From 777a71dd4d901f055967ddbd038d2a74ffce0eb8 Mon Sep 17 00:00:00 2001 From: Erez Zilber Date: Mon, 11 Sep 2006 12:19:17 +0300 Subject: IB/iser: fix a check of SG alignment for RDMA dma mapping may include a "compaction" of the sg associated with scsi command. Hence, the size of the maximal prefix of the SG which is aligned for rdma must be compared against the length of the dma mapped sg (mem->dma_nents) and not against the size of it before it was mapped (mem->size). Signed-off-by: Erez Zilber Signed-off-by: Roland Dreier --- drivers/infiniband/ulp/iser/iser_memory.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/infiniband/ulp/iser/iser_memory.c') diff --git a/drivers/infiniband/ulp/iser/iser_memory.c b/drivers/infiniband/ulp/iser/iser_memory.c index 31950a522a1c..53af9567632e 100644 --- a/drivers/infiniband/ulp/iser/iser_memory.c +++ b/drivers/infiniband/ulp/iser/iser_memory.c @@ -378,7 +378,7 @@ int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask, regd_buf = &iser_ctask->rdma_regd[cmd_dir]; aligned_len = iser_data_buf_aligned_len(mem); - if (aligned_len != mem->size) { + if (aligned_len != mem->dma_nents) { iser_err("rdma alignment violation %d/%d aligned\n", aligned_len, mem->size); iser_data_buf_dump(mem); -- cgit v1.2.3 From 8dfa0876d3dde5f9c1818a4c35caaabc3ddba78b Mon Sep 17 00:00:00 2001 From: Erez Zilber Date: Mon, 11 Sep 2006 12:22:30 +0300 Subject: IB/iser: make FMR "page size" be 4K and not PAGE_SIZE As iser is able to use at most one rdma operation for the execution of a scsi command, and registration of the sg associated with scsi command has its restrictions, the code checks if an sg is "aligned for rdma". Alignment for rdma is measured in "fmr page" units whose possible resolutions are different between HCAs and can be smaller, equal or bigger to the system page size. When the system page size is bigger than 4KB (eg the default with ia64 kernels) there a bigger chance that an sg would be aligned for rdma if the fmr page size is 4KB. Change the code to create FMR whose pages are of size 4KB and to take that into account when processing the sg. Signed-off-by: Erez Zilber Signed-off-by: Roland Dreier --- drivers/infiniband/ulp/iser/iscsi_iser.h | 6 +++++- drivers/infiniband/ulp/iser/iser_memory.c | 31 ++++++++++++++++++++----------- drivers/infiniband/ulp/iser/iser_verbs.c | 4 ++-- 3 files changed, 27 insertions(+), 14 deletions(-) (limited to 'drivers/infiniband/ulp/iser/iser_memory.c') diff --git a/drivers/infiniband/ulp/iser/iscsi_iser.h b/drivers/infiniband/ulp/iser/iscsi_iser.h index 3350ba690cfe..0ba02abb0414 100644 --- a/drivers/infiniband/ulp/iser/iscsi_iser.h +++ b/drivers/infiniband/ulp/iser/iscsi_iser.h @@ -82,8 +82,12 @@ __func__ , ## arg); \ } while (0) +#define SHIFT_4K 12 +#define SIZE_4K (1UL << SHIFT_4K) +#define MASK_4K (~(SIZE_4K-1)) + /* support upto 512KB in one RDMA */ -#define ISCSI_ISER_SG_TABLESIZE (0x80000 >> PAGE_SHIFT) +#define ISCSI_ISER_SG_TABLESIZE (0x80000 >> SHIFT_4K) #define ISCSI_ISER_MAX_LUN 256 #define ISCSI_ISER_MAX_CMD_LEN 16 diff --git a/drivers/infiniband/ulp/iser/iser_memory.c b/drivers/infiniband/ulp/iser/iser_memory.c index 53af9567632e..bcef0d31f756 100644 --- a/drivers/infiniband/ulp/iser/iser_memory.c +++ b/drivers/infiniband/ulp/iser/iser_memory.c @@ -42,6 +42,7 @@ #include "iscsi_iser.h" #define ISER_KMALLOC_THRESHOLD 0x20000 /* 128K - kmalloc limit */ + /** * Decrements the reference count for the * registered buffer & releases it @@ -239,7 +240,7 @@ static int iser_sg_to_page_vec(struct iser_data_buf *data, int i; /* compute the offset of first element */ - page_vec->offset = (u64) sg[0].offset; + page_vec->offset = (u64) sg[0].offset & ~MASK_4K; for (i = 0; i < data->dma_nents; i++) { total_sz += sg_dma_len(&sg[i]); @@ -247,21 +248,30 @@ static int iser_sg_to_page_vec(struct iser_data_buf *data, first_addr = sg_dma_address(&sg[i]); last_addr = first_addr + sg_dma_len(&sg[i]); - start_aligned = !(first_addr & ~PAGE_MASK); - end_aligned = !(last_addr & ~PAGE_MASK); + start_aligned = !(first_addr & ~MASK_4K); + end_aligned = !(last_addr & ~MASK_4K); /* continue to collect page fragments till aligned or SG ends */ while (!end_aligned && (i + 1 < data->dma_nents)) { i++; total_sz += sg_dma_len(&sg[i]); last_addr = sg_dma_address(&sg[i]) + sg_dma_len(&sg[i]); - end_aligned = !(last_addr & ~PAGE_MASK); + end_aligned = !(last_addr & ~MASK_4K); } - first_addr = first_addr & PAGE_MASK; - - for (page = first_addr; page < last_addr; page += PAGE_SIZE) - page_vec->pages[cur_page++] = page; + /* handle the 1st page in the 1st DMA element */ + if (cur_page == 0) { + page = first_addr & MASK_4K; + page_vec->pages[cur_page] = page; + cur_page++; + page += SIZE_4K; + } else + page = first_addr; + + for (; page < last_addr; page += SIZE_4K) { + page_vec->pages[cur_page] = page; + cur_page++; + } } page_vec->data_size = total_sz; @@ -269,8 +279,7 @@ static int iser_sg_to_page_vec(struct iser_data_buf *data, return cur_page; } -#define MASK_4K ((1UL << 12) - 1) /* 0xFFF */ -#define IS_4K_ALIGNED(addr) ((((unsigned long)addr) & MASK_4K) == 0) +#define IS_4K_ALIGNED(addr) ((((unsigned long)addr) & ~MASK_4K) == 0) /** * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned @@ -352,7 +361,7 @@ static void iser_page_vec_build(struct iser_data_buf *data, page_vec->length = page_vec_len; - if (page_vec_len * PAGE_SIZE < page_vec->data_size) { + if (page_vec_len * SIZE_4K < page_vec->data_size) { iser_err("page_vec too short to hold this SG\n"); iser_data_buf_dump(data); iser_dump_page_vec(page_vec); diff --git a/drivers/infiniband/ulp/iser/iser_verbs.c b/drivers/infiniband/ulp/iser/iser_verbs.c index 72febf1f8ff8..9b27a7c26aa8 100644 --- a/drivers/infiniband/ulp/iser/iser_verbs.c +++ b/drivers/infiniband/ulp/iser/iser_verbs.c @@ -150,7 +150,7 @@ static int iser_create_ib_conn_res(struct iser_conn *ib_conn) } ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1); - params.page_shift = PAGE_SHIFT; + params.page_shift = SHIFT_4K; /* when the first/last SG element are not start/end * * page aligned, the map whould be of N+1 pages */ params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1; @@ -604,7 +604,7 @@ int iser_reg_page_vec(struct iser_conn *ib_conn, mem_reg->lkey = mem->fmr->lkey; mem_reg->rkey = mem->fmr->rkey; - mem_reg->len = page_vec->length * PAGE_SIZE; + mem_reg->len = page_vec->length * SIZE_4K; mem_reg->va = io_addr; mem_reg->mem_h = (void *)mem; -- cgit v1.2.3 From e981f1d4b8288072ba7cf6b7141cd4aefb404383 Mon Sep 17 00:00:00 2001 From: Erez Zilber Date: Mon, 11 Sep 2006 12:24:00 +0300 Subject: IB/iser: fix some debug prints fix and add some debug prints related to iser handling of memory for rdma. Signed-off-by: Erez Zilber Signed-off-by: Roland Dreier --- drivers/infiniband/ulp/iser/iser_memory.c | 17 ++++++++++++++--- 1 file changed, 14 insertions(+), 3 deletions(-) (limited to 'drivers/infiniband/ulp/iser/iser_memory.c') diff --git a/drivers/infiniband/ulp/iser/iser_memory.c b/drivers/infiniband/ulp/iser/iser_memory.c index bcef0d31f756..8fea0bce5042 100644 --- a/drivers/infiniband/ulp/iser/iser_memory.c +++ b/drivers/infiniband/ulp/iser/iser_memory.c @@ -329,9 +329,9 @@ static void iser_data_buf_dump(struct iser_data_buf *data) struct scatterlist *sg = (struct scatterlist *)data->buf; int i; - for (i = 0; i < data->size; i++) + for (i = 0; i < data->dma_nents; i++) iser_err("sg[%d] dma_addr:0x%lX page:0x%p " - "off:%d sz:%d dma_len:%d\n", + "off:0x%x sz:0x%x dma_len:0x%x\n", i, (unsigned long)sg_dma_address(&sg[i]), sg[i].page, sg[i].offset, sg[i].length,sg_dma_len(&sg[i])); @@ -383,6 +383,7 @@ int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask, struct iser_regd_buf *regd_buf; int aligned_len; int err; + int i; regd_buf = &iser_ctask->rdma_regd[cmd_dir]; @@ -400,8 +401,18 @@ int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask, iser_page_vec_build(mem, ib_conn->page_vec); err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, ®d_buf->reg); - if (err) + if (err) { + iser_data_buf_dump(mem); + iser_err("mem->dma_nents = %d (dlength = 0x%x)\n", mem->dma_nents, + ntoh24(iser_ctask->desc.iscsi_header.dlength)); + iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n", + ib_conn->page_vec->data_size, ib_conn->page_vec->length, + ib_conn->page_vec->offset); + for (i=0 ; ipage_vec->length ; i++) { + iser_err("page_vec[%d] = 0x%lx\n", i, ib_conn->page_vec->pages[i]); + } return err; + } /* take a reference on this regd buf such that it will not be released * * (eg in send dto completion) before we get the scsi response */ -- cgit v1.2.3 From d81110285f7f6c07a0ce8f99a5ff158a647cd649 Mon Sep 17 00:00:00 2001 From: Erez Zilber Date: Mon, 11 Sep 2006 12:26:33 +0300 Subject: IB/iser: Do not use FMR for a single dma entry sg Fast Memory Registration (fmr) is used to register for rdma an sg whose elements are not linearly sequential after dma mapping. The IB verbs layer provides an "all dma memory MR (memory region)" which can be used for RDMA-ing a dma linearly sequential buffer. Change the code to use the dma mr instead of doing fmr when dma mapping produces a single dma entry sg. Signed-off-by: Erez Zilber Signed-off-by: Roland Dreier --- drivers/infiniband/ulp/iser/iscsi_iser.h | 1 + drivers/infiniband/ulp/iser/iser_memory.c | 48 ++++++++++++++++++++++--------- drivers/infiniband/ulp/iser/iser_verbs.c | 6 ++-- 3 files changed, 39 insertions(+), 16 deletions(-) (limited to 'drivers/infiniband/ulp/iser/iser_memory.c') diff --git a/drivers/infiniband/ulp/iser/iscsi_iser.h b/drivers/infiniband/ulp/iser/iscsi_iser.h index 0ba02abb0414..7e1a411db2a3 100644 --- a/drivers/infiniband/ulp/iser/iscsi_iser.h +++ b/drivers/infiniband/ulp/iser/iscsi_iser.h @@ -175,6 +175,7 @@ struct iser_mem_reg { u64 va; u64 len; void *mem_h; + int is_fmr; }; struct iser_regd_buf { diff --git a/drivers/infiniband/ulp/iser/iser_memory.c b/drivers/infiniband/ulp/iser/iser_memory.c index 8fea0bce5042..d0b03f426581 100644 --- a/drivers/infiniband/ulp/iser/iser_memory.c +++ b/drivers/infiniband/ulp/iser/iser_memory.c @@ -56,7 +56,7 @@ int iser_regd_buff_release(struct iser_regd_buf *regd_buf) if ((atomic_read(®d_buf->ref_count) == 0) || atomic_dec_and_test(®d_buf->ref_count)) { /* if we used the dma mr, unreg is just NOP */ - if (regd_buf->reg.rkey != 0) + if (regd_buf->reg.is_fmr) iser_unreg_mem(®d_buf->reg); if (regd_buf->dma_addr) { @@ -91,9 +91,9 @@ void iser_reg_single(struct iser_device *device, BUG_ON(dma_mapping_error(dma_addr)); regd_buf->reg.lkey = device->mr->lkey; - regd_buf->reg.rkey = 0; /* indicate there's no need to unreg */ regd_buf->reg.len = regd_buf->data_size; regd_buf->reg.va = dma_addr; + regd_buf->reg.is_fmr = 0; regd_buf->dma_addr = dma_addr; regd_buf->direction = direction; @@ -379,11 +379,13 @@ int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask, enum iser_data_dir cmd_dir) { struct iser_conn *ib_conn = iser_ctask->iser_conn->ib_conn; + struct iser_device *device = ib_conn->device; struct iser_data_buf *mem = &iser_ctask->data[cmd_dir]; struct iser_regd_buf *regd_buf; int aligned_len; int err; int i; + struct scatterlist *sg; regd_buf = &iser_ctask->rdma_regd[cmd_dir]; @@ -399,19 +401,37 @@ int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask, mem = &iser_ctask->data_copy[cmd_dir]; } - iser_page_vec_build(mem, ib_conn->page_vec); - err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, ®d_buf->reg); - if (err) { - iser_data_buf_dump(mem); - iser_err("mem->dma_nents = %d (dlength = 0x%x)\n", mem->dma_nents, - ntoh24(iser_ctask->desc.iscsi_header.dlength)); - iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n", - ib_conn->page_vec->data_size, ib_conn->page_vec->length, - ib_conn->page_vec->offset); - for (i=0 ; ipage_vec->length ; i++) { - iser_err("page_vec[%d] = 0x%lx\n", i, ib_conn->page_vec->pages[i]); + /* if there a single dma entry, FMR is not needed */ + if (mem->dma_nents == 1) { + sg = (struct scatterlist *)mem->buf; + + regd_buf->reg.lkey = device->mr->lkey; + regd_buf->reg.rkey = device->mr->rkey; + regd_buf->reg.len = sg_dma_len(&sg[0]); + regd_buf->reg.va = sg_dma_address(&sg[0]); + regd_buf->reg.is_fmr = 0; + + iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X " + "va: 0x%08lX sz: %ld]\n", + (unsigned int)regd_buf->reg.lkey, + (unsigned int)regd_buf->reg.rkey, + (unsigned long)regd_buf->reg.va, + (unsigned long)regd_buf->reg.len); + } else { /* use FMR for multiple dma entries */ + iser_page_vec_build(mem, ib_conn->page_vec); + err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, ®d_buf->reg); + if (err) { + iser_data_buf_dump(mem); + iser_err("mem->dma_nents = %d (dlength = 0x%x)\n", mem->dma_nents, + ntoh24(iser_ctask->desc.iscsi_header.dlength)); + iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n", + ib_conn->page_vec->data_size, ib_conn->page_vec->length, + ib_conn->page_vec->offset); + for (i=0 ; ipage_vec->length ; i++) + iser_err("page_vec[%d] = 0x%llx\n", i, + (unsigned long long) ib_conn->page_vec->pages[i]); + return err; } - return err; } /* take a reference on this regd buf such that it will not be released * diff --git a/drivers/infiniband/ulp/iser/iser_verbs.c b/drivers/infiniband/ulp/iser/iser_verbs.c index 9b27a7c26aa8..ecdca7fc1e4c 100644 --- a/drivers/infiniband/ulp/iser/iser_verbs.c +++ b/drivers/infiniband/ulp/iser/iser_verbs.c @@ -88,8 +88,9 @@ static int iser_create_device_ib_res(struct iser_device *device) iser_cq_tasklet_fn, (unsigned long)device); - device->mr = ib_get_dma_mr(device->pd, - IB_ACCESS_LOCAL_WRITE); + device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE | + IB_ACCESS_REMOTE_READ); if (IS_ERR(device->mr)) goto dma_mr_err; @@ -606,6 +607,7 @@ int iser_reg_page_vec(struct iser_conn *ib_conn, mem_reg->rkey = mem->fmr->rkey; mem_reg->len = page_vec->length * SIZE_4K; mem_reg->va = io_addr; + mem_reg->is_fmr = 1; mem_reg->mem_h = (void *)mem; mem_reg->va += page_vec->offset; -- cgit v1.2.3