               ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
+               read_si.obj_offset += min_p * PAGE_SIZE;
+               offset += min_p * PAGE_SIZE;
+               for (p = min_p; p <= max_p; p++) {
+                       struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
+                       struct page **pp = &_1ps->pages[c];
+                       bool uptodate;
+
+                       if (*pp)
+                               /* to-be-written pages start here */
+                               goto read_last_stripe;
+
+                       *pp = ios->r4w->get_page(ios->private, offset,
+                                                &uptodate);
+                       if (unlikely(!*pp))
+                               return -ENOMEM;
+
+                       if (!uptodate)
+                               _add_to_read_4_write(ios, &read_si, *pp);
+
+                       /* Mark read-pages to be cache_released */
+                       _1ps->page_is_read[c] = true;
+                       read_si.obj_offset += PAGE_SIZE;
+                       offset += PAGE_SIZE;
+               }
+               offset += (sp2d->pages_in_unit - p) * PAGE_SIZE;
+       }
+
+read_last_stripe:
+       offset = ios->offset + (ios->length + PAGE_SIZE - 1) /
+                               PAGE_SIZE * PAGE_SIZE;
+       last_stripe_end = div_u64(offset + bytes_in_stripe - 1, bytes_in_stripe)
+                                * bytes_in_stripe;
+       if (offset == last_stripe_end) /* Optimize for the aligned case */
+               goto read_it;
+
+       ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
+       p = read_si.unit_off / PAGE_SIZE;
+       c = _dev_order(ios->layout->group_width * ios->layout->mirrors_p1,
+                      ios->layout->mirrors_p1, read_si.par_dev, read_si.dev);
+
+       BUG_ON(ios->si.first_stripe_start + bytes_in_stripe != last_stripe_end);
+       /* unaligned IO must be within a single stripe */
+
+       if (min_p == sp2d->pages_in_unit) {
+               /* Didn't do it yet */
+               min_p = _sp2d_min_pg(sp2d);
+               max_p = _sp2d_max_pg(sp2d);
+       }
+
+       while (offset < last_stripe_end) {
+               struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
+
+               if ((min_p <= p) && (p <= max_p)) {
+                       struct page *page;
+                       bool uptodate;
+
+                       BUG_ON(_1ps->pages[c]);
+                       page = ios->r4w->get_page(ios->private, offset,
+                                                 &uptodate);
+                       if (unlikely(!page))
+                               return -ENOMEM;
+
+                       _1ps->pages[c] = page;
+                       /* Mark read-pages to be cache_released */
+                       _1ps->page_is_read[c] = true;
+                       if (!uptodate)
+                               _add_to_read_4_write(ios, &read_si, page);
+               }
+
+               offset += PAGE_SIZE;
+               if (p == (sp2d->pages_in_unit - 1)) {
+                       ++c;
+                       p = 0;
+                       ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
+               } else {
+                       read_si.obj_offset += PAGE_SIZE;
+                       ++p;
+               }
+       }
+
+read_it:
+       ios_read = ios->ios_read_4_write;
+       if (!ios_read)
+               return 0;
+
+       /* FIXME: Ugly to signal _sbi_read_mirror that we have bio(s). Change
+        * to check for per_dev->bio
+        */
+       ios_read->pages = ios->pages;
+
+       /* Now read these devices */
+       for (i = 0; i < ios_read->numdevs; i += ios_read->layout->mirrors_p1) {
+               ret = _ore_read_mirror(ios_read, i);
+               if (unlikely(ret))
+                       return ret;
+       }
+
+       ret = ore_io_execute(ios_read); /* Synchronus execution */
+       if (unlikely(ret)) {
+               ORE_DBGMSG("!! ore_io_execute => %d\n", ret);
+               return ret;
+       }
+
+       _mark_read4write_pages_uptodate(ios_read, ret);
+       return 0;
+}
+
 /* In writes @cur_len means length left. .i.e cur_len==0 is the last parity U */
 int _ore_add_parity_unit(struct ore_io_state *ios,
                            struct ore_striping_info *si,
@@ -86,42 +554,89 @@ int _ore_add_parity_unit(struct ore_io_state *ios,
                BUG_ON(per_dev->cur_sg >= ios->sgs_per_dev);
                _ore_add_sg_seg(per_dev, cur_len, true);
        } else {
+               struct __stripe_pages_2d *sp2d = ios->sp2d;
                struct page **pages = ios->parity_pages + ios->cur_par_page;
-               unsigned num_pages = ios->layout->stripe_unit / PAGE_SIZE;
+               unsigned num_pages;
                unsigned array_start = 0;
                unsigned i;
                int ret;
 
+               si->cur_pg = _sp2d_min_pg(sp2d);
+               num_pages  = _sp2d_max_pg(sp2d) + 1 - si->cur_pg;
+
+               if (!cur_len) /* If last stripe operate on parity comp */
+                       si->cur_comp = sp2d->data_devs;
+
+               if (!per_dev->length) {
+                       per_dev->offset += si->cur_pg * PAGE_SIZE;
+                       /* If first stripe, Read in all read4write pages
+                        * (if needed) before we calculate the first parity.
+                        */
+                       _read_4_write(ios);
+               }
+
                for (i = 0; i < num_pages; i++) {
                        pages[i] = _raid_page_alloc();
                        if (unlikely(!pages[i]))
                                return -ENOMEM;
 
                        ++(ios->cur_par_page);
-                       /* TODO: only read support for now */
-                       clear_highpage(pages[i]);
                }
 
-               ORE_DBGMSG("writing dev=%d num_pages=%d cur_par_page=%d",
-                            per_dev->dev, num_pages, ios->cur_par_page);
+               BUG_ON(si->cur_comp != sp2d->data_devs);
+               BUG_ON(si->cur_pg + num_pages > sp2d->pages_in_unit);
 
                ret = _ore_add_stripe_unit(ios,  &array_start, 0, pages,
                                           per_dev, num_pages * PAGE_SIZE);
                if (unlikely(ret))
                        return ret;
+
+               /* TODO: raid6 if (last_parity_dev) */
+               _gen_xor_unit(sp2d);
+               _sp2d_reset(sp2d, ios->r4w, ios->private);
        }
        return 0;
 }
 
 int _ore_post_alloc_raid_stuff(struct ore_io_state *ios)
 {
-       /*TODO: Only raid writes has stuff to add here */
+       struct ore_layout *layout = ios->layout;
+
+       if (ios->parity_pages) {
+               unsigned pages_in_unit = layout->stripe_unit / PAGE_SIZE;
+               unsigned stripe_size = ios->si.bytes_in_stripe;
+               u64 last_stripe, first_stripe;
+
+               if (_sp2d_alloc(pages_in_unit, layout->group_width,
+                               layout->parity, &ios->sp2d)) {
+                       return -ENOMEM;
+               }
+
+               BUG_ON(ios->offset % PAGE_SIZE);
+
+               /* Round io down to last full strip */
+               first_stripe = div_u64(ios->offset, stripe_size);
+               last_stripe = div_u64(ios->offset + ios->length, stripe_size);
+
+               /* If an IO spans more then a single stripe it must end at
+                * a stripe boundary. The reminder at the end is pushed into the
+                * next IO.
+                */
+               if (last_stripe != first_stripe) {
+                       ios->length = last_stripe * stripe_size - ios->offset;
+
+                       BUG_ON(!ios->length);
+                       ios->nr_pages = (ios->length + PAGE_SIZE - 1) /
+                                       PAGE_SIZE;
+                       ios->si.length = ios->length; /*make it consistent */
+               }
+       }
        return 0;
 }
 
 void _ore_free_raid_stuff(struct ore_io_state *ios)
 {
-       if (ios->parity_pages) { /* writing and raid */
+       if (ios->sp2d) { /* writing and raid */
                unsigned i;
 
                for (i = 0; i < ios->cur_par_page; i++) {
@@ -132,9 +647,14 @@ void _ore_free_raid_stuff(struct ore_io_state *ios)
                }
                if (ios->extra_part_alloc)
                        kfree(ios->parity_pages);
+               /* If IO returned an error pages might need unlocking */
+               _sp2d_reset(ios->sp2d, ios->r4w, ios->private);
+               _sp2d_free(ios->sp2d);
        } else {
                /* Will only be set if raid reading && sglist is big */
                if (ios->extra_part_alloc)
                        kfree(ios->per_dev[0].sglist);
        }
+       if (ios->ios_read_4_write)
+               ore_put_io_state(ios->ios_read_4_write);
 }