diff options
| -rw-r--r-- | drivers/accel/habanalabs/Kconfig | 23 | ||||
| -rw-r--r-- | drivers/accel/habanalabs/common/Makefile | 5 | ||||
| -rw-r--r-- | drivers/accel/habanalabs/common/debugfs.c | 3 | ||||
| -rw-r--r-- | drivers/accel/habanalabs/common/habanalabs.h | 12 | ||||
| -rw-r--r-- | drivers/accel/habanalabs/common/hldio.c | 437 | ||||
| -rw-r--r-- | drivers/accel/habanalabs/common/hldio.h | 146 |
6 files changed, 624 insertions, 2 deletions
diff --git a/drivers/accel/habanalabs/Kconfig b/drivers/accel/habanalabs/Kconfig index 1919fbb169c7..6d1506acbd72 100644 --- a/drivers/accel/habanalabs/Kconfig +++ b/drivers/accel/habanalabs/Kconfig @@ -27,3 +27,26 @@ config DRM_ACCEL_HABANALABS To compile this driver as a module, choose M here: the module will be called habanalabs. + +if DRM_ACCEL_HABANALABS + +config HL_HLDIO + bool "Habanalabs NVMe Direct I/O (HLDIO)" + depends on PCI_P2PDMA + depends on BLOCK + help + Enable NVMe peer-to-peer direct I/O support for Habanalabs AI + accelerators. + + This allows direct data transfers between NVMe storage devices + and Habanalabs accelerators without involving system memory, + using PCI peer-to-peer DMA capabilities. + + Requirements: + - CONFIG_PCI_P2PDMA=y + - NVMe device and Habanalabs accelerator under same PCI root complex + - IOMMU disabled or in passthrough mode + - Hardware supporting PCI P2P DMA + + If unsure, say N +endif # DRM_ACCEL_HABANALABS diff --git a/drivers/accel/habanalabs/common/Makefile b/drivers/accel/habanalabs/common/Makefile index e6abffea9f87..b6d00de09db5 100644 --- a/drivers/accel/habanalabs/common/Makefile +++ b/drivers/accel/habanalabs/common/Makefile @@ -13,3 +13,8 @@ HL_COMMON_FILES := common/habanalabs_drv.o common/device.o common/context.o \ common/command_submission.o common/firmware_if.o \ common/security.o common/state_dump.o \ common/memory_mgr.o common/decoder.o + +# Conditionally add HLDIO support +ifdef CONFIG_HL_HLDIO +HL_COMMON_FILES += common/hldio.o +endif
\ No newline at end of file diff --git a/drivers/accel/habanalabs/common/debugfs.c b/drivers/accel/habanalabs/common/debugfs.c index 4d975e8059ca..f3600df064d1 100644 --- a/drivers/accel/habanalabs/common/debugfs.c +++ b/drivers/accel/habanalabs/common/debugfs.c @@ -1891,6 +1891,7 @@ void hl_debugfs_device_fini(struct hl_device *hdev) vfree(entry->state_dump[i]); kfree(entry->entry_arr); + } void hl_debugfs_add_device(struct hl_device *hdev) @@ -1903,6 +1904,7 @@ void hl_debugfs_add_device(struct hl_device *hdev) if (!hdev->asic_prop.fw_security_enabled) add_secured_nodes(dev_entry, dev_entry->root); + } void hl_debugfs_add_file(struct hl_fpriv *hpriv) @@ -2035,3 +2037,4 @@ void hl_debugfs_set_state_dump(struct hl_device *hdev, char *data, up_write(&dev_entry->state_dump_sem); } + diff --git a/drivers/accel/habanalabs/common/habanalabs.h b/drivers/accel/habanalabs/common/habanalabs.h index 33aa385c2693..6be185d7be58 100644 --- a/drivers/accel/habanalabs/common/habanalabs.h +++ b/drivers/accel/habanalabs/common/habanalabs.h @@ -704,6 +704,7 @@ struct hl_hints_range { * @supports_advanced_cpucp_rc: true if new cpucp opcodes are supported. * @supports_engine_modes: true if changing engines/engine_cores modes is supported. * @support_dynamic_resereved_fw_size: true if we support dynamic reserved size for fw. + * @supports_nvme: indicates whether the asic supports NVMe P2P DMA. */ struct asic_fixed_properties { struct hw_queue_properties *hw_queues_props; @@ -824,6 +825,7 @@ struct asic_fixed_properties { u8 supports_advanced_cpucp_rc; u8 supports_engine_modes; u8 support_dynamic_resereved_fw_size; + u8 supports_nvme; }; /** @@ -2276,6 +2278,9 @@ struct hl_vm { u8 init_done; }; +#ifdef CONFIG_HL_HLDIO +#include "hldio.h" +#endif /* * DEBUG, PROFILING STRUCTURE @@ -2346,7 +2351,6 @@ struct hl_fpriv { struct mutex ctx_lock; }; - /* * DebugFS */ @@ -2374,6 +2378,7 @@ struct hl_debugfs_entry { struct hl_dbg_device_entry *dev_entry; }; + /** * struct hl_dbg_device_entry - ASIC specific debugfs manager. * @root: root dentry. @@ -3334,6 +3339,7 @@ struct eq_heartbeat_debug_info { * @captured_err_info: holds information about errors. * @reset_info: holds current device reset information. * @heartbeat_debug_info: counters used to debug heartbeat failures. + * @hldio: describes habanalabs direct storage interaction interface. * @irq_affinity_mask: mask of available CPU cores for user and decoder interrupt handling. * @stream_master_qid_arr: pointer to array with QIDs of master streams. * @fw_inner_major_ver: the major of current loaded preboot inner version. @@ -3527,7 +3533,9 @@ struct hl_device { struct hl_reset_info reset_info; struct eq_heartbeat_debug_info heartbeat_debug_info; - +#ifdef CONFIG_HL_HLDIO + struct hl_dio hldio; +#endif cpumask_t irq_affinity_mask; u32 *stream_master_qid_arr; diff --git a/drivers/accel/habanalabs/common/hldio.c b/drivers/accel/habanalabs/common/hldio.c new file mode 100644 index 000000000000..083ae5610875 --- /dev/null +++ b/drivers/accel/habanalabs/common/hldio.c @@ -0,0 +1,437 @@ +// SPDX-License-Identifier: GPL-2.0 + +/* + * Copyright 2024 HabanaLabs, Ltd. + * All Rights Reserved. + */ + +#include "habanalabs.h" +#include "hldio.h" +#include <generated/uapi/linux/version.h> +#include <linux/pci-p2pdma.h> +#include <linux/blkdev.h> +#include <linux/vmalloc.h> + +/* + * NVMe Direct I/O implementation for habanalabs driver + * + * ASSUMPTIONS + * =========== + * 1. No IOMMU (well, technically it can work with IOMMU, but it is *almost useless). + * 2. Only READ operations (can extend in the future). + * 3. No sparse files (can overcome this in the future). + * 4. Kernel version >= 6.9 + * 5. Requiring page alignment is OK (I don't see a solution to this one right, + * now, how do we read partial pages?) + * 6. Kernel compiled with CONFIG_PCI_P2PDMA. This requires a CUSTOM kernel. + * Theoretically I have a slight idea on how this could be solvable, but it + * is probably inacceptable for the upstream. Also may not work in the end. + * 7. Either make sure our cards and disks are under the same PCI bridge, or + * compile a custom kernel to hack around this. + */ + +#define IO_STABILIZE_TIMEOUT 10000000 /* 10 seconds in microseconds */ + +/* + * This struct contains all the useful data I could milk out of the file handle + * provided by the user. + * @TODO: right now it is retrieved on each IO, but can be done once with some + * dedicated IOCTL, call it for example HL_REGISTER_HANDLE. + */ +struct hl_dio_fd { + /* Back pointer in case we need it in async completion */ + struct hl_ctx *ctx; + /* Associated fd struct */ + struct file *filp; +}; + +/* + * This is a single IO descriptor + */ +struct hl_direct_io { + struct hl_dio_fd f; + struct kiocb kio; + struct bio_vec *bv; + struct iov_iter iter; + u64 device_va; + u64 off_bytes; + u64 len_bytes; + u32 type; +}; + +bool hl_device_supports_nvme(struct hl_device *hdev) +{ + return hdev->asic_prop.supports_nvme; +} + +static int hl_dio_fd_register(struct hl_ctx *ctx, int fd, struct hl_dio_fd *f) +{ + struct hl_device *hdev = ctx->hdev; + struct block_device *bd; + struct super_block *sb; + struct inode *inode; + struct gendisk *gd; + struct device *disk_dev; + int rc; + + f->filp = fget(fd); + if (!f->filp) { + rc = -ENOENT; + goto out; + } + + if (!(f->filp->f_flags & O_DIRECT)) { + dev_err(hdev->dev, "file is not in the direct mode\n"); + rc = -EINVAL; + goto fput; + } + + if (!f->filp->f_op->read_iter) { + dev_err(hdev->dev, "read iter is not supported, need to fall back to legacy\n"); + rc = -EINVAL; + goto fput; + } + + inode = file_inode(f->filp); + sb = inode->i_sb; + bd = sb->s_bdev; + gd = bd->bd_disk; + + if (inode->i_blocks << sb->s_blocksize_bits < i_size_read(inode)) { + dev_err(hdev->dev, "sparse files are not currently supported\n"); + rc = -EINVAL; + goto fput; + } + + if (!bd || !gd) { + dev_err(hdev->dev, "invalid block device\n"); + rc = -ENODEV; + goto fput; + } + /* Get the underlying device from the block device */ + disk_dev = disk_to_dev(gd); + if (!dma_pci_p2pdma_supported(disk_dev)) { + dev_err(hdev->dev, "device does not support PCI P2P DMA\n"); + rc = -EOPNOTSUPP; + goto fput; + } + + /* + * @TODO: Maybe we need additional checks here + */ + + f->ctx = ctx; + rc = 0; + + goto out; +fput: + fput(f->filp); +out: + return rc; +} + +static void hl_dio_fd_unregister(struct hl_dio_fd *f) +{ + fput(f->filp); +} + +static long hl_dio_count_io(struct hl_device *hdev) +{ + s64 sum = 0; + int i; + + for_each_possible_cpu(i) + sum += per_cpu(*hdev->hldio.inflight_ios, i); + + return sum; +} + +static bool hl_dio_get_iopath(struct hl_ctx *ctx) +{ + struct hl_device *hdev = ctx->hdev; + + if (hdev->hldio.io_enabled) { + this_cpu_inc(*hdev->hldio.inflight_ios); + + /* Avoid race conditions */ + if (!hdev->hldio.io_enabled) { + this_cpu_dec(*hdev->hldio.inflight_ios); + return false; + } + + hl_ctx_get(ctx); + + return true; + } + + return false; +} + +static void hl_dio_put_iopath(struct hl_ctx *ctx) +{ + struct hl_device *hdev = ctx->hdev; + + hl_ctx_put(ctx); + this_cpu_dec(*hdev->hldio.inflight_ios); +} + +static void hl_dio_set_io_enabled(struct hl_device *hdev, bool enabled) +{ + hdev->hldio.io_enabled = enabled; +} + +static bool hl_dio_validate_io(struct hl_device *hdev, struct hl_direct_io *io) +{ + if ((u64)io->device_va & ~PAGE_MASK) { + dev_dbg(hdev->dev, "device address must be 4K aligned\n"); + return false; + } + + if (io->len_bytes & ~PAGE_MASK) { + dev_dbg(hdev->dev, "IO length must be 4K aligned\n"); + return false; + } + + if (io->off_bytes & ~PAGE_MASK) { + dev_dbg(hdev->dev, "IO offset must be 4K aligned\n"); + return false; + } + + return true; +} + +static struct page *hl_dio_va2page(struct hl_device *hdev, struct hl_ctx *ctx, u64 device_va) +{ + struct hl_dio *hldio = &hdev->hldio; + u64 device_pa; + int rc, i; + + rc = hl_mmu_va_to_pa(ctx, device_va, &device_pa); + if (rc) { + dev_err(hdev->dev, "device virtual address translation error: %#llx (%d)", + device_va, rc); + return NULL; + } + + for (i = 0 ; i < hldio->np2prs ; ++i) { + if (device_pa >= hldio->p2prs[i].device_pa && + device_pa < hldio->p2prs[i].device_pa + hldio->p2prs[i].size) + return hldio->p2prs[i].p2ppages[(device_pa - hldio->p2prs[i].device_pa) >> + PAGE_SHIFT]; + } + + return NULL; +} + +static ssize_t hl_direct_io(struct hl_device *hdev, struct hl_direct_io *io) +{ + u64 npages, device_va; + ssize_t rc; + int i; + + if (!hl_dio_validate_io(hdev, io)) + return -EINVAL; + + if (!hl_dio_get_iopath(io->f.ctx)) { + dev_info(hdev->dev, "can't schedule a new IO, IO is disabled\n"); + return -ESHUTDOWN; + } + + init_sync_kiocb(&io->kio, io->f.filp); + io->kio.ki_pos = io->off_bytes; + + npages = (io->len_bytes >> PAGE_SHIFT); + + /* @TODO: this can be implemented smarter, vmalloc in iopath is not + * ideal. Maybe some variation of genpool. Number of pages may differ + * greatly, so maybe even use pools of different sizes and chose the + * closest one. + */ + io->bv = vzalloc(npages * sizeof(struct bio_vec)); + if (!io->bv) + return -ENOMEM; + + for (i = 0, device_va = io->device_va; i < npages ; ++i, device_va += PAGE_SIZE) { + io->bv[i].bv_page = hl_dio_va2page(hdev, io->f.ctx, device_va); + if (!io->bv[i].bv_page) { + dev_err(hdev->dev, "error getting page struct for device va %#llx", + device_va); + rc = -EFAULT; + goto cleanup; + } + io->bv[i].bv_offset = 0; + io->bv[i].bv_len = PAGE_SIZE; + } + + iov_iter_bvec(&io->iter, io->type, io->bv, 1, io->len_bytes); + if (io->f.filp->f_op && io->f.filp->f_op->read_iter) + rc = io->f.filp->f_op->read_iter(&io->kio, &io->iter); + else + rc = -EINVAL; + +cleanup: + vfree(io->bv); + hl_dio_put_iopath(io->f.ctx); + + dev_dbg(hdev->dev, "IO ended with %ld\n", rc); + + return rc; +} + +/* + * @TODO: This function can be used as a callback for io completion under + * kio->ki_complete in order to implement async IO. + * Note that on more recent kernels there is no ret2. + */ +__maybe_unused static void hl_direct_io_complete(struct kiocb *kio, long ret, long ret2) +{ + struct hl_direct_io *io = container_of(kio, struct hl_direct_io, kio); + + dev_dbg(io->f.ctx->hdev->dev, "IO completed with %ld\n", ret); + + /* Do something to copy result to user / notify completion */ + + hl_dio_put_iopath(io->f.ctx); + + hl_dio_fd_unregister(&io->f); +} + +/* + * DMA disk to ASIC, wait for results. Must be invoked from the user context + */ +int hl_dio_ssd2hl(struct hl_device *hdev, struct hl_ctx *ctx, int fd, + u64 device_va, off_t off_bytes, size_t len_bytes, + size_t *len_read) +{ + struct hl_direct_io *io; + ssize_t rc; + + dev_dbg(hdev->dev, "SSD2HL fd=%d va=%#llx len=%#lx\n", fd, device_va, len_bytes); + + io = kzalloc(sizeof(*io), GFP_KERNEL); + if (!io) { + rc = -ENOMEM; + goto out; + } + + *io = (struct hl_direct_io){ + .device_va = device_va, + .len_bytes = len_bytes, + .off_bytes = off_bytes, + .type = READ, + }; + + rc = hl_dio_fd_register(ctx, fd, &io->f); + if (rc) + goto kfree_io; + + rc = hl_direct_io(hdev, io); + if (rc >= 0) { + *len_read = rc; + rc = 0; + } + + /* This shall be called only in the case of a sync IO */ + hl_dio_fd_unregister(&io->f); +kfree_io: + kfree(io); +out: + return rc; +} + +static void hl_p2p_region_fini(struct hl_device *hdev, struct hl_p2p_region *p2pr) +{ + if (p2pr->p2ppages) { + vfree(p2pr->p2ppages); + p2pr->p2ppages = NULL; + } + + if (p2pr->p2pmem) { + dev_dbg(hdev->dev, "freeing P2P mem from %p, size=%#llx\n", + p2pr->p2pmem, p2pr->size); + pci_free_p2pmem(hdev->pdev, p2pr->p2pmem, p2pr->size); + p2pr->p2pmem = NULL; + } +} + +void hl_p2p_region_fini_all(struct hl_device *hdev) +{ + int i; + + for (i = 0 ; i < hdev->hldio.np2prs ; ++i) + hl_p2p_region_fini(hdev, &hdev->hldio.p2prs[i]); + + kvfree(hdev->hldio.p2prs); + hdev->hldio.p2prs = NULL; + hdev->hldio.np2prs = 0; +} + +int hl_p2p_region_init(struct hl_device *hdev, struct hl_p2p_region *p2pr) +{ + void *addr; + int rc, i; + + /* Start by publishing our p2p memory */ + rc = pci_p2pdma_add_resource(hdev->pdev, p2pr->bar, p2pr->size, p2pr->bar_offset); + if (rc) { + dev_err(hdev->dev, "error adding p2p resource: %d\n", rc); + goto err; + } + + /* Alloc all p2p mem */ + p2pr->p2pmem = pci_alloc_p2pmem(hdev->pdev, p2pr->size); + if (!p2pr->p2pmem) { + dev_err(hdev->dev, "error allocating p2p memory\n"); + rc = -ENOMEM; + goto err; + } + + p2pr->p2ppages = vmalloc((p2pr->size >> PAGE_SHIFT) * sizeof(struct page *)); + if (!p2pr->p2ppages) { + rc = -ENOMEM; + goto err; + } + + for (i = 0, addr = p2pr->p2pmem ; i < (p2pr->size >> PAGE_SHIFT) ; ++i, addr += PAGE_SIZE) { + p2pr->p2ppages[i] = virt_to_page(addr); + if (!p2pr->p2ppages[i]) { + rc = -EFAULT; + goto err; + } + } + + return 0; +err: + hl_p2p_region_fini(hdev, p2pr); + return rc; +} + +int hl_dio_start(struct hl_device *hdev) +{ + dev_dbg(hdev->dev, "initializing HLDIO\n"); + + /* Initialize the IO counter and enable IO */ + hdev->hldio.inflight_ios = alloc_percpu(s64); + if (!hdev->hldio.inflight_ios) + return -ENOMEM; + + hl_dio_set_io_enabled(hdev, true); + + return 0; +} + +void hl_dio_stop(struct hl_device *hdev) +{ + dev_dbg(hdev->dev, "deinitializing HLDIO\n"); + + if (hdev->hldio.io_enabled) { + /* Wait for all the IO to finish */ + hl_dio_set_io_enabled(hdev, false); + hl_poll_timeout_condition(hdev, !hl_dio_count_io(hdev), 1000, IO_STABILIZE_TIMEOUT); + } + + if (hdev->hldio.inflight_ios) { + free_percpu(hdev->hldio.inflight_ios); + hdev->hldio.inflight_ios = NULL; + } +} diff --git a/drivers/accel/habanalabs/common/hldio.h b/drivers/accel/habanalabs/common/hldio.h new file mode 100644 index 000000000000..2874388f2851 --- /dev/null +++ b/drivers/accel/habanalabs/common/hldio.h @@ -0,0 +1,146 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ +/* + * hldio.h - NVMe Direct I/O (HLDIO) infrastructure for Habana Labs Driver + * + * This feature requires specific hardware setup and must not be built + * under COMPILE_TEST. + */ + +#ifndef __HL_HLDIO_H__ +#define __HL_HLDIO_H__ + +#include <linux/types.h> +#include <linux/fs.h> +#include <linux/seq_file.h> +#include <linux/ktime.h> /* ktime functions */ +#include <linux/delay.h> /* usleep_range */ +#include <linux/kernel.h> /* might_sleep_if */ +#include <linux/errno.h> /* error codes */ + +/* Forward declarations */ +struct hl_device; +struct file; + +/* Enable only if Kconfig selected */ +#ifdef CONFIG_HL_HLDIO +/** + * struct hl_p2p_region - describes a single P2P memory region + * @p2ppages: array of page structs for the P2P memory + * @p2pmem: virtual address of the P2P memory region + * @device_pa: physical address on the device + * @bar_offset: offset within the BAR + * @size: size of the region in bytes + * @bar: BAR number containing this region + */ +struct hl_p2p_region { + struct page **p2ppages; + void *p2pmem; + u64 device_pa; + u64 bar_offset; + u64 size; + int bar; +}; + +/** + * struct hl_dio_stats - Direct I/O statistics + * @total_ops: total number of operations attempted + * @successful_ops: number of successful operations + * @failed_ops: number of failed operations + * @bytes_transferred: total bytes successfully transferred + * @last_len_read: length of the last read operation + */ +struct hl_dio_stats { + u64 total_ops; + u64 successful_ops; + u64 failed_ops; + u64 bytes_transferred; + size_t last_len_read; +}; + +/** + * struct hl_dio - describes habanalabs direct storage interaction interface + * @p2prs: array of p2p regions + * @inflight_ios: percpu counter for inflight ios + * @np2prs: number of elements in p2prs + * @io_enabled: 1 if io is enabled 0 otherwise + */ +struct hl_dio { + struct hl_p2p_region *p2prs; + s64 __percpu *inflight_ios; + u8 np2prs; + u8 io_enabled; +}; + +int hl_dio_ssd2hl(struct hl_device *hdev, struct hl_ctx *ctx, int fd, + u64 device_va, off_t off_bytes, size_t len_bytes, + size_t *len_read); +void hl_p2p_region_fini_all(struct hl_device *hdev); +int hl_p2p_region_init(struct hl_device *hdev, struct hl_p2p_region *p2pr); +int hl_dio_start(struct hl_device *hdev); +void hl_dio_stop(struct hl_device *hdev); + +/* Init/teardown */ +int hl_hldio_init(struct hl_device *hdev); +void hl_hldio_fini(struct hl_device *hdev); + +/* File operations */ +long hl_hldio_ioctl(struct file *filep, unsigned int cmd, unsigned long arg); + +/* DebugFS hooks */ +#ifdef CONFIG_DEBUG_FS +void hl_hldio_debugfs_init(struct hl_device *hdev); +void hl_hldio_debugfs_fini(struct hl_device *hdev); +#else +static inline void hl_hldio_debugfs_init(struct hl_device *hdev) { } +static inline void hl_hldio_debugfs_fini(struct hl_device *hdev) { } +#endif + +#else /* !CONFIG_HL_HLDIO */ + +struct hl_p2p_region; +/* Stubs when HLDIO is disabled */ +static inline int hl_dio_ssd2hl(struct hl_device *hdev, struct hl_ctx *ctx, int fd, + u64 device_va, off_t off_bytes, size_t len_bytes, + size_t *len_read) +{ return -EOPNOTSUPP; } +static inline void hl_p2p_region_fini_all(struct hl_device *hdev) {} +static inline int hl_p2p_region_init(struct hl_device *hdev, struct hl_p2p_region *p2pr) +{ return -EOPNOTSUPP; } +static inline int hl_dio_start(struct hl_device *hdev) { return -EOPNOTSUPP; } +static inline void hl_dio_stop(struct hl_device *hdev) {} + +static inline int hl_hldio_init(struct hl_device *hdev) { return 0; } +static inline void hl_hldio_fini(struct hl_device *hdev) { } +static inline long hl_hldio_ioctl(struct file *f, unsigned int c, + unsigned long a) +{ return -ENOTTY; } +static inline void hl_hldio_debugfs_init(struct hl_device *hdev) { } +static inline void hl_hldio_debugfs_fini(struct hl_device *hdev) { } + +#endif /* CONFIG_HL_HLDIO */ + +/* Simplified polling macro for HLDIO (no simulator support) */ +#define hl_poll_timeout_condition(hdev, cond, sleep_us, timeout_us) \ +({ \ + ktime_t __timeout = ktime_add_us(ktime_get(), timeout_us); \ + might_sleep_if(sleep_us); \ + (void)(hdev); /* keep signature consistent, hdev unused */ \ + for (;;) { \ + mb(); /* ensure ordering of memory operations */ \ + if (cond) \ + break; \ + if (timeout_us && ktime_compare(ktime_get(), __timeout) > 0) \ + break; \ + if (sleep_us) \ + usleep_range((sleep_us >> 2) + 1, sleep_us); \ + } \ + (cond) ? 0 : -ETIMEDOUT; \ +}) + +#ifdef CONFIG_HL_HLDIO +bool hl_device_supports_nvme(struct hl_device *hdev); +#else +static inline bool hl_device_supports_nvme(struct hl_device *hdev) { return false; } +#endif + +#endif /* __HL_HLDIO_H__ */ |