IB/ehca: Support large page MRs
authorHoang-Nam Nguyen <hnguyen@de.ibm.com>
Fri, 20 Jul 2007 14:01:51 +0000 (16:01 +0200)
committerRoland Dreier <rolandd@cisco.com>
Sat, 21 Jul 2007 04:19:43 +0000 (21:19 -0700)
Add support for MR pages larger than 4K on eHCA2. This reduces
firmware memory consumption.  If enabled via the mr_largepage module
parameter, the MR page size will be determined based on the MR length
and the hardware capabilities -- if the MR is >= 16M, 16M pages are
used, for example.

Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ehca/ehca_classes.h
drivers/infiniband/hw/ehca/ehca_main.c
drivers/infiniband/hw/ehca/ehca_mrmw.c
drivers/infiniband/hw/ehca/ehca_mrmw.h
drivers/infiniband/hw/ehca/hcp_if.c

index 043e4fb23fb093bb92f800518d38e8bd1f6dbc4b..63b8b9f7c4fcbd17c1851ec74607f39163a06be5 100644 (file)
@@ -100,6 +100,11 @@ struct ehca_sport {
        struct ehca_sma_attr saved_attr;
 };
 
+#define HCA_CAP_MR_PGSIZE_4K  1
+#define HCA_CAP_MR_PGSIZE_64K 2
+#define HCA_CAP_MR_PGSIZE_1M  4
+#define HCA_CAP_MR_PGSIZE_16M 8
+
 struct ehca_shca {
        struct ib_device ib_device;
        struct ibmebus_dev *ibmebus_dev;
@@ -115,6 +120,8 @@ struct ehca_shca {
        struct h_galpas galpas;
        struct mutex modify_mutex;
        u64 hca_cap;
+       /* MR pgsize: bit 0-3 means 4K, 64K, 1M, 16M respectively */
+       u32 hca_cap_mr_pgsize;
        int max_mtu;
 };
 
@@ -206,6 +213,7 @@ struct ehca_mr {
        enum ehca_mr_flag flags;
        u32 num_kpages;         /* number of kernel pages */
        u32 num_hwpages;        /* number of hw pages to form MR */
+       u64 hwpage_size;        /* hw page size used for this MR */
        int acl;                /* ACL (stored here for usage in reregister) */
        u64 *start;             /* virtual start address (stored here for */
                                /* usage in reregister) */
@@ -240,6 +248,7 @@ struct ehca_mr_pginfo {
        enum ehca_mr_pgi_type type;
        u64 num_kpages;
        u64 kpage_cnt;
+       u64 hwpage_size;     /* hw page size used for this MR */
        u64 num_hwpages;     /* number of hw pages */
        u64 hwpage_cnt;      /* counter for hw pages */
        u64 next_hwpage;     /* next hw page in buffer/chunk/listelem */
index 04c324330b7cca4519d6618951893630ace4a17e..ec0145192228fdcb56de58a865f60af90b897fd7 100644 (file)
@@ -63,6 +63,7 @@ int ehca_port_act_time = 30;
 int ehca_poll_all_eqs  = 1;
 int ehca_static_rate   = -1;
 int ehca_scaling_code  = 0;
+int ehca_mr_largepage  = 0;
 
 module_param_named(open_aqp1,     ehca_open_aqp1,     int, 0);
 module_param_named(debug_level,   ehca_debug_level,   int, 0);
@@ -72,7 +73,8 @@ module_param_named(use_hp_mr,     ehca_use_hp_mr,     int, 0);
 module_param_named(port_act_time, ehca_port_act_time, int, 0);
 module_param_named(poll_all_eqs,  ehca_poll_all_eqs,  int, 0);
 module_param_named(static_rate,   ehca_static_rate,   int, 0);
-module_param_named(scaling_code,   ehca_scaling_code,   int, 0);
+module_param_named(scaling_code,  ehca_scaling_code,  int, 0);
+module_param_named(mr_largepage,  ehca_mr_largepage,  int, 0);
 
 MODULE_PARM_DESC(open_aqp1,
                 "AQP1 on startup (0: no (default), 1: yes)");
@@ -95,6 +97,9 @@ MODULE_PARM_DESC(static_rate,
                 "set permanent static rate (default: disabled)");
 MODULE_PARM_DESC(scaling_code,
                 "set scaling code (0: disabled/default, 1: enabled)");
+MODULE_PARM_DESC(mr_largepage,
+                "use large page for MR (0: use PAGE_SIZE (default), "
+                "1: use large page depending on MR size");
 
 DEFINE_RWLOCK(ehca_qp_idr_lock);
 DEFINE_RWLOCK(ehca_cq_idr_lock);
@@ -295,6 +300,8 @@ int ehca_sense_attributes(struct ehca_shca *shca)
                if (EHCA_BMASK_GET(hca_cap_descr[i].mask, shca->hca_cap))
                        ehca_gen_dbg("   %s", hca_cap_descr[i].descr);
 
+       shca->hca_cap_mr_pgsize = rblock->memory_page_size_supported;
+
        port = (struct hipz_query_port *)rblock;
        h_ret = hipz_h_query_port(shca->ipz_hca_handle, 1, port);
        if (h_ret != H_SUCCESS) {
@@ -590,6 +597,14 @@ static ssize_t ehca_show_adapter_handle(struct device *dev,
 }
 static DEVICE_ATTR(adapter_handle, S_IRUGO, ehca_show_adapter_handle, NULL);
 
+static ssize_t ehca_show_mr_largepage(struct device *dev,
+                                     struct device_attribute *attr,
+                                     char *buf)
+{
+       return sprintf(buf, "%d\n", ehca_mr_largepage);
+}
+static DEVICE_ATTR(mr_largepage, S_IRUGO, ehca_show_mr_largepage, NULL);
+
 static struct attribute *ehca_dev_attrs[] = {
        &dev_attr_adapter_handle.attr,
        &dev_attr_num_ports.attr,
@@ -606,6 +621,7 @@ static struct attribute *ehca_dev_attrs[] = {
        &dev_attr_cur_mw.attr,
        &dev_attr_max_pd.attr,
        &dev_attr_max_ah.attr,
+       &dev_attr_mr_largepage.attr,
        NULL
 };
 
index 9f4c9d46e8ef89b1df9a6a405c3ba011125d1834..c1b868b79d677bf6a8caf0d2ff0532d860082e71 100644 (file)
@@ -5,6 +5,7 @@
  *
  *  Authors: Dietmar Decker <ddecker@de.ibm.com>
  *           Christoph Raisch <raisch@de.ibm.com>
+ *           Hoang-Nam Nguyen <hnguyen@de.ibm.com>
  *
  *  Copyright (c) 2005 IBM Corporation
  *
 static struct kmem_cache *mr_cache;
 static struct kmem_cache *mw_cache;
 
+enum ehca_mr_pgsize {
+       EHCA_MR_PGSIZE4K  = 0x1000L,
+       EHCA_MR_PGSIZE64K = 0x10000L,
+       EHCA_MR_PGSIZE1M  = 0x100000L,
+       EHCA_MR_PGSIZE16M = 0x1000000L
+};
+
+extern int ehca_mr_largepage;
+
+static u32 ehca_encode_hwpage_size(u32 pgsize)
+{
+       u32 idx = 0;
+       pgsize >>= 12;
+       /*
+        * map mr page size into hw code:
+        * 0, 1, 2, 3 for 4K, 64K, 1M, 64M
+        */
+       while (!(pgsize & 1)) {
+               idx++;
+               pgsize >>= 4;
+       }
+       return idx;
+}
+
+static u64 ehca_get_max_hwpage_size(struct ehca_shca *shca)
+{
+       if (shca->hca_cap_mr_pgsize & HCA_CAP_MR_PGSIZE_16M)
+               return EHCA_MR_PGSIZE16M;
+       return EHCA_MR_PGSIZE4K;
+}
+
 static struct ehca_mr *ehca_mr_new(void)
 {
        struct ehca_mr *me;
@@ -207,19 +239,23 @@ struct ib_mr *ehca_reg_phys_mr(struct ib_pd *pd,
                struct ehca_mr_pginfo pginfo;
                u32 num_kpages;
                u32 num_hwpages;
+               u64 hw_pgsize;
 
                num_kpages = NUM_CHUNKS(((u64)iova_start % PAGE_SIZE) + size,
                                        PAGE_SIZE);
-               num_hwpages = NUM_CHUNKS(((u64)iova_start % EHCA_PAGESIZE) +
-                                        size, EHCA_PAGESIZE);
+               /* for kernel space we try most possible pgsize */
+               hw_pgsize = ehca_get_max_hwpage_size(shca);
+               num_hwpages = NUM_CHUNKS(((u64)iova_start % hw_pgsize) + size,
+                                        hw_pgsize);
                memset(&pginfo, 0, sizeof(pginfo));
                pginfo.type = EHCA_MR_PGI_PHYS;
                pginfo.num_kpages = num_kpages;
+               pginfo.hwpage_size = hw_pgsize;
                pginfo.num_hwpages = num_hwpages;
                pginfo.u.phy.num_phys_buf = num_phys_buf;
                pginfo.u.phy.phys_buf_array = phys_buf_array;
-               pginfo.next_hwpage = (((u64)iova_start & ~PAGE_MASK) /
-                                     EHCA_PAGESIZE);
+               pginfo.next_hwpage =
+                       ((u64)iova_start & ~(hw_pgsize - 1)) / hw_pgsize;
 
                ret = ehca_reg_mr(shca, e_mr, iova_start, size, mr_access_flags,
                                  e_pd, &pginfo, &e_mr->ib.ib_mr.lkey,
@@ -259,6 +295,7 @@ struct ib_mr *ehca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
        int ret;
        u32 num_kpages;
        u32 num_hwpages;
+       u64 hwpage_size;
 
        if (!pd) {
                ehca_gen_err("bad pd=%p", pd);
@@ -309,16 +346,32 @@ struct ib_mr *ehca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 
        /* determine number of MR pages */
        num_kpages = NUM_CHUNKS((virt % PAGE_SIZE) + length, PAGE_SIZE);
-       num_hwpages = NUM_CHUNKS((virt % EHCA_PAGESIZE) + length,
-                                EHCA_PAGESIZE);
+       /* select proper hw_pgsize */
+       if (ehca_mr_largepage &&
+           (shca->hca_cap_mr_pgsize & HCA_CAP_MR_PGSIZE_16M)) {
+               if (length <= EHCA_MR_PGSIZE4K
+                   && PAGE_SIZE == EHCA_MR_PGSIZE4K)
+                       hwpage_size = EHCA_MR_PGSIZE4K;
+               else if (length <= EHCA_MR_PGSIZE64K)
+                       hwpage_size = EHCA_MR_PGSIZE64K;
+               else if (length <= EHCA_MR_PGSIZE1M)
+                       hwpage_size = EHCA_MR_PGSIZE1M;
+               else
+                       hwpage_size = EHCA_MR_PGSIZE16M;
+       } else
+               hwpage_size = EHCA_MR_PGSIZE4K;
+       ehca_dbg(pd->device, "hwpage_size=%lx", hwpage_size);
 
+reg_user_mr_fallback:
+       num_hwpages = NUM_CHUNKS((virt % hwpage_size) + length, hwpage_size);
        /* register MR on HCA */
        memset(&pginfo, 0, sizeof(pginfo));
        pginfo.type = EHCA_MR_PGI_USER;
+       pginfo.hwpage_size = hwpage_size;
        pginfo.num_kpages = num_kpages;
        pginfo.num_hwpages = num_hwpages;
        pginfo.u.usr.region = e_mr->umem;
-       pginfo.next_hwpage = e_mr->umem->offset / EHCA_PAGESIZE;
+       pginfo.next_hwpage = e_mr->umem->offset / hwpage_size;
        pginfo.u.usr.next_chunk = list_prepare_entry(pginfo.u.usr.next_chunk,
                                                     (&e_mr->umem->chunk_list),
                                                     list);
@@ -326,6 +379,18 @@ struct ib_mr *ehca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
        ret = ehca_reg_mr(shca, e_mr, (u64 *)virt, length, mr_access_flags,
                          e_pd, &pginfo, &e_mr->ib.ib_mr.lkey,
                          &e_mr->ib.ib_mr.rkey);
+       if (ret == -EINVAL && pginfo.hwpage_size > PAGE_SIZE) {
+               ehca_warn(pd->device, "failed to register mr "
+                         "with hwpage_size=%lx", hwpage_size);
+               ehca_info(pd->device, "try to register mr with "
+                         "kpage_size=%lx", PAGE_SIZE);
+               /*
+                * this means kpages are not contiguous for a hw page
+                * try kernel page size as fallback solution
+                */
+               hwpage_size = PAGE_SIZE;
+               goto reg_user_mr_fallback;
+       }
        if (ret) {
                ib_mr = ERR_PTR(ret);
                goto reg_user_mr_exit2;
@@ -452,6 +517,8 @@ int ehca_rereg_phys_mr(struct ib_mr *mr,
        new_pd = container_of(mr->pd, struct ehca_pd, ib_pd);
 
        if (mr_rereg_mask & IB_MR_REREG_TRANS) {
+               u64 hw_pgsize = ehca_get_max_hwpage_size(shca);
+
                new_start = iova_start; /* change address */
                /* check physical buffer list and calculate size */
                ret = ehca_mr_chk_buf_and_calc_size(phys_buf_array,
@@ -468,16 +535,17 @@ int ehca_rereg_phys_mr(struct ib_mr *mr,
                }
                num_kpages = NUM_CHUNKS(((u64)new_start % PAGE_SIZE) +
                                        new_size, PAGE_SIZE);
-               num_hwpages = NUM_CHUNKS(((u64)new_start % EHCA_PAGESIZE) +
-                                        new_size, EHCA_PAGESIZE);
+               num_hwpages = NUM_CHUNKS(((u64)new_start % hw_pgsize) +
+                                        new_size, hw_pgsize);
                memset(&pginfo, 0, sizeof(pginfo));
                pginfo.type = EHCA_MR_PGI_PHYS;
                pginfo.num_kpages = num_kpages;
+               pginfo.hwpage_size = hw_pgsize;
                pginfo.num_hwpages = num_hwpages;
                pginfo.u.phy.num_phys_buf = num_phys_buf;
                pginfo.u.phy.phys_buf_array = phys_buf_array;
-               pginfo.next_hwpage = (((u64)iova_start & ~PAGE_MASK) /
-                                     EHCA_PAGESIZE);
+               pginfo.next_hwpage =
+                       ((u64)iova_start & ~(hw_pgsize - 1)) / hw_pgsize;
        }
        if (mr_rereg_mask & IB_MR_REREG_ACCESS)
                new_acl = mr_access_flags;
@@ -709,6 +777,7 @@ struct ib_fmr *ehca_alloc_fmr(struct ib_pd *pd,
        int ret;
        u32 tmp_lkey, tmp_rkey;
        struct ehca_mr_pginfo pginfo;
+       u64 hw_pgsize;
 
        /* check other parameters */
        if (((mr_access_flags & IB_ACCESS_REMOTE_WRITE) &&
@@ -738,8 +807,8 @@ struct ib_fmr *ehca_alloc_fmr(struct ib_pd *pd,
                ib_fmr = ERR_PTR(-EINVAL);
                goto alloc_fmr_exit0;
        }
-       if (((1 << fmr_attr->page_shift) != EHCA_PAGESIZE) &&
-           ((1 << fmr_attr->page_shift) != PAGE_SIZE)) {
+       hw_pgsize = ehca_get_max_hwpage_size(shca);
+       if ((1 << fmr_attr->page_shift) != hw_pgsize) {
                ehca_err(pd->device, "unsupported fmr_attr->page_shift=%x",
                         fmr_attr->page_shift);
                ib_fmr = ERR_PTR(-EINVAL);
@@ -755,6 +824,10 @@ struct ib_fmr *ehca_alloc_fmr(struct ib_pd *pd,
 
        /* register MR on HCA */
        memset(&pginfo, 0, sizeof(pginfo));
+       /*
+        * pginfo.num_hwpages==0, ie register_rpages() will not be called
+        * but deferred to map_phys_fmr()
+        */
        ret = ehca_reg_mr(shca, e_fmr, NULL,
                          fmr_attr->max_pages * (1 << fmr_attr->page_shift),
                          mr_access_flags, e_pd, &pginfo,
@@ -765,6 +838,7 @@ struct ib_fmr *ehca_alloc_fmr(struct ib_pd *pd,
        }
 
        /* successful */
+       e_fmr->hwpage_size = hw_pgsize;
        e_fmr->fmr_page_size = 1 << fmr_attr->page_shift;
        e_fmr->fmr_max_pages = fmr_attr->max_pages;
        e_fmr->fmr_max_maps = fmr_attr->max_maps;
@@ -822,10 +896,12 @@ int ehca_map_phys_fmr(struct ib_fmr *fmr,
        memset(&pginfo, 0, sizeof(pginfo));
        pginfo.type = EHCA_MR_PGI_FMR;
        pginfo.num_kpages = list_len;
-       pginfo.num_hwpages = list_len * (e_fmr->fmr_page_size / EHCA_PAGESIZE);
+       pginfo.hwpage_size = e_fmr->hwpage_size;
+       pginfo.num_hwpages =
+               list_len * e_fmr->fmr_page_size / pginfo.hwpage_size;
        pginfo.u.fmr.page_list = page_list;
-       pginfo.next_hwpage = ((iova & (e_fmr->fmr_page_size-1)) /
-                             EHCA_PAGESIZE);
+       pginfo.next_hwpage =
+               (iova & (e_fmr->fmr_page_size-1)) / pginfo.hwpage_size;
        pginfo.u.fmr.fmr_pgsize = e_fmr->fmr_page_size;
 
        ret = ehca_rereg_mr(shca, e_fmr, (u64 *)iova,
@@ -964,7 +1040,7 @@ int ehca_reg_mr(struct ehca_shca *shca,
        struct ehca_mr_hipzout_parms hipzout;
 
        ehca_mrmw_map_acl(acl, &hipz_acl);
-       ehca_mrmw_set_pgsize_hipz_acl(&hipz_acl);
+       ehca_mrmw_set_pgsize_hipz_acl(pginfo->hwpage_size, &hipz_acl);
        if (ehca_use_hp_mr == 1)
                hipz_acl |= 0x00000001;
 
@@ -987,6 +1063,7 @@ int ehca_reg_mr(struct ehca_shca *shca,
        /* successful registration */
        e_mr->num_kpages = pginfo->num_kpages;
        e_mr->num_hwpages = pginfo->num_hwpages;
+       e_mr->hwpage_size = pginfo->hwpage_size;
        e_mr->start = iova_start;
        e_mr->size = size;
        e_mr->acl = acl;
@@ -1029,6 +1106,9 @@ int ehca_reg_mr_rpages(struct ehca_shca *shca,
        u32 i;
        u64 *kpage;
 
+       if (!pginfo->num_hwpages) /* in case of fmr */
+               return 0;
+
        kpage = ehca_alloc_fw_ctrlblock(GFP_KERNEL);
        if (!kpage) {
                ehca_err(&shca->ib_device, "kpage alloc failed");
@@ -1036,7 +1116,7 @@ int ehca_reg_mr_rpages(struct ehca_shca *shca,
                goto ehca_reg_mr_rpages_exit0;
        }
 
-       /* max 512 pages per shot */
+       /* max MAX_RPAGES ehca mr pages per register call */
        for (i = 0; i < NUM_CHUNKS(pginfo->num_hwpages, MAX_RPAGES); i++) {
 
                if (i == NUM_CHUNKS(pginfo->num_hwpages, MAX_RPAGES) - 1) {
@@ -1049,8 +1129,8 @@ int ehca_reg_mr_rpages(struct ehca_shca *shca,
                ret = ehca_set_pagebuf(pginfo, rnum, kpage);
                if (ret) {
                        ehca_err(&shca->ib_device, "ehca_set_pagebuf "
-                                        "bad rc, ret=%x rnum=%x kpage=%p",
-                                        ret, rnum, kpage);
+                                "bad rc, ret=%x rnum=%x kpage=%p",
+                                ret, rnum, kpage);
                        goto ehca_reg_mr_rpages_exit1;
                }
 
@@ -1065,9 +1145,10 @@ int ehca_reg_mr_rpages(struct ehca_shca *shca,
                } else
                        rpage = *kpage;
 
-               h_ret = hipz_h_register_rpage_mr(shca->ipz_hca_handle, e_mr,
-                                                0, /* pagesize 4k */
-                                                0, rpage, rnum);
+               h_ret = hipz_h_register_rpage_mr(
+                       shca->ipz_hca_handle, e_mr,
+                       ehca_encode_hwpage_size(pginfo->hwpage_size),
+                       0, rpage, rnum);
 
                if (i == NUM_CHUNKS(pginfo->num_hwpages, MAX_RPAGES) - 1) {
                        /*
@@ -1131,7 +1212,7 @@ inline int ehca_rereg_mr_rereg1(struct ehca_shca *shca,
        struct ehca_mr_hipzout_parms hipzout;
 
        ehca_mrmw_map_acl(acl, &hipz_acl);
-       ehca_mrmw_set_pgsize_hipz_acl(&hipz_acl);
+       ehca_mrmw_set_pgsize_hipz_acl(pginfo->hwpage_size, &hipz_acl);
 
        kpage = ehca_alloc_fw_ctrlblock(GFP_KERNEL);
        if (!kpage) {
@@ -1182,6 +1263,7 @@ inline int ehca_rereg_mr_rereg1(struct ehca_shca *shca,
                 */
                e_mr->num_kpages = pginfo->num_kpages;
                e_mr->num_hwpages = pginfo->num_hwpages;
+               e_mr->hwpage_size = pginfo->hwpage_size;
                e_mr->start = iova_start;
                e_mr->size = size;
                e_mr->acl = acl;
@@ -1268,13 +1350,14 @@ int ehca_rereg_mr(struct ehca_shca *shca,
 
                /* set some MR values */
                e_mr->flags = save_mr.flags;
+               e_mr->hwpage_size = save_mr.hwpage_size;
                e_mr->fmr_page_size = save_mr.fmr_page_size;
                e_mr->fmr_max_pages = save_mr.fmr_max_pages;
                e_mr->fmr_max_maps = save_mr.fmr_max_maps;
                e_mr->fmr_map_cnt = save_mr.fmr_map_cnt;
 
                ret = ehca_reg_mr(shca, e_mr, iova_start, size, acl,
-                                     e_pd, pginfo, lkey, rkey);
+                                 e_pd, pginfo, lkey, rkey);
                if (ret) {
                        u32 offset = (u64)(&e_mr->flags) - (u64)e_mr;
                        memcpy(&e_mr->flags, &(save_mr.flags),
@@ -1355,6 +1438,7 @@ int ehca_unmap_one_fmr(struct ehca_shca *shca,
 
        /* set some MR values */
        e_fmr->flags = save_fmr.flags;
+       e_fmr->hwpage_size = save_fmr.hwpage_size;
        e_fmr->fmr_page_size = save_fmr.fmr_page_size;
        e_fmr->fmr_max_pages = save_fmr.fmr_max_pages;
        e_fmr->fmr_max_maps = save_fmr.fmr_max_maps;
@@ -1363,8 +1447,6 @@ int ehca_unmap_one_fmr(struct ehca_shca *shca,
 
        memset(&pginfo, 0, sizeof(pginfo));
        pginfo.type = EHCA_MR_PGI_FMR;
-       pginfo.num_kpages = 0;
-       pginfo.num_hwpages = 0;
        ret = ehca_reg_mr(shca, e_fmr, NULL,
                          (e_fmr->fmr_max_pages * e_fmr->fmr_page_size),
                          e_fmr->acl, e_pd, &pginfo, &tmp_lkey,
@@ -1373,7 +1455,6 @@ int ehca_unmap_one_fmr(struct ehca_shca *shca,
                u32 offset = (u64)(&e_fmr->flags) - (u64)e_fmr;
                memcpy(&e_fmr->flags, &(save_mr.flags),
                       sizeof(struct ehca_mr) - offset);
-               goto ehca_unmap_one_fmr_exit0;
        }
 
 ehca_unmap_one_fmr_exit0:
@@ -1401,7 +1482,7 @@ int ehca_reg_smr(struct ehca_shca *shca,
        struct ehca_mr_hipzout_parms hipzout;
 
        ehca_mrmw_map_acl(acl, &hipz_acl);
-       ehca_mrmw_set_pgsize_hipz_acl(&hipz_acl);
+       ehca_mrmw_set_pgsize_hipz_acl(e_origmr->hwpage_size, &hipz_acl);
 
        h_ret = hipz_h_register_smr(shca->ipz_hca_handle, e_newmr, e_origmr,
                                    (u64)iova_start, hipz_acl, e_pd->fw_pd,
@@ -1420,6 +1501,7 @@ int ehca_reg_smr(struct ehca_shca *shca,
        /* successful registration */
        e_newmr->num_kpages = e_origmr->num_kpages;
        e_newmr->num_hwpages = e_origmr->num_hwpages;
+       e_newmr->hwpage_size   = e_origmr->hwpage_size;
        e_newmr->start = iova_start;
        e_newmr->size = e_origmr->size;
        e_newmr->acl = acl;
@@ -1452,6 +1534,7 @@ int ehca_reg_internal_maxmr(
        struct ib_phys_buf ib_pbuf;
        u32 num_kpages;
        u32 num_hwpages;
+       u64 hw_pgsize;
 
        e_mr = ehca_mr_new();
        if (!e_mr) {
@@ -1468,13 +1551,15 @@ int ehca_reg_internal_maxmr(
        ib_pbuf.size = size_maxmr;
        num_kpages = NUM_CHUNKS(((u64)iova_start % PAGE_SIZE) + size_maxmr,
                                PAGE_SIZE);
-       num_hwpages = NUM_CHUNKS(((u64)iova_start % EHCA_PAGESIZE) + size_maxmr,
-                                EHCA_PAGESIZE);
+       hw_pgsize = ehca_get_max_hwpage_size(shca);
+       num_hwpages = NUM_CHUNKS(((u64)iova_start % hw_pgsize) + size_maxmr,
+                                hw_pgsize);
 
        memset(&pginfo, 0, sizeof(pginfo));
        pginfo.type = EHCA_MR_PGI_PHYS;
        pginfo.num_kpages = num_kpages;
        pginfo.num_hwpages = num_hwpages;
+       pginfo.hwpage_size = hw_pgsize;
        pginfo.u.phy.num_phys_buf = 1;
        pginfo.u.phy.phys_buf_array = &ib_pbuf;
 
@@ -1523,7 +1608,7 @@ int ehca_reg_maxmr(struct ehca_shca *shca,
        struct ehca_mr_hipzout_parms hipzout;
 
        ehca_mrmw_map_acl(acl, &hipz_acl);
-       ehca_mrmw_set_pgsize_hipz_acl(&hipz_acl);
+       ehca_mrmw_set_pgsize_hipz_acl(e_origmr->hwpage_size, &hipz_acl);
 
        h_ret = hipz_h_register_smr(shca->ipz_hca_handle, e_newmr, e_origmr,
                                    (u64)iova_start, hipz_acl, e_pd->fw_pd,
@@ -1539,6 +1624,7 @@ int ehca_reg_maxmr(struct ehca_shca *shca,
        /* successful registration */
        e_newmr->num_kpages = e_origmr->num_kpages;
        e_newmr->num_hwpages = e_origmr->num_hwpages;
+       e_newmr->hwpage_size = e_origmr->hwpage_size;
        e_newmr->start = iova_start;
        e_newmr->size = e_origmr->size;
        e_newmr->acl = acl;
@@ -1684,6 +1770,7 @@ static int ehca_set_pagebuf_user1(struct ehca_mr_pginfo *pginfo,
        u64 pgaddr;
        u32 i = 0;
        u32 j = 0;
+       int hwpages_per_kpage = PAGE_SIZE / pginfo->hwpage_size;
 
        /* loop over desired chunk entries */
        chunk      = pginfo->u.usr.next_chunk;
@@ -1695,7 +1782,7 @@ static int ehca_set_pagebuf_user1(struct ehca_mr_pginfo *pginfo,
                                << PAGE_SHIFT ;
                        *kpage = phys_to_abs(pgaddr +
                                             (pginfo->next_hwpage *
-                                             EHCA_PAGESIZE));
+                                             pginfo->hwpage_size));
                        if ( !(*kpage) ) {
                                ehca_gen_err("pgaddr=%lx "
                                             "chunk->page_list[i]=%lx "
@@ -1708,8 +1795,7 @@ static int ehca_set_pagebuf_user1(struct ehca_mr_pginfo *pginfo,
                        (pginfo->hwpage_cnt)++;
                        (pginfo->next_hwpage)++;
                        kpage++;
-                       if (pginfo->next_hwpage %
-                           (PAGE_SIZE / EHCA_PAGESIZE) == 0) {
+                       if (pginfo->next_hwpage % hwpages_per_kpage == 0) {
                                (pginfo->kpage_cnt)++;
                                (pginfo->u.usr.next_nmap)++;
                                pginfo->next_hwpage = 0;
@@ -1738,6 +1824,143 @@ static int ehca_set_pagebuf_user1(struct ehca_mr_pginfo *pginfo,
        return ret;
 }
 
+/*
+ * check given pages for contiguous layout
+ * last page addr is returned in prev_pgaddr for further check
+ */
+static int ehca_check_kpages_per_ate(struct scatterlist *page_list,
+                                    int start_idx, int end_idx,
+                                    u64 *prev_pgaddr)
+{
+       int t;
+       for (t = start_idx; t <= end_idx; t++) {
+               u64 pgaddr = page_to_pfn(page_list[t].page) << PAGE_SHIFT;
+               ehca_gen_dbg("chunk_page=%lx value=%016lx", pgaddr,
+                            *(u64 *)abs_to_virt(phys_to_abs(pgaddr)));
+               if (pgaddr - PAGE_SIZE != *prev_pgaddr) {
+                       ehca_gen_err("uncontiguous page found pgaddr=%lx "
+                                    "prev_pgaddr=%lx page_list_i=%x",
+                                    pgaddr, *prev_pgaddr, t);
+                       return -EINVAL;
+               }
+               *prev_pgaddr = pgaddr;
+       }
+       return 0;
+}
+
+/* PAGE_SIZE < pginfo->hwpage_size */
+static int ehca_set_pagebuf_user2(struct ehca_mr_pginfo *pginfo,
+                                 u32 number,
+                                 u64 *kpage)
+{
+       int ret = 0;
+       struct ib_umem_chunk *prev_chunk;
+       struct ib_umem_chunk *chunk;
+       u64 pgaddr, prev_pgaddr;
+       u32 i = 0;
+       u32 j = 0;
+       int kpages_per_hwpage = pginfo->hwpage_size / PAGE_SIZE;
+       int nr_kpages = kpages_per_hwpage;
+
+       /* loop over desired chunk entries */
+       chunk      = pginfo->u.usr.next_chunk;
+       prev_chunk = pginfo->u.usr.next_chunk;
+       list_for_each_entry_continue(
+               chunk, (&(pginfo->u.usr.region->chunk_list)), list) {
+               for (i = pginfo->u.usr.next_nmap; i < chunk->nmap; ) {
+                       if (nr_kpages == kpages_per_hwpage) {
+                               pgaddr = ( page_to_pfn(chunk->page_list[i].page)
+                                          << PAGE_SHIFT );
+                               *kpage = phys_to_abs(pgaddr);
+                               if ( !(*kpage) ) {
+                                       ehca_gen_err("pgaddr=%lx i=%x",
+                                                    pgaddr, i);
+                                       ret = -EFAULT;
+                                       return ret;
+                               }
+                               /*
+                                * The first page in a hwpage must be aligned;
+                                * the first MR page is exempt from this rule.
+                                */
+                               if (pgaddr & (pginfo->hwpage_size - 1)) {
+                                       if (pginfo->hwpage_cnt) {
+                                               ehca_gen_err(
+                                                       "invalid alignment "
+                                                       "pgaddr=%lx i=%x "
+                                                       "mr_pgsize=%lx",
+                                                       pgaddr, i,
+                                                       pginfo->hwpage_size);
+                                               ret = -EFAULT;
+                                               return ret;
+                                       }
+                                       /* first MR page */
+                                       pginfo->kpage_cnt =
+                                               (pgaddr &
+                                                (pginfo->hwpage_size - 1)) >>
+                                               PAGE_SHIFT;
+                                       nr_kpages -= pginfo->kpage_cnt;
+                                       *kpage = phys_to_abs(
+                                               pgaddr &
+                                               ~(pginfo->hwpage_size - 1));
+                               }
+                               ehca_gen_dbg("kpage=%lx chunk_page=%lx "
+                                            "value=%016lx", *kpage, pgaddr,
+                                            *(u64 *)abs_to_virt(
+                                                    phys_to_abs(pgaddr)));
+                               prev_pgaddr = pgaddr;
+                               i++;
+                               pginfo->kpage_cnt++;
+                               pginfo->u.usr.next_nmap++;
+                               nr_kpages--;
+                               if (!nr_kpages)
+                                       goto next_kpage;
+                               continue;
+                       }
+                       if (i + nr_kpages > chunk->nmap) {
+                               ret = ehca_check_kpages_per_ate(
+                                       chunk->page_list, i,
+                                       chunk->nmap - 1, &prev_pgaddr);
+                               if (ret) return ret;
+                               pginfo->kpage_cnt += chunk->nmap - i;
+                               pginfo->u.usr.next_nmap += chunk->nmap - i;
+                               nr_kpages -= chunk->nmap - i;
+                               break;
+                       }
+
+                       ret = ehca_check_kpages_per_ate(chunk->page_list, i,
+                                                       i + nr_kpages - 1,
+                                                       &prev_pgaddr);
+                       if (ret) return ret;
+                       i += nr_kpages;
+                       pginfo->kpage_cnt += nr_kpages;
+                       pginfo->u.usr.next_nmap += nr_kpages;
+next_kpage:
+                       nr_kpages = kpages_per_hwpage;
+                       (pginfo->hwpage_cnt)++;
+                       kpage++;
+                       j++;
+                       if (j >= number) break;
+               }
+               if ((pginfo->u.usr.next_nmap >= chunk->nmap) &&
+                   (j >= number)) {
+                       pginfo->u.usr.next_nmap = 0;
+                       prev_chunk = chunk;
+                       break;
+               } else if (pginfo->u.usr.next_nmap >= chunk->nmap) {
+                       pginfo->u.usr.next_nmap = 0;
+                       prev_chunk = chunk;
+               } else if (j >= number)
+                       break;
+               else
+                       prev_chunk = chunk;
+       }
+       pginfo->u.usr.next_chunk =
+               list_prepare_entry(prev_chunk,
+                                  (&(pginfo->u.usr.region->chunk_list)),
+                                  list);
+       return ret;
+}
+
 int ehca_set_pagebuf_phys(struct ehca_mr_pginfo *pginfo,
                          u32 number,
                          u64 *kpage)
@@ -1750,9 +1973,10 @@ int ehca_set_pagebuf_phys(struct ehca_mr_pginfo *pginfo,
        /* loop over desired phys_buf_array entries */
        while (i < number) {
                pbuf   = pginfo->u.phy.phys_buf_array + pginfo->u.phy.next_buf;
-               num_hw  = NUM_CHUNKS((pbuf->addr % EHCA_PAGESIZE) +
-                                    pbuf->size, EHCA_PAGESIZE);
-               offs_hw = (pbuf->addr & ~PAGE_MASK) / EHCA_PAGESIZE;
+               num_hw  = NUM_CHUNKS((pbuf->addr % pginfo->hwpage_size) +
+                                    pbuf->size, pginfo->hwpage_size);
+               offs_hw = (pbuf->addr & ~(pginfo->hwpage_size - 1)) /
+                       pginfo->hwpage_size;
                while (pginfo->next_hwpage < offs_hw + num_hw) {
                        /* sanity check */
                        if ((pginfo->kpage_cnt >= pginfo->num_kpages) ||
@@ -1768,21 +1992,23 @@ int ehca_set_pagebuf_phys(struct ehca_mr_pginfo *pginfo,
                                return -EFAULT;
                        }
                        *kpage = phys_to_abs(
-                               (pbuf->addr & EHCA_PAGEMASK)
-                               + (pginfo->next_hwpage * EHCA_PAGESIZE));
+                               (pbuf->addr & ~(pginfo->hwpage_size - 1)) +
+                               (pginfo->next_hwpage * pginfo->hwpage_size));
                        if ( !(*kpage) && pbuf->addr ) {
-                               ehca_gen_err("pbuf->addr=%lx "
-                                            "pbuf->size=%lx "
+                               ehca_gen_err("pbuf->addr=%lx pbuf->size=%lx "
                                             "next_hwpage=%lx", pbuf->addr,
-                                            pbuf->size,
-                                            pginfo->next_hwpage);
+                                            pbuf->size, pginfo->next_hwpage);
                                return -EFAULT;
                        }
                        (pginfo->hwpage_cnt)++;
                        (pginfo->next_hwpage)++;
-                       if (pginfo->next_hwpage %
-                           (PAGE_SIZE / EHCA_PAGESIZE) == 0)
-                               (pginfo->kpage_cnt)++;
+                       if (PAGE_SIZE >= pginfo->hwpage_size) {
+                               if (pginfo->next_hwpage %
+                                   (PAGE_SIZE / pginfo->hwpage_size) == 0)
+                                       (pginfo->kpage_cnt)++;
+                       } else
+                               pginfo->kpage_cnt += pginfo->hwpage_size /
+                                       PAGE_SIZE;
                        kpage++;
                        i++;
                        if (i >= number) break;
@@ -1806,8 +2032,8 @@ int ehca_set_pagebuf_fmr(struct ehca_mr_pginfo *pginfo,
        /* loop over desired page_list entries */
        fmrlist = pginfo->u.fmr.page_list + pginfo->u.fmr.next_listelem;
        for (i = 0; i < number; i++) {
-               *kpage = phys_to_abs((*fmrlist & EHCA_PAGEMASK) +
-                                    pginfo->next_hwpage * EHCA_PAGESIZE);
+               *kpage = phys_to_abs((*fmrlist & ~(pginfo->hwpage_size - 1)) +
+                                    pginfo->next_hwpage * pginfo->hwpage_size);
                if ( !(*kpage) ) {
                        ehca_gen_err("*fmrlist=%lx fmrlist=%p "
                                     "next_listelem=%lx next_hwpage=%lx",
@@ -1817,15 +2043,38 @@ int ehca_set_pagebuf_fmr(struct ehca_mr_pginfo *pginfo,
                        return -EFAULT;
                }
                (pginfo->hwpage_cnt)++;
-               (pginfo->next_hwpage)++;
-               kpage++;
-               if (pginfo->next_hwpage %
-                   (pginfo->u.fmr.fmr_pgsize / EHCA_PAGESIZE) == 0) {
-                       (pginfo->kpage_cnt)++;
-                       (pginfo->u.fmr.next_listelem)++;
-                       fmrlist++;
-                       pginfo->next_hwpage = 0;
+               if (pginfo->u.fmr.fmr_pgsize >= pginfo->hwpage_size) {
+                       if (pginfo->next_hwpage %
+                           (pginfo->u.fmr.fmr_pgsize /
+                            pginfo->hwpage_size) == 0) {
+                               (pginfo->kpage_cnt)++;
+                               (pginfo->u.fmr.next_listelem)++;
+                               fmrlist++;
+                               pginfo->next_hwpage = 0;
+                       } else
+                               (pginfo->next_hwpage)++;
+               } else {
+                       unsigned int cnt_per_hwpage = pginfo->hwpage_size /
+                               pginfo->u.fmr.fmr_pgsize;
+                       unsigned int j;
+                       u64 prev = *kpage;
+                       /* check if adrs are contiguous */
+                       for (j = 1; j < cnt_per_hwpage; j++) {
+                               u64 p = phys_to_abs(fmrlist[j] &
+                                                   ~(pginfo->hwpage_size - 1));
+                               if (prev + pginfo->u.fmr.fmr_pgsize != p) {
+                                       ehca_gen_err("uncontiguous fmr pages "
+                                                    "found prev=%lx p=%lx "
+                                                    "idx=%x", prev, p, i + j);
+                                       return -EINVAL;
+                               }
+                               prev = p;
+                       }
+                       pginfo->kpage_cnt += cnt_per_hwpage;
+                       pginfo->u.fmr.next_listelem += cnt_per_hwpage;
+                       fmrlist += cnt_per_hwpage;
                }
+               kpage++;
        }
        return ret;
 }
@@ -1842,7 +2091,9 @@ int ehca_set_pagebuf(struct ehca_mr_pginfo *pginfo,
                ret = ehca_set_pagebuf_phys(pginfo, number, kpage);
                break;
        case EHCA_MR_PGI_USER:
-               ret = ehca_set_pagebuf_user1(pginfo, number, kpage);
+               ret = PAGE_SIZE >= pginfo->hwpage_size ?
+                       ehca_set_pagebuf_user1(pginfo, number, kpage) :
+                       ehca_set_pagebuf_user2(pginfo, number, kpage);
                break;
        case EHCA_MR_PGI_FMR:
                ret = ehca_set_pagebuf_fmr(pginfo, number, kpage);
@@ -1895,9 +2146,9 @@ void ehca_mrmw_map_acl(int ib_acl,
 /*----------------------------------------------------------------------*/
 
 /* sets page size in hipz access control for MR/MW. */
-void ehca_mrmw_set_pgsize_hipz_acl(u32 *hipz_acl) /*INOUT*/
+void ehca_mrmw_set_pgsize_hipz_acl(u32 pgsize, u32 *hipz_acl) /*INOUT*/
 {
-       return; /* HCA supports only 4k */
+       *hipz_acl |= (ehca_encode_hwpage_size(pgsize) << 24);
 } /* end ehca_mrmw_set_pgsize_hipz_acl() */
 
 /*----------------------------------------------------------------------*/
index 24f13fe3708bf4910f202b2ff93a47fc19222a17..bc8f4e31c1235c75655d84d9e1c1d60ca6a315f1 100644 (file)
@@ -111,7 +111,7 @@ int ehca_mr_is_maxmr(u64 size,
 void ehca_mrmw_map_acl(int ib_acl,
                       u32 *hipz_acl);
 
-void ehca_mrmw_set_pgsize_hipz_acl(u32 *hipz_acl);
+void ehca_mrmw_set_pgsize_hipz_acl(u32 pgsize, u32 *hipz_acl);
 
 void ehca_mrmw_reverse_map_acl(const u32 *hipz_acl,
                               int *ib_acl);
index 3394e05f4b4fcbc0d1a879ef705320ca9efe77a7..358796ccf00894607bc37e75f94cf2afe134372b 100644 (file)
@@ -427,7 +427,8 @@ u64 hipz_h_register_rpage(const struct ipz_adapter_handle adapter_handle,
 {
        return ehca_plpar_hcall_norets(H_REGISTER_RPAGES,
                                       adapter_handle.handle,      /* r4  */
-                                      queue_type | pagesize << 8, /* r5  */
+                                      (u64)queue_type | ((u64)pagesize) << 8,
+                                      /* r5  */
                                       resource_handle,            /* r6  */
                                       logical_address_of_page,    /* r7  */
                                       count,                      /* r8  */
@@ -724,6 +725,9 @@ u64 hipz_h_alloc_resource_mr(const struct ipz_adapter_handle adapter_handle,
        u64 ret;
        u64 outs[PLPAR_HCALL9_BUFSIZE];
 
+       ehca_gen_dbg("kernel PAGE_SIZE=%x access_ctrl=%016x "
+                    "vaddr=%lx length=%lx",
+                    (u32)PAGE_SIZE, access_ctrl, vaddr, length);
        ret = ehca_plpar_hcall9(H_ALLOC_RESOURCE, outs,
                                adapter_handle.handle,            /* r4 */
                                5,                                /* r5 */
@@ -746,8 +750,22 @@ u64 hipz_h_register_rpage_mr(const struct ipz_adapter_handle adapter_handle,
                             const u64 logical_address_of_page,
                             const u64 count)
 {
+       extern int ehca_debug_level;
        u64 ret;
 
+       if (unlikely(ehca_debug_level >= 2)) {
+               if (count > 1) {
+                       u64 *kpage;
+                       int i;
+                       kpage = (u64 *)abs_to_virt(logical_address_of_page);
+                       for (i = 0; i < count; i++)
+                               ehca_gen_dbg("kpage[%d]=%p",
+                                            i, (void *)kpage[i]);
+               } else
+                       ehca_gen_dbg("kpage=%p",
+                                    (void *)logical_address_of_page);
+       }
+
        if ((count > 1) && (logical_address_of_page & (EHCA_PAGESIZE-1))) {
                ehca_gen_err("logical_address_of_page not on a 4k boundary "
                             "adapter_handle=%lx mr=%p mr_handle=%lx "