/** * * @file kmem.c * @brief This file contains all functions dealing with kernel memory. * @author Guillermo Marcus * @date 2009-04-05 * */ #include #include #include #include #include #include #include #include #include #include #include "config.h" /* compile-time configuration */ #include "compat.h" /* compatibility definitions for older linux */ #include "pciDriver.h" /* external interface for the driver */ #include "common.h" /* internal definitions for all parts */ #include "kmem.h" /* prototypes for kernel memory */ #include "sysfs.h" /* prototypes for sysfs */ /** * * Allocates new kernel memory including the corresponding management structure, makes * it available via sysfs if possible. * */ int pcidriver_kmem_alloc(pcidriver_privdata_t *privdata, kmem_handle_t *kmem_handle) { pcidriver_kmem_entry_t *kmem_entry; void *retptr; if (kmem_handle->flags&KMEM_FLAG_REUSE) { /* kmem_entry = pcidriver_kmem_find_entry_use(privdata, kmem_handle->use, kmem_handle->item); if (kmem_entry) { unsigned long flags = kmem_handle->flags; if (kmem_handle->type != kmem_entry->type) { mod_info("Invalid type of reusable kmem_entry\n"); return -EINVAL; } if (kmem_handle->type == PCILIB_KMEM_TYPE_PAGE) { kmem_handle->size = kmem_entry->size; } else if (kmem_handle->size != kmem_entry->size) { mod_info("Invalid size of reusable kmem_entry\n"); return -EINVAL; } if (((kmem_entry->mode&KMEM_MODE_EXCLUSIVE)?1:0) != ((flags&KMEM_FLAG_EXCLUSIVE)?1:0)) { mod_info("Invalid mode of reusable kmem_entry\n"); return -EINVAL; } if ((kmem_entry->mode&KMEM_MODE_COUNT)==KMEM_MODE_COUNT) { mod_info("Reuse counter of kmem_entry is overflown"); return -EBUSY; } kmem_handle->handle_id = kmem_entry->id; kmem_handle->pa = (unsigned long)(kmem_entry->dma_handle); kmem_handle->flags = KMEM_FLAG_REUSED; if (kmem_entry->refs&KMEM_REF_HW) kmem_handle->flags |= KMEM_FLAG_REUSED_HW; if (kmem_entry->mode&KMEM_MODE_PERSISTENT) kmem_handle->flags |= KMEM_FLAG_REUSED_PERSISTENT; kmem_entry->mode += 1; if (flags&KMEM_FLAG_HW) kmem_entry->refs |= KMEM_REF_HW; if (flags&KMEM_FLAG_PERSISTENT) kmem_entry->mode |= KMEM_MODE_PERSISTENT; privdata->kmem_cur_id = kmem_entry->id; return 0; }*/ } /* First, allocate zeroed memory for the kmem_entry */ if ((kmem_entry = kcalloc(1, sizeof(pcidriver_kmem_entry_t), GFP_KERNEL)) == NULL) goto kmem_alloc_entry_fail; /* Initialize the kmem_entry */ kmem_entry->id = atomic_inc_return(&privdata->kmem_count) - 1; privdata->kmem_cur_id = kmem_entry->id; kmem_handle->handle_id = kmem_entry->id; kmem_entry->use = kmem_handle->use; kmem_entry->item = kmem_handle->item; kmem_entry->type = kmem_handle->type; /* Initialize sysfs if possible */ if (pcidriver_sysfs_initialize_kmem(privdata, kmem_entry->id, &(kmem_entry->sysfs_attr)) != 0) goto kmem_alloc_mem_fail; /* ...and allocate the DMA memory */ /* note this is a memory pair, referencing the same area: the cpu address (cpua) * and the PCI bus address (pa). The CPU and PCI addresses may not be the same. * The CPU sees only CPU addresses, while the device sees only PCI addresses. * CPU address is used for the mmap (internal to the driver), and * PCI address is the address passed to the DMA Controller in the device. */ switch (kmem_entry->type) { case PCILIB_KMEM_TYPE_CONSISTENT: retptr = pci_alloc_consistent( privdata->pdev, kmem_handle->size, &(kmem_entry->dma_handle) ); break; case PCILIB_KMEM_TYPE_PAGE: retptr = (void*)__get_free_pages(GFP_KERNEL, get_order(PAGE_SIZE)); kmem_entry->dma_handle = 0; kmem_handle->size = PAGE_SIZE; // kmem_entry->dma_handle = pci_map_single(privdata->pdev, retptr, PAGE_SIZE, PCI_DMA_FROMDEVICE); // printk("%llx %lx\n", kmem_entry->dma_handle, retptr); break; default: goto kmem_alloc_mem_fail; } if (retptr == NULL) goto kmem_alloc_mem_fail; kmem_entry->size = kmem_handle->size; kmem_entry->cpua = (unsigned long)retptr; kmem_handle->pa = (unsigned long)(kmem_entry->dma_handle); kmem_entry->mode = 1; if (kmem_handle->flags&KMEM_FLAG_REUSE) { kmem_entry->mode |= KMEM_MODE_REUSABLE; if (kmem_handle->flags&KMEM_FLAG_EXCLUSIVE) kmem_entry->mode |= KMEM_MODE_EXCLUSIVE; if (kmem_handle->flags&KMEM_FLAG_PERSISTENT) kmem_entry->mode |= KMEM_MODE_PERSISTENT; } kmem_entry->refs = 0; if (kmem_handle->flags&KMEM_FLAG_HW) kmem_entry->refs |= KMEM_REF_HW; kmem_handle->flags = 0; set_pages_reserved_compat(kmem_entry->cpua, kmem_entry->size); /* Add the kmem_entry to the list of the device */ spin_lock( &(privdata->kmemlist_lock) ); list_add_tail( &(kmem_entry->list), &(privdata->kmem_list) ); spin_unlock( &(privdata->kmemlist_lock) ); return 0; kmem_alloc_mem_fail: kfree(kmem_entry); kmem_alloc_entry_fail: return -ENOMEM; } /** * * Called via sysfs, frees kernel memory and the corresponding management structure * */ int pcidriver_kmem_free( pcidriver_privdata_t *privdata, kmem_handle_t *kmem_handle ) { pcidriver_kmem_entry_t *kmem_entry; /* Find the associated kmem_entry for this buffer */ if ((kmem_entry = pcidriver_kmem_find_entry(privdata, kmem_handle)) == NULL) return -EINVAL; /* kmem_handle is not valid */ if (kmem_entry->mode&KMEM_MODE_COUNT) kmem_entry->mode -= 1; if (kmem_handle->flags&KMEM_FLAG_HW) kmem_entry->refs &= ~KMEM_REF_HW; if (kmem_handle->flags&KMEM_FLAG_PERSISTENT) kmem_entry->mode &= ~KMEM_MODE_PERSISTENT; if (kmem_handle->flags&KMEM_FLAG_REUSE) return 0; if (kmem_entry->refs) { mod_info("can't free referenced kmem_entry\n"); kmem_entry->mode += 1; return -EBUSY; } if (kmem_entry->mode & KMEM_MODE_PERSISTENT) { mod_info("can't free persistent kmem_entry\n"); return -EBUSY; } if (((kmem_entry->mode&KMEM_MODE_EXCLUSIVE)==0)&&(kmem_entry->mode&KMEM_MODE_COUNT)) return 0; return pcidriver_kmem_free_entry(privdata, kmem_entry); } /** * * Called when cleaning up, frees all kernel memory and their corresponding management structure * */ int pcidriver_kmem_free_all(pcidriver_privdata_t *privdata) { int failed = 0; struct list_head *ptr, *next; pcidriver_kmem_entry_t *kmem_entry; /* iterate safely over the entries and delete them */ list_for_each_safe(ptr, next, &(privdata->kmem_list)) { kmem_entry = list_entry(ptr, pcidriver_kmem_entry_t, list); /* if (kmem_entry->refs) failed = 1; else*/ pcidriver_kmem_free_entry(privdata, kmem_entry); /* spin lock inside! */ } if (failed) { mod_info("Some kmem_entries are still referenced\n"); return -EBUSY; } return 0; } /** * * Synchronize memory to/from the device (or in both directions). * */ int pcidriver_kmem_sync( pcidriver_privdata_t *privdata, kmem_sync_t *kmem_sync ) { pcidriver_kmem_entry_t *kmem_entry; /* Find the associated kmem_entry for this buffer */ if ((kmem_entry = pcidriver_kmem_find_entry(privdata, &(kmem_sync->handle))) == NULL) return -EINVAL; /* kmem_handle is not valid */ if (!kmem_entry->dma_handle) { mod_info_dbg("Instead of synchronization, we are mapping kmem_entry with id: %d\n", kmem_entry->id); if (kmem_sync->dir == PCIDRIVER_DMA_TODEVICE) kmem_entry->dma_handle = pci_map_single(privdata->pdev, (void*)kmem_entry->cpua, kmem_entry->size, PCI_DMA_TODEVICE); else kmem_entry->dma_handle = pci_map_single(privdata->pdev, (void*)kmem_entry->cpua, kmem_entry->size, PCI_DMA_FROMDEVICE); kmem_sync->handle.pa = kmem_entry->dma_handle; } #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,11) switch (kmem_sync->dir) { case PCIDRIVER_DMA_TODEVICE: pci_dma_sync_single_for_device( privdata->pdev, kmem_entry->dma_handle, kmem_entry->size, PCI_DMA_TODEVICE ); break; case PCIDRIVER_DMA_FROMDEVICE: pci_dma_sync_single_for_cpu( privdata->pdev, kmem_entry->dma_handle, kmem_entry->size, PCI_DMA_FROMDEVICE ); break; case PCIDRIVER_DMA_BIDIRECTIONAL: pci_dma_sync_single_for_device( privdata->pdev, kmem_entry->dma_handle, kmem_entry->size, PCI_DMA_BIDIRECTIONAL ); pci_dma_sync_single_for_cpu( privdata->pdev, kmem_entry->dma_handle, kmem_entry->size, PCI_DMA_BIDIRECTIONAL ); break; default: return -EINVAL; /* wrong direction parameter */ } #else switch (kmem_sync->dir) { case PCIDRIVER_DMA_TODEVICE: pci_dma_sync_single( privdata->pdev, kmem_entry->dma_handle, kmem_entry->size, PCI_DMA_TODEVICE ); break; case PCIDRIVER_DMA_FROMDEVICE: pci_dma_sync_single( privdata->pdev, kmem_entry->dma_handle, kmem_entry->size, PCI_DMA_FROMDEVICE ); break; case PCIDRIVER_DMA_BIDIRECTIONAL: pci_dma_sync_single( privdata->pdev, kmem_entry->dma_handle, kmem_entry->size, PCI_DMA_BIDIRECTIONAL ); break; default: return -EINVAL; /* wrong direction parameter */ } #endif return 0; /* success */ } /** * * Free the given kmem_entry and its memory. * */ int pcidriver_kmem_free_entry(pcidriver_privdata_t *privdata, pcidriver_kmem_entry_t *kmem_entry) { pcidriver_sysfs_remove(privdata, &(kmem_entry->sysfs_attr)); /* Go over the pages of the kmem buffer, and mark them as not reserved */ #if 0 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,15) /* * This code is DISABLED. * Apparently, it is not needed to unreserve them. Doing so here * hangs the machine. Why? * * Uhm.. see links: * * http://lwn.net/Articles/161204/ * http://lists.openfabrics.org/pipermail/general/2007-March/034101.html * * I insist, this should be enabled, but doing so hangs the machine. * Literature supports the point, and there is even a similar problem (see link) * But this is not the case. It seems right to me. but obviously is not. * * Anyway, this goes away in kernel >=2.6.15. */ unsigned long start = __pa(kmem_entry->cpua) >> PAGE_SHIFT; unsigned long end = __pa(kmem_entry->cpua + kmem_entry->size) >> PAGE_SHIFT; unsigned long i; for(i=start;itype) { case PCILIB_KMEM_TYPE_CONSISTENT: pci_free_consistent( privdata->pdev, kmem_entry->size, (void *)(kmem_entry->cpua), kmem_entry->dma_handle ); break; case PCILIB_KMEM_TYPE_PAGE: if (kmem_entry->dma_handle) pci_unmap_single(privdata->pdev, kmem_entry->dma_handle, kmem_entry->size, PCI_DMA_TODEVICE); free_page((unsigned long)kmem_entry->cpua); break; } /* Remove the kmem list entry */ spin_lock( &(privdata->kmemlist_lock) ); list_del( &(kmem_entry->list) ); spin_unlock( &(privdata->kmemlist_lock) ); /* Release kmem_entry memory */ kfree(kmem_entry); return 0; } /** * * Find the corresponding kmem_entry for the given kmem_handle. * */ pcidriver_kmem_entry_t *pcidriver_kmem_find_entry(pcidriver_privdata_t *privdata, kmem_handle_t *kmem_handle) { struct list_head *ptr; pcidriver_kmem_entry_t *entry, *result = NULL; /* should I implement it better using the handle_id? */ spin_lock(&(privdata->kmemlist_lock)); list_for_each(ptr, &(privdata->kmem_list)) { entry = list_entry(ptr, pcidriver_kmem_entry_t, list); if (entry->id == kmem_handle->handle_id) { result = entry; break; } } spin_unlock(&(privdata->kmemlist_lock)); return result; } /** * * find the corresponding kmem_entry for the given id. * */ pcidriver_kmem_entry_t *pcidriver_kmem_find_entry_id(pcidriver_privdata_t *privdata, int id) { struct list_head *ptr; pcidriver_kmem_entry_t *entry, *result = NULL; spin_lock(&(privdata->kmemlist_lock)); list_for_each(ptr, &(privdata->kmem_list)) { entry = list_entry(ptr, pcidriver_kmem_entry_t, list); if (entry->id == id) { result = entry; break; } } spin_unlock(&(privdata->kmemlist_lock)); return result; } /** * * find the corresponding kmem_entry for the given use and item. * */ pcidriver_kmem_entry_t *pcidriver_kmem_find_entry_use(pcidriver_privdata_t *privdata, unsigned long use, unsigned long item) { struct list_head *ptr; pcidriver_kmem_entry_t *entry, *result = NULL; spin_lock(&(privdata->kmemlist_lock)); list_for_each(ptr, &(privdata->kmem_list)) { entry = list_entry(ptr, pcidriver_kmem_entry_t, list); if ((entry->use == use)&&(entry->item == item)&&(entry->mode&KMEM_MODE_REUSABLE)) { result = entry; break; } } spin_unlock(&(privdata->kmemlist_lock)); return result; } void pcidriver_kmem_mmap_close(struct vm_area_struct *vma) { pcidriver_kmem_entry_t *kmem_entry = (pcidriver_kmem_entry_t*)vma->vm_private_data; if (kmem_entry) kmem_entry->refs -= 1; } static struct vm_operations_struct pcidriver_kmem_mmap_ops = { .close = pcidriver_kmem_mmap_close }; /** * * mmap() kernel memory to userspace. * */ int pcidriver_mmap_kmem(pcidriver_privdata_t *privdata, struct vm_area_struct *vma) { unsigned long vma_size; pcidriver_kmem_entry_t *kmem_entry; int ret; mod_info_dbg("Entering mmap_kmem\n"); /* FIXME: Is this really right? Always just the latest one? Can't we identify one? */ /* Get latest entry on the kmem_list */ kmem_entry = pcidriver_kmem_find_entry_id(privdata, privdata->kmem_cur_id); if (!kmem_entry) { mod_info("Trying to mmap a kernel memory buffer without creating it first!\n"); return -EFAULT; } mod_info_dbg("Got kmem_entry with id: %d\n", kmem_entry->id); /* Check sizes */ vma_size = (vma->vm_end - vma->vm_start); if ((vma_size != kmem_entry->size) && ((kmem_entry->size < PAGE_SIZE) && (vma_size != PAGE_SIZE))) { mod_info("kem_entry size(%lu) and vma size do not match(%lu)\n", kmem_entry->size, vma_size); return -EINVAL; } /* reference counting */ if ((kmem_entry->mode&KMEM_MODE_EXCLUSIVE)&&(kmem_entry->refs&KMEM_REF_COUNT)) { mod_info("can't make second mmaping for exclusive kmem_entry\n"); return -EBUSY; } if ((kmem_entry->refs&KMEM_REF_COUNT)==KMEM_REF_COUNT) { mod_info("maximal amount of references is reached by kmem_entry\n"); return -EBUSY; } kmem_entry->refs += 1; vma->vm_flags |= (VM_RESERVED); #ifdef pgprot_noncached // This is coherent memory, so it must not be cached. vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); #endif mod_info_dbg("Mapping address %08lx / PFN %08lx\n", virt_to_phys((void*)kmem_entry->cpua), page_to_pfn(virt_to_page((void*)kmem_entry->cpua))); ret = remap_pfn_range_cpua_compat( vma, vma->vm_start, kmem_entry->cpua, kmem_entry->size, vma->vm_page_prot ); if (ret) { mod_info("kmem remap failed: %d (%lx)\n", ret,kmem_entry->cpua); kmem_entry->refs -= 1; return -EAGAIN; } vma->vm_ops = &pcidriver_kmem_mmap_ops; vma->vm_private_data = (void*)kmem_entry; return ret; }