From: Rusty Russell Date: Thu, 26 Jul 2007 17:41:03 +0000 (-0700) Subject: lguest: documentation III: Drivers X-Git-Tag: v2.6.23-rc2~222 X-Git-Url: https://err.no/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=e2c9784325490c878b7f69aeec1bed98b288bd97;p=linux-2.6 lguest: documentation III: Drivers Documentation: The Drivers Signed-off-by: Rusty Russell Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- diff --git a/drivers/block/lguest_blk.c b/drivers/block/lguest_blk.c index 5b79d07241..93e3c4001b 100644 --- a/drivers/block/lguest_blk.c +++ b/drivers/block/lguest_blk.c @@ -1,6 +1,12 @@ -/* A simple block driver for lguest. +/*D:400 + * The Guest block driver * - * Copyright 2006 Rusty Russell IBM Corporation + * This is a simple block driver, which appears as /dev/lgba, lgbb, lgbc etc. + * The mechanism is simple: we place the information about the request in the + * device page, then use SEND_DMA (containing the data for a write, or an empty + * "ping" DMA for a read). + :*/ +/* Copyright 2006 Rusty Russell IBM Corporation * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -25,27 +31,50 @@ static char next_block_index = 'a'; +/*D:420 Here is the structure which holds all the information we need about + * each Guest block device. + * + * I'm sure at this stage, you're wondering "hey, where was the adventure I was + * promised?" and thinking "Rusty sucks, I shall say nasty things about him on + * my blog". I think Real adventures have boring bits, too, and you're in the + * middle of one. But it gets better. Just not quite yet. */ struct blockdev { + /* The block queue infrastructure wants a spinlock: it is held while it + * calls our block request function. We grab it in our interrupt + * handler so the responses don't mess with new requests. */ spinlock_t lock; - /* The disk structure for the kernel. */ + /* The disk structure registered with kernel. */ struct gendisk *disk; - /* The major number for this disk. */ + /* The major device number for this disk, and the interrupt. We only + * really keep them here for completeness; we'd need them if we + * supported device unplugging. */ int major; int irq; + /* The physical address of this device's memory page */ unsigned long phys_addr; - /* The mapped block page. */ + /* The mapped memory page for convenient acces. */ struct lguest_block_page *lb_page; - /* We only have a single request outstanding at a time. */ + /* We only have a single request outstanding at a time: this is it. */ struct lguest_dma dma; struct request *req; }; -/* Jens gave me this nice helper to end all chunks of a request. */ +/*D:495 We originally used end_request() throughout the driver, but it turns + * out that end_request() is deprecated, and doesn't actually end the request + * (which seems like a good reason to deprecate it!). It simply ends the first + * bio. So if we had 3 bios in a "struct request" we would do all 3, + * end_request(), do 2, end_request(), do 1 and end_request(): twice as much + * work as we needed to do. + * + * This reinforced to me that I do not understand the block layer. + * + * Nonetheless, Jens Axboe gave me this nice helper to end all chunks of a + * request. This improved disk speed by 130%. */ static void end_entire_request(struct request *req, int uptodate) { if (end_that_request_first(req, uptodate, req->hard_nr_sectors)) @@ -55,30 +84,62 @@ static void end_entire_request(struct request *req, int uptodate) end_that_request_last(req, uptodate); } +/* I'm told there are only two stories in the world worth telling: love and + * hate. So there used to be a love scene here like this: + * + * Launcher: We could make beautiful I/O together, you and I. + * Guest: My, that's a big disk! + * + * Unfortunately, it was just too raunchy for our otherwise-gentle tale. */ + +/*D:490 This is the interrupt handler, called when a block read or write has + * been completed for us. */ static irqreturn_t lgb_irq(int irq, void *_bd) { + /* We handed our "struct blockdev" as the argument to request_irq(), so + * it is passed through to us here. This tells us which device we're + * dealing with in case we have more than one. */ struct blockdev *bd = _bd; unsigned long flags; + /* We weren't doing anything? Strange, but could happen if we shared + * interrupts (we don't!). */ if (!bd->req) { pr_debug("No work!\n"); return IRQ_NONE; } + /* Not done yet? That's equally strange. */ if (!bd->lb_page->result) { pr_debug("No result!\n"); return IRQ_NONE; } + /* We have to grab the lock before ending the request. */ spin_lock_irqsave(&bd->lock, flags); + /* "result" is 1 for success, 2 for failure: end_entire_request() wants + * to know whether this succeeded or not. */ end_entire_request(bd->req, bd->lb_page->result == 1); + /* Clear out request, it's done. */ bd->req = NULL; + /* Reset incoming DMA for next time. */ bd->dma.used_len = 0; + /* Ready for more reads or writes */ blk_start_queue(bd->disk->queue); spin_unlock_irqrestore(&bd->lock, flags); + + /* The interrupt was for us, we dealt with it. */ return IRQ_HANDLED; } +/*D:480 The block layer's "struct request" contains a number of "struct bio"s, + * each of which contains "struct bio_vec"s, each of which contains a page, an + * offset and a length. + * + * Fortunately there are iterators to help us walk through the "struct + * request". Even more fortunately, there were plenty of places to steal the + * code from. We pack the "struct request" into our "struct lguest_dma" and + * return the total length. */ static unsigned int req_to_dma(struct request *req, struct lguest_dma *dma) { unsigned int i = 0, idx, len = 0; @@ -87,8 +148,13 @@ static unsigned int req_to_dma(struct request *req, struct lguest_dma *dma) rq_for_each_bio(bio, req) { struct bio_vec *bvec; bio_for_each_segment(bvec, bio, idx) { + /* We told the block layer not to give us too many. */ BUG_ON(i == LGUEST_MAX_DMA_SECTIONS); + /* If we had a zero-length segment, it would look like + * the end of the data referred to by the "struct + * lguest_dma", so make sure that doesn't happen. */ BUG_ON(!bvec->bv_len); + /* Convert page & offset to a physical address */ dma->addr[i] = page_to_phys(bvec->bv_page) + bvec->bv_offset; dma->len[i] = bvec->bv_len; @@ -96,26 +162,39 @@ static unsigned int req_to_dma(struct request *req, struct lguest_dma *dma) i++; } } + /* If the array isn't full, we mark the end with a 0 length */ if (i < LGUEST_MAX_DMA_SECTIONS) dma->len[i] = 0; return len; } +/* This creates an empty DMA, useful for prodding the Host without sending data + * (ie. when we want to do a read) */ static void empty_dma(struct lguest_dma *dma) { dma->len[0] = 0; } +/*D:470 Setting up a request is fairly easy: */ static void setup_req(struct blockdev *bd, int type, struct request *req, struct lguest_dma *dma) { + /* The type is 1 (write) or 0 (read). */ bd->lb_page->type = type; + /* The sector on disk where the read or write starts. */ bd->lb_page->sector = req->sector; + /* The result is initialized to 0 (unfinished). */ bd->lb_page->result = 0; + /* The current request (so we can end it in the interrupt handler). */ bd->req = req; + /* The number of bytes: returned as a side-effect of req_to_dma(), + * which packs the block layer's "struct request" into our "struct + * lguest_dma" */ bd->lb_page->bytes = req_to_dma(req, dma); } +/*D:450 Write is pretty straightforward: we pack the request into a "struct + * lguest_dma", then use SEND_DMA to send the request. */ static void do_write(struct blockdev *bd, struct request *req) { struct lguest_dma send; @@ -126,6 +205,9 @@ static void do_write(struct blockdev *bd, struct request *req) lguest_send_dma(bd->phys_addr, &send); } +/* Read is similar to write, except we pack the request into our receive + * "struct lguest_dma" and send through an empty DMA just to tell the Host that + * there's a request pending. */ static void do_read(struct blockdev *bd, struct request *req) { struct lguest_dma ping; @@ -137,21 +219,30 @@ static void do_read(struct blockdev *bd, struct request *req) lguest_send_dma(bd->phys_addr, &ping); } +/*D:440 This where requests come in: we get handed the request queue and are + * expected to pull a "struct request" off it until we've finished them or + * we're waiting for a reply: */ static void do_lgb_request(struct request_queue *q) { struct blockdev *bd; struct request *req; again: + /* This sometimes returns NULL even on the very first time around. I + * wonder if it's something to do with letting elves handle the request + * queue... */ req = elv_next_request(q); if (!req) return; + /* We attached the struct blockdev to the disk: get it back */ bd = req->rq_disk->private_data; - /* Sometimes we get repeated requests after blk_stop_queue. */ + /* Sometimes we get repeated requests after blk_stop_queue(), but we + * can only handle one at a time. */ if (bd->req) return; + /* We only do reads and writes: no tricky business! */ if (!blk_fs_request(req)) { pr_debug("Got non-command 0x%08x\n", req->cmd_type); req->errors++; @@ -164,20 +255,31 @@ again: else do_read(bd, req); - /* Wait for interrupt to tell us it's done. */ + /* We've put out the request, so stop any more coming in until we get + * an interrupt, which takes us to lgb_irq() to re-enable the queue. */ blk_stop_queue(q); } +/*D:430 This is the "struct block_device_operations" we attach to the disk at + * the end of lguestblk_probe(). It doesn't seem to want much. */ static struct block_device_operations lguestblk_fops = { .owner = THIS_MODULE, }; +/*D:425 Setting up a disk device seems to involve a lot of code. I'm not sure + * quite why. I do know that the IDE code sent two or three of the maintainers + * insane, perhaps this is the fringe of the same disease? + * + * As in the console code, the probe function gets handed the generic + * lguest_device from lguest_bus.c: */ static int lguestblk_probe(struct lguest_device *lgdev) { struct blockdev *bd; int err; int irqflags = IRQF_SHARED; + /* First we allocate our own "struct blockdev" and initialize the easy + * fields. */ bd = kmalloc(sizeof(*bd), GFP_KERNEL); if (!bd) return -ENOMEM; @@ -187,59 +289,100 @@ static int lguestblk_probe(struct lguest_device *lgdev) bd->req = NULL; bd->dma.used_len = 0; bd->dma.len[0] = 0; + /* The descriptor in the lguest_devices array provided by the Host + * gives the Guest the physical page number of the device's page. */ bd->phys_addr = (lguest_devices[lgdev->index].pfn << PAGE_SHIFT); + /* We use lguest_map() to get a pointer to the device page */ bd->lb_page = lguest_map(bd->phys_addr, 1); if (!bd->lb_page) { err = -ENOMEM; goto out_free_bd; } + /* We need a major device number: 0 means "assign one dynamically". */ bd->major = register_blkdev(0, "lguestblk"); if (bd->major < 0) { err = bd->major; goto out_unmap; } + /* This allocates a "struct gendisk" where we pack all the information + * about the disk which the rest of Linux sees. We ask for one minor + * number; I do wonder if we should be asking for more. */ bd->disk = alloc_disk(1); if (!bd->disk) { err = -ENOMEM; goto out_unregister_blkdev; } + /* Every disk needs a queue for requests to come in: we set up the + * queue with a callback function (the core of our driver) and the lock + * to use. */ bd->disk->queue = blk_init_queue(do_lgb_request, &bd->lock); if (!bd->disk->queue) { err = -ENOMEM; goto out_put_disk; } - /* We can only handle a certain number of sg entries */ + /* We can only handle a certain number of pointers in our SEND_DMA + * call, so we set that with blk_queue_max_hw_segments(). This is not + * to be confused with blk_queue_max_phys_segments() of course! I + * know, who could possibly confuse the two? + * + * Well, it's simple to tell them apart: this one seems to work and the + * other one didn't. */ blk_queue_max_hw_segments(bd->disk->queue, LGUEST_MAX_DMA_SECTIONS); - /* Buffers must not cross page boundaries */ + + /* Due to technical limitations of our Host (and simple coding) we + * can't have a single buffer which crosses a page boundary. Tell it + * here. This means that our maximum request size is 16 + * (LGUEST_MAX_DMA_SECTIONS) pages. */ blk_queue_segment_boundary(bd->disk->queue, PAGE_SIZE-1); + /* We name our disk: this becomes the device name when udev does its + * magic thing and creates the device node, such as /dev/lgba. + * next_block_index is a global which starts at 'a'. Unfortunately + * this simple increment logic means that the 27th disk will be called + * "/dev/lgb{". In that case, I recommend having at least 29 disks, so + * your /dev directory will be balanced. */ sprintf(bd->disk->disk_name, "lgb%c", next_block_index++); + + /* We look to the device descriptor again to see if this device's + * interrupts are expected to be random. If they are, we tell the irq + * subsystem. At the moment this bit is always set. */ if (lguest_devices[lgdev->index].features & LGUEST_DEVICE_F_RANDOMNESS) irqflags |= IRQF_SAMPLE_RANDOM; + + /* Now we have the name and irqflags, we can request the interrupt; we + * give it the "struct blockdev" we have set up to pass to lgb_irq() + * when there is an interrupt. */ err = request_irq(bd->irq, lgb_irq, irqflags, bd->disk->disk_name, bd); if (err) goto out_cleanup_queue; + /* We bind our one-entry DMA pool to the key for this block device so + * the Host can reply to our requests. The key is equal to the + * physical address of the device's page, which is conveniently + * unique. */ err = lguest_bind_dma(bd->phys_addr, &bd->dma, 1, bd->irq); if (err) goto out_free_irq; + /* We finish our disk initialization and add the disk to the system. */ bd->disk->major = bd->major; bd->disk->first_minor = 0; bd->disk->private_data = bd; bd->disk->fops = &lguestblk_fops; - /* This is initialized to the disk size by the other end. */ + /* This is initialized to the disk size by the Launcher. */ set_capacity(bd->disk, bd->lb_page->num_sectors); add_disk(bd->disk); printk(KERN_INFO "%s: device %i at major %d\n", bd->disk->disk_name, lgdev->index, bd->major); + /* We don't need to keep the "struct blockdev" around, but if we ever + * implemented device removal, we'd need this. */ lgdev->private = bd; return 0; @@ -258,6 +401,8 @@ out_free_bd: return err; } +/*D:410 The boilerplate code for registering the lguest block driver is just + * like the console: */ static struct lguest_driver lguestblk_drv = { .name = "lguestblk", .owner = THIS_MODULE, diff --git a/drivers/char/hvc_lguest.c b/drivers/char/hvc_lguest.c index e7b889e404..1de8967cce 100644 --- a/drivers/char/hvc_lguest.c +++ b/drivers/char/hvc_lguest.c @@ -1,6 +1,19 @@ -/* Simple console for lguest. +/*D:300 + * The Guest console driver * - * Copyright (C) 2006 Rusty Russell, IBM Corporation + * This is a trivial console driver: we use lguest's DMA mechanism to send + * bytes out, and register a DMA buffer to receive bytes in. It is assumed to + * be present and available from the very beginning of boot. + * + * Writing console drivers is one of the few remaining Dark Arts in Linux. + * Fortunately for us, the path of virtual consoles has been well-trodden by + * the PowerPC folks, who wrote "hvc_console.c" to generically support any + * virtual console. We use that infrastructure which only requires us to write + * the basic put_chars and get_chars functions and call the right register + * functions. + :*/ + +/* Copyright (C) 2006 Rusty Russell, IBM Corporation * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -21,49 +34,81 @@ #include #include "hvc_console.h" +/*D:340 This is our single console input buffer, with associated "struct + * lguest_dma" referring to it. Note the 0-terminated length array, and the + * use of physical address for the buffer itself. */ static char inbuf[256]; static struct lguest_dma cons_input = { .used_len = 0, .addr[0] = __pa(inbuf), .len[0] = sizeof(inbuf), .len[1] = 0 }; +/*D:310 The put_chars() callback is pretty straightforward. + * + * First we put the pointer and length in a "struct lguest_dma": we only have + * one pointer, so we set the second length to 0. Then we use SEND_DMA to send + * the data to (Host) buffers attached to the console key. Usually a device's + * key is a physical address within the device's memory, but because the + * console device doesn't have any associated physical memory, we use the + * LGUEST_CONSOLE_DMA_KEY constant (aka 0). */ static int put_chars(u32 vtermno, const char *buf, int count) { struct lguest_dma dma; - /* FIXME: what if it's over a page boundary? */ + /* FIXME: DMA buffers in a "struct lguest_dma" are not allowed + * to go over page boundaries. This never seems to happen, + * but if it did we'd need to fix this code. */ dma.len[0] = count; dma.len[1] = 0; dma.addr[0] = __pa(buf); lguest_send_dma(LGUEST_CONSOLE_DMA_KEY, &dma); + /* We're expected to return the amount of data we wrote: all of it. */ return count; } +/*D:350 get_chars() is the callback from the hvc_console infrastructure when + * an interrupt is received. + * + * Firstly we see if our buffer has been filled: if not, we return. The rest + * of the code deals with the fact that the hvc_console() infrastructure only + * asks us for 16 bytes at a time. We keep a "cons_offset" variable for + * partially-read buffers. */ static int get_chars(u32 vtermno, char *buf, int count) { static int cons_offset; + /* Nothing left to see here... */ if (!cons_input.used_len) return 0; + /* You want more than we have to give? Well, try wanting less! */ if (cons_input.used_len - cons_offset < count) count = cons_input.used_len - cons_offset; + /* Copy across to their buffer and increment offset. */ memcpy(buf, inbuf + cons_offset, count); cons_offset += count; + + /* Finished? Zero offset, and reset cons_input so Host will use it + * again. */ if (cons_offset == cons_input.used_len) { cons_offset = 0; cons_input.used_len = 0; } return count; } +/*:*/ static struct hv_ops lguest_cons = { .get_chars = get_chars, .put_chars = put_chars, }; +/*D:320 Console drivers are initialized very early so boot messages can go + * out. At this stage, the console is output-only. Our driver checks we're a + * Guest, and if so hands hvc_instantiate() the console number (0), priority + * (0), and the struct hv_ops containing the put_chars() function. */ static int __init cons_init(void) { if (strcmp(paravirt_ops.name, "lguest") != 0) @@ -73,21 +118,46 @@ static int __init cons_init(void) } console_initcall(cons_init); +/*D:370 To set up and manage our virtual console, we call hvc_alloc() and + * stash the result in the private pointer of the "struct lguest_device". + * Since we never remove the console device we never need this pointer again, + * but using ->private is considered good form, and you never know who's going + * to copy your driver. + * + * Once the console is set up, we bind our input buffer ready for input. */ static int lguestcons_probe(struct lguest_device *lgdev) { int err; + /* The first argument of hvc_alloc() is the virtual console number, so + * we use zero. The second argument is the interrupt number. + * + * The third argument is a "struct hv_ops" containing the put_chars() + * and get_chars() pointers. The final argument is the output buffer + * size: we use 256 and expect the Host to have room for us to send + * that much. */ lgdev->private = hvc_alloc(0, lgdev_irq(lgdev), &lguest_cons, 256); if (IS_ERR(lgdev->private)) return PTR_ERR(lgdev->private); + /* We bind a single DMA buffer at key LGUEST_CONSOLE_DMA_KEY. + * "cons_input" is that statically-initialized global DMA buffer we saw + * above, and we also give the interrupt we want. */ err = lguest_bind_dma(LGUEST_CONSOLE_DMA_KEY, &cons_input, 1, lgdev_irq(lgdev)); if (err) printk("lguest console: failed to bind buffer.\n"); return err; } +/* Note the use of lgdev_irq() for the interrupt number. We tell hvc_alloc() + * to expect input when this interrupt is triggered, and then tell + * lguest_bind_dma() that is the interrupt to send us when input comes in. */ +/*D:360 From now on the console driver follows standard Guest driver form: + * register_lguest_driver() registers the device type and probe function, and + * the probe function sets up the device. + * + * The standard "struct lguest_driver": */ static struct lguest_driver lguestcons_drv = { .name = "lguestcons", .owner = THIS_MODULE, @@ -95,6 +165,7 @@ static struct lguest_driver lguestcons_drv = { .probe = lguestcons_probe, }; +/* The standard init function */ static int __init hvc_lguest_init(void) { return register_lguest_driver(&lguestcons_drv); diff --git a/drivers/lguest/lguest_bus.c b/drivers/lguest/lguest_bus.c index 9a22d19950..55a7940ca7 100644 --- a/drivers/lguest/lguest_bus.c +++ b/drivers/lguest/lguest_bus.c @@ -46,6 +46,10 @@ static struct device_attribute lguest_dev_attrs[] = { __ATTR_NULL }; +/*D:130 The generic bus infrastructure requires a function which says whether a + * device matches a driver. For us, it is simple: "struct lguest_driver" + * contains a "device_type" field which indicates what type of device it can + * handle, so we just cast the args and compare: */ static int lguest_dev_match(struct device *_dev, struct device_driver *_drv) { struct lguest_device *dev = container_of(_dev,struct lguest_device,dev); @@ -53,6 +57,7 @@ static int lguest_dev_match(struct device *_dev, struct device_driver *_drv) return (drv->device_type == lguest_devices[dev->index].type); } +/*:*/ struct lguest_bus { struct bus_type bus; @@ -71,11 +76,24 @@ static struct lguest_bus lguest_bus = { } }; +/*D:140 This is the callback which occurs once the bus infrastructure matches + * up a device and driver, ie. in response to add_lguest_device() calling + * device_register(), or register_lguest_driver() calling driver_register(). + * + * At the moment it's always the latter: the devices are added first, since + * scan_devices() is called from a "core_initcall", and the drivers themselves + * called later as a normal "initcall". But it would work the other way too. + * + * So now we have the happy couple, we add the status bit to indicate that we + * found a driver. If the driver truly loves the device, it will return + * happiness from its probe function (ok, perhaps this wasn't my greatest + * analogy), and we set the final "driver ok" bit so the Host sees it's all + * green. */ static int lguest_dev_probe(struct device *_dev) { int ret; - struct lguest_device *dev = container_of(_dev,struct lguest_device,dev); - struct lguest_driver *drv = container_of(dev->dev.driver, + struct lguest_device*dev = container_of(_dev,struct lguest_device,dev); + struct lguest_driver*drv = container_of(dev->dev.driver, struct lguest_driver, drv); lguest_devices[dev->index].status |= LGUEST_DEVICE_S_DRIVER; @@ -85,6 +103,10 @@ static int lguest_dev_probe(struct device *_dev) return ret; } +/* The last part of the bus infrastructure is the function lguest drivers use + * to register themselves. Firstly, we do nothing if there's no lguest bus + * (ie. this is not a Guest), otherwise we fill in the embedded generic "struct + * driver" fields and call the generic driver_register(). */ int register_lguest_driver(struct lguest_driver *drv) { if (!lguest_devices) @@ -97,12 +119,36 @@ int register_lguest_driver(struct lguest_driver *drv) return driver_register(&drv->drv); } + +/* At the moment we build all the drivers into the kernel because they're so + * simple: 8144 bytes for all three of them as I type this. And as the console + * really needs to be built in, it's actually only 3527 bytes for the network + * and block drivers. + * + * If they get complex it will make sense for them to be modularized, so we + * need to explicitly export the symbol. + * + * I don't think non-GPL modules make sense, so it's a GPL-only export. + */ EXPORT_SYMBOL_GPL(register_lguest_driver); +/*D:120 This is the core of the lguest bus: actually adding a new device. + * It's a separate function because it's neater that way, and because an + * earlier version of the code supported hotplug and unplug. They were removed + * early on because they were never used. + * + * As Andrew Tridgell says, "Untested code is buggy code". + * + * It's worth reading this carefully: we start with an index into the array of + * "struct lguest_device_desc"s indicating the device which is new: */ static void add_lguest_device(unsigned int index) { struct lguest_device *new; + /* Each "struct lguest_device_desc" has a "status" field, which the + * Guest updates as the device is probed. In the worst case, the Host + * can look at these bits to tell what part of device setup failed, + * even if the console isn't available. */ lguest_devices[index].status |= LGUEST_DEVICE_S_ACKNOWLEDGE; new = kmalloc(sizeof(struct lguest_device), GFP_KERNEL); if (!new) { @@ -111,12 +157,17 @@ static void add_lguest_device(unsigned int index) return; } + /* The "struct lguest_device" setup is pretty straight-forward example + * code. */ new->index = index; new->private = NULL; memset(&new->dev, 0, sizeof(new->dev)); new->dev.parent = &lguest_bus.dev; new->dev.bus = &lguest_bus.bus; sprintf(new->dev.bus_id, "%u", index); + + /* device_register() causes the bus infrastructure to look for a + * matching driver. */ if (device_register(&new->dev) != 0) { printk(KERN_EMERG "Cannot register lguest device %u\n", index); lguest_devices[index].status |= LGUEST_DEVICE_S_FAILED; @@ -124,6 +175,9 @@ static void add_lguest_device(unsigned int index) } } +/*D:110 scan_devices() simply iterates through the device array. The type 0 + * is reserved to mean "no device", and anything else means we have found a + * device: add it. */ static void scan_devices(void) { unsigned int i; @@ -133,12 +187,23 @@ static void scan_devices(void) add_lguest_device(i); } +/*D:100 Fairly early in boot, lguest_bus_init() is called to set up the lguest + * bus. We check that we are a Guest by checking paravirt_ops.name: there are + * other ways of checking, but this seems most obvious to me. + * + * So we can access the array of "struct lguest_device_desc"s easily, we map + * that memory and store the pointer in the global "lguest_devices". Then we + * register the bus with the core. Doing two registrations seems clunky to me, + * but it seems to be the correct sysfs incantation. + * + * Finally we call scan_devices() which adds all the devices found in the + * "struct lguest_device_desc" array. */ static int __init lguest_bus_init(void) { if (strcmp(paravirt_ops.name, "lguest") != 0) return 0; - /* Devices are in page above top of "normal" mem. */ + /* Devices are in a single page above top of "normal" mem */ lguest_devices = lguest_map(max_pfn< IBM Corporation + * This is very simple a virtual network driver, and our last Guest driver. + * The only trick is that it can talk directly to multiple other recipients + * (ie. other Guests on the same network). It can also be used with only the + * Host on the network. + :*/ + +/* Copyright 2006 Rusty Russell IBM Corporation * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -28,23 +35,28 @@ #define MAX_LANS 4 #define NUM_SKBS 8 +/*D:530 The "struct lguestnet_info" contains all the information we need to + * know about the network device. */ struct lguestnet_info { - /* The shared page(s). */ + /* The mapped device page(s) (an array of "struct lguest_net"). */ struct lguest_net *peer; + /* The physical address of the device page(s) */ unsigned long peer_phys; + /* The size of the device page(s). */ unsigned long mapsize; /* The lguest_device I come from */ struct lguest_device *lgdev; - /* My peerid. */ + /* My peerid (ie. my slot in the array). */ unsigned int me; - /* Receive queue. */ + /* Receive queue: the network packets waiting to be filled. */ struct sk_buff *skb[NUM_SKBS]; struct lguest_dma dma[NUM_SKBS]; }; +/*:*/ /* How many bytes left in this page. */ static unsigned int rest_of_page(void *data) @@ -52,39 +64,82 @@ static unsigned int rest_of_page(void *data) return PAGE_SIZE - ((unsigned long)data % PAGE_SIZE); } -/* Simple convention: offset 4 * peernum. */ +/*D:570 Each peer (ie. Guest or Host) on the network binds their receive + * buffers to a different key: we simply use the physical address of the + * device's memory page plus the peer number. The Host insists that all keys + * be a multiple of 4, so we multiply the peer number by 4. */ static unsigned long peer_key(struct lguestnet_info *info, unsigned peernum) { return info->peer_phys + 4 * peernum; } +/* This is the routine which sets up a "struct lguest_dma" to point to a + * network packet, similar to req_to_dma() in lguest_blk.c. The structure of a + * "struct sk_buff" has grown complex over the years: it consists of a "head" + * linear section pointed to by "skb->data", and possibly an array of + * "fragments" in the case of a non-linear packet. + * + * Our receive buffers don't use fragments at all but outgoing skbs might, so + * we handle it. */ static void skb_to_dma(const struct sk_buff *skb, unsigned int headlen, struct lguest_dma *dma) { unsigned int i, seg; + /* First, we put the linear region into the "struct lguest_dma". Each + * entry can't go over a page boundary, so even though all our packets + * are 1514 bytes or less, we might need to use two entries here: */ for (i = seg = 0; i < headlen; seg++, i += rest_of_page(skb->data+i)) { dma->addr[seg] = virt_to_phys(skb->data + i); dma->len[seg] = min((unsigned)(headlen - i), rest_of_page(skb->data + i)); } + + /* Now we handle the fragments: at least they're guaranteed not to go + * over a page. skb_shinfo(skb) returns a pointer to the structure + * which tells us about the number of fragments and the fragment + * array. */ for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, seg++) { const skb_frag_t *f = &skb_shinfo(skb)->frags[i]; /* Should not happen with MTU less than 64k - 2 * PAGE_SIZE. */ if (seg == LGUEST_MAX_DMA_SECTIONS) { + /* We will end up sending a truncated packet should + * this ever happen. Plus, a cool log message! */ printk("Woah dude! Megapacket!\n"); break; } dma->addr[seg] = page_to_phys(f->page) + f->page_offset; dma->len[seg] = f->size; } + + /* If after all that we didn't use the entire "struct lguest_dma" + * array, we terminate it with a 0 length. */ if (seg < LGUEST_MAX_DMA_SECTIONS) dma->len[seg] = 0; } -/* We overload multicast bit to show promiscuous mode. */ +/* + * Packet transmission. + * + * Our packet transmission is a little unusual. A real network card would just + * send out the packet and leave the receivers to decide if they're interested. + * Instead, we look through the network device memory page and see if any of + * the ethernet addresses match the packet destination, and if so we send it to + * that Guest. + * + * This is made a little more complicated in two cases. The first case is + * broadcast packets: for that we send the packet to all Guests on the network, + * one at a time. The second case is "promiscuous" mode, where a Guest wants + * to see all the packets on the network. We need a way for the Guest to tell + * us it wants to see all packets, so it sets the "multicast" bit on its + * published MAC address, which is never valid in a real ethernet address. + */ #define PROMISC_BIT 0x01 +/* This is the callback which is summoned whenever the network device's + * multicast or promiscuous state changes. If the card is in promiscuous mode, + * we advertise that in our ethernet address in the device's memory. We do the + * same if Linux wants any or all multicast traffic. */ static void lguestnet_set_multicast(struct net_device *dev) { struct lguestnet_info *info = netdev_priv(dev); @@ -95,11 +150,14 @@ static void lguestnet_set_multicast(struct net_device *dev) info->peer[info->me].mac[0] &= ~PROMISC_BIT; } +/* A simple test function to see if a peer wants to see all packets.*/ static int promisc(struct lguestnet_info *info, unsigned int peer) { return info->peer[peer].mac[0] & PROMISC_BIT; } +/* Another simple function to see if a peer's advertised ethernet address + * matches a packet's destination ethernet address. */ static int mac_eq(const unsigned char mac[ETH_ALEN], struct lguestnet_info *info, unsigned int peer) { @@ -109,6 +167,8 @@ static int mac_eq(const unsigned char mac[ETH_ALEN], return memcmp(mac+1, info->peer[peer].mac+1, ETH_ALEN-1) == 0; } +/* This is the function which actually sends a packet once we've decided a + * peer wants it: */ static void transfer_packet(struct net_device *dev, struct sk_buff *skb, unsigned int peernum) @@ -116,76 +176,134 @@ static void transfer_packet(struct net_device *dev, struct lguestnet_info *info = netdev_priv(dev); struct lguest_dma dma; + /* We use our handy "struct lguest_dma" packing function to prepare + * the skb for sending. */ skb_to_dma(skb, skb_headlen(skb), &dma); pr_debug("xfer length %04x (%u)\n", htons(skb->len), skb->len); + /* This is the actual send call which copies the packet. */ lguest_send_dma(peer_key(info, peernum), &dma); + + /* Check that the entire packet was transmitted. If not, it could mean + * that the other Guest registered a short receive buffer, but this + * driver should never do that. More likely, the peer is dead. */ if (dma.used_len != skb->len) { dev->stats.tx_carrier_errors++; pr_debug("Bad xfer to peer %i: %i of %i (dma %p/%i)\n", peernum, dma.used_len, skb->len, (void *)dma.addr[0], dma.len[0]); } else { + /* On success we update the stats. */ dev->stats.tx_bytes += skb->len; dev->stats.tx_packets++; } } +/* Another helper function to tell is if a slot in the device memory is unused. + * Since we always set the Local Assignment bit in the ethernet address, the + * first byte can never be 0. */ static int unused_peer(const struct lguest_net peer[], unsigned int num) { return peer[num].mac[0] == 0; } +/* Finally, here is the routine which handles an outgoing packet. It's called + * "start_xmit" for traditional reasons. */ static int lguestnet_start_xmit(struct sk_buff *skb, struct net_device *dev) { unsigned int i; int broadcast; struct lguestnet_info *info = netdev_priv(dev); + /* Extract the destination ethernet address from the packet. */ const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest; pr_debug("%s: xmit %02x:%02x:%02x:%02x:%02x:%02x\n", dev->name, dest[0],dest[1],dest[2],dest[3],dest[4],dest[5]); + /* If it's a multicast packet, we broadcast to everyone. That's not + * very efficient, but there are very few applications which actually + * use multicast, which is a shame really. + * + * As etherdevice.h points out: "By definition the broadcast address is + * also a multicast address." So we don't have to test for broadcast + * packets separately. */ broadcast = is_multicast_ether_addr(dest); + + /* Look through all the published ethernet addresses to see if we + * should send this packet. */ for (i = 0; i < info->mapsize/sizeof(struct lguest_net); i++) { + /* We don't send to ourselves (we actually can't SEND_DMA to + * ourselves anyway), and don't send to unused slots.*/ if (i == info->me || unused_peer(info->peer, i)) continue; + /* If it's broadcast we send it. If they want every packet we + * send it. If the destination matches their address we send + * it. Otherwise we go to the next peer. */ if (!broadcast && !promisc(info, i) && !mac_eq(dest, info, i)) continue; pr_debug("lguestnet %s: sending from %i to %i\n", dev->name, info->me, i); + /* Our routine which actually does the transfer. */ transfer_packet(dev, skb, i); } + + /* An xmit routine is expected to dispose of the packet, so we do. */ dev_kfree_skb(skb); + + /* As per kernel convention, 0 means success. This is why I love + * networking: even if we never sent to anyone, that's still + * success! */ return 0; } -/* Find a new skb to put in this slot in shared mem. */ +/*D:560 + * Packet receiving. + * + * First, here's a helper routine which fills one of our array of receive + * buffers: */ static int fill_slot(struct net_device *dev, unsigned int slot) { struct lguestnet_info *info = netdev_priv(dev); - /* Try to create and register a new one. */ + + /* We can receive ETH_DATA_LEN (1500) byte packets, plus a standard + * ethernet header of ETH_HLEN (14) bytes. */ info->skb[slot] = netdev_alloc_skb(dev, ETH_HLEN + ETH_DATA_LEN); if (!info->skb[slot]) { printk("%s: could not fill slot %i\n", dev->name, slot); return -ENOMEM; } + /* skb_to_dma() is a helper which sets up the "struct lguest_dma" to + * point to the data in the skb: we also use it for sending out a + * packet. */ skb_to_dma(info->skb[slot], ETH_HLEN + ETH_DATA_LEN, &info->dma[slot]); + + /* This is a Write Memory Barrier: it ensures that the entry in the + * receive buffer array is written *before* we set the "used_len" entry + * to 0. If the Host were looking at the receive buffer array from a + * different CPU, it could potentially see "used_len = 0" and not see + * the updated receive buffer information. This would be a horribly + * nasty bug, so make sure the compiler and CPU know this has to happen + * first. */ wmb(); - /* Now we tell hypervisor it can use the slot. */ + /* Writing 0 to "used_len" tells the Host it can use this receive + * buffer now. */ info->dma[slot].used_len = 0; return 0; } +/* This is the actual receive routine. When we receive an interrupt from the + * Host to tell us a packet has been delivered, we arrive here: */ static irqreturn_t lguestnet_rcv(int irq, void *dev_id) { struct net_device *dev = dev_id; struct lguestnet_info *info = netdev_priv(dev); unsigned int i, done = 0; + /* Look through our entire receive array for an entry which has data + * in it. */ for (i = 0; i < ARRAY_SIZE(info->dma); i++) { unsigned int length; struct sk_buff *skb; @@ -194,10 +312,16 @@ static irqreturn_t lguestnet_rcv(int irq, void *dev_id) if (length == 0) continue; + /* We've found one! Remember the skb (we grabbed the length + * above), and immediately refill the slot we've taken it + * from. */ done++; skb = info->skb[i]; fill_slot(dev, i); + /* This shouldn't happen: micropackets could be sent by a + * badly-behaved Guest on the network, but the Host will never + * stuff more data in the buffer than the buffer length. */ if (length < ETH_HLEN || length > ETH_HLEN + ETH_DATA_LEN) { pr_debug(KERN_WARNING "%s: unbelievable skb len: %i\n", dev->name, length); @@ -205,36 +329,72 @@ static irqreturn_t lguestnet_rcv(int irq, void *dev_id) continue; } + /* skb_put(), what a great function! I've ranted about this + * function before (http://lkml.org/lkml/1999/9/26/24). You + * call it after you've added data to the end of an skb (in + * this case, it was the Host which wrote the data). */ skb_put(skb, length); + + /* The ethernet header contains a protocol field: we use the + * standard helper to extract it, and place the result in + * skb->protocol. The helper also sets up skb->pkt_type and + * eats up the ethernet header from the front of the packet. */ skb->protocol = eth_type_trans(skb, dev); - /* This is a reliable transport. */ + + /* If this device doesn't need checksums for sending, we also + * don't need to check the packets when they come in. */ if (dev->features & NETIF_F_NO_CSUM) skb->ip_summed = CHECKSUM_UNNECESSARY; + + /* As a last resort for debugging the driver or the lguest I/O + * subsystem, you can uncomment the "#define DEBUG" at the top + * of this file, which turns all the pr_debug() into printk() + * and floods the logs. */ pr_debug("Receiving skb proto 0x%04x len %i type %i\n", ntohs(skb->protocol), skb->len, skb->pkt_type); + /* Update the packet and byte counts (visible from ifconfig, + * and good for debugging). */ dev->stats.rx_bytes += skb->len; dev->stats.rx_packets++; + + /* Hand our fresh network packet into the stack's "network + * interface receive" routine. That will free the packet + * itself when it's finished. */ netif_rx(skb); } + + /* If we found any packets, we assume the interrupt was for us. */ return done ? IRQ_HANDLED : IRQ_NONE; } +/*D:550 This is where we start: when the device is brought up by dhcpd or + * ifconfig. At this point we advertise our MAC address to the rest of the + * network, and register receive buffers ready for incoming packets. */ static int lguestnet_open(struct net_device *dev) { int i; struct lguestnet_info *info = netdev_priv(dev); - /* Set up our MAC address */ + /* Copy our MAC address into the device page, so others on the network + * can find us. */ memcpy(info->peer[info->me].mac, dev->dev_addr, ETH_ALEN); - /* Turn on promisc mode if needed */ + /* We might already be in promisc mode (dev->flags & IFF_PROMISC). Our + * set_multicast callback handles this already, so we call it now. */ lguestnet_set_multicast(dev); + /* Allocate packets and put them into our "struct lguest_dma" array. + * If we fail to allocate all the packets we could still limp along, + * but it's a sign of real stress so we should probably give up now. */ for (i = 0; i < ARRAY_SIZE(info->dma); i++) { if (fill_slot(dev, i) != 0) goto cleanup; } + + /* Finally we tell the Host where our array of "struct lguest_dma" + * receive buffers is, binding it to the key corresponding to the + * device's physical memory plus our peerid. */ if (lguest_bind_dma(peer_key(info,info->me), info->dma, NUM_SKBS, lgdev_irq(info->lgdev)) != 0) goto cleanup; @@ -245,22 +405,29 @@ cleanup: dev_kfree_skb(info->skb[i]); return -ENOMEM; } +/*:*/ +/* The close routine is called when the device is no longer in use: we clean up + * elegantly. */ static int lguestnet_close(struct net_device *dev) { unsigned int i; struct lguestnet_info *info = netdev_priv(dev); - /* Clear all trace: others might deliver packets, we'll ignore it. */ + /* Clear all trace of our existence out of the device memory by setting + * the slot which held our MAC address to 0 (unused). */ memset(&info->peer[info->me], 0, sizeof(info->peer[info->me])); - /* Deregister sg lists. */ + /* Unregister our array of receive buffers */ lguest_unbind_dma(peer_key(info, info->me), info->dma); for (i = 0; i < ARRAY_SIZE(info->dma); i++) dev_kfree_skb(info->skb[i]); return 0; } +/*D:510 The network device probe function is basically a standard ethernet + * device setup. It reads the "struct lguest_device_desc" and sets the "struct + * net_device". Oh, the line-by-line excitement! Let's skip over it. :*/ static int lguestnet_probe(struct lguest_device *lgdev) { int err, irqf = IRQF_SHARED; @@ -290,10 +457,16 @@ static int lguestnet_probe(struct lguest_device *lgdev) dev->stop = lguestnet_close; dev->hard_start_xmit = lguestnet_start_xmit; - /* Turning on/off promisc will call dev->set_multicast_list. - * We don't actually support multicast yet */ + /* We don't actually support multicast yet, but turning on/off + * promisc also calls dev->set_multicast_list. */ dev->set_multicast_list = lguestnet_set_multicast; SET_NETDEV_DEV(dev, &lgdev->dev); + + /* The network code complains if you have "scatter-gather" capability + * if you don't also handle checksums (it seem that would be + * "illogical"). So we use a lie of omission and don't tell it that we + * can handle scattered packets unless we also don't want checksums, + * even though to us they're completely independent. */ if (desc->features & LGUEST_NET_F_NOCSUM) dev->features = NETIF_F_SG|NETIF_F_NO_CSUM; @@ -325,6 +498,9 @@ static int lguestnet_probe(struct lguest_device *lgdev) } pr_debug("lguestnet: registered device %s\n", dev->name); + /* Finally, we put the "struct net_device" in the generic "struct + * lguest_device"s private pointer. Again, it's not necessary, but + * makes sure the cool kernel kids don't tease us. */ lgdev->private = dev; return 0; @@ -352,3 +528,11 @@ module_init(lguestnet_init); MODULE_DESCRIPTION("Lguest network driver"); MODULE_LICENSE("GPL"); + +/*D:580 + * This is the last of the Drivers, and with this we have covered the many and + * wonderous and fine (and boring) details of the Guest. + * + * "make Launcher" beckons, where we answer questions like "Where do Guests + * come from?", and "What do you do when someone asks for optimization?" + */ diff --git a/include/linux/lguest_bus.h b/include/linux/lguest_bus.h index c9b4e05fee..d27853ddc6 100644 --- a/include/linux/lguest_bus.h +++ b/include/linux/lguest_bus.h @@ -15,11 +15,14 @@ struct lguest_device { void *private; }; -/* By convention, each device can use irq index+1 if it wants to. */ +/*D:380 Since interrupt numbers are arbitrary, we use a convention: each device + * can use the interrupt number corresponding to its index. The +1 is because + * interrupt 0 is not usable (it's actually the timer interrupt). */ static inline int lgdev_irq(const struct lguest_device *dev) { return dev->index + 1; } +/*:*/ /* dma args must not be vmalloced! */ void lguest_send_dma(unsigned long key, struct lguest_dma *dma); diff --git a/include/linux/lguest_launcher.h b/include/linux/lguest_launcher.h index 0ba414a40c..6416705794 100644 --- a/include/linux/lguest_launcher.h +++ b/include/linux/lguest_launcher.h @@ -9,14 +9,45 @@ /* How many devices? Assume each one wants up to two dma arrays per device. */ #define LGUEST_MAX_DEVICES (LGUEST_MAX_DMA/2) +/*D:200 + * Lguest I/O + * + * The lguest I/O mechanism is the only way Guests can talk to devices. There + * are two hypercalls involved: SEND_DMA for output and BIND_DMA for input. In + * each case, "struct lguest_dma" describes the buffer: this contains 16 + * addr/len pairs, and if there are fewer buffer elements the len array is + * terminated with a 0. + * + * I/O is organized by keys: BIND_DMA attaches buffers to a particular key, and + * SEND_DMA transfers to buffers bound to particular key. By convention, keys + * correspond to a physical address within the device's page. This means that + * devices will never accidentally end up with the same keys, and allows the + * Host use The Futex Trick (as we'll see later in our journey). + * + * SEND_DMA simply indicates a key to send to, and the physical address of the + * "struct lguest_dma" to send. The Host will write the number of bytes + * transferred into the "struct lguest_dma"'s used_len member. + * + * BIND_DMA indicates a key to bind to, a pointer to an array of "struct + * lguest_dma"s ready for receiving, the size of that array, and an interrupt + * to trigger when data is received. The Host will only allow transfers into + * buffers with a used_len of zero: it then sets used_len to the number of + * bytes transferred and triggers the interrupt for the Guest to process the + * new input. */ struct lguest_dma { - /* 0 if free to be used, filled by hypervisor. */ + /* 0 if free to be used, filled by the Host. */ u32 used_len; unsigned long addr[LGUEST_MAX_DMA_SECTIONS]; u16 len[LGUEST_MAX_DMA_SECTIONS]; }; +/*:*/ +/*D:460 This is the layout of a block device memory page. The Launcher sets up + * the num_sectors initially to tell the Guest the size of the disk. The Guest + * puts the type, sector and length of the request in the first three fields, + * then DMAs to the Host. The Host processes the request, sets up the result, + * then DMAs back to the Guest. */ struct lguest_block_page { /* 0 is a read, 1 is a write. */ @@ -28,27 +59,47 @@ struct lguest_block_page u32 num_sectors; /* Disk length = num_sectors * 512 */ }; -/* There is a shared page of these. */ +/*D:520 The network device is basically a memory page where all the Guests on + * the network publish their MAC (ethernet) addresses: it's an array of "struct + * lguest_net": */ struct lguest_net { /* Simply the mac address (with multicast bit meaning promisc). */ unsigned char mac[6]; }; +/*:*/ /* Where the Host expects the Guest to SEND_DMA console output to. */ #define LGUEST_CONSOLE_DMA_KEY 0 -/* We have a page of these descriptors in the lguest_device page. */ +/*D:010 + * Drivers + * + * The Guest needs devices to do anything useful. Since we don't let it touch + * real devices (think of the damage it could do!) we provide virtual devices. + * We could emulate a PCI bus with various devices on it, but that is a fairly + * complex burden for the Host and suboptimal for the Guest, so we have our own + * "lguest" bus and simple drivers. + * + * Devices are described by an array of LGUEST_MAX_DEVICES of these structs, + * placed by the Launcher just above the top of physical memory: + */ struct lguest_device_desc { + /* The device type: console, network, disk etc. */ u16 type; #define LGUEST_DEVICE_T_CONSOLE 1 #define LGUEST_DEVICE_T_NET 2 #define LGUEST_DEVICE_T_BLOCK 3 + /* The specific features of this device: these depends on device type + * except for LGUEST_DEVICE_F_RANDOMNESS. */ u16 features; #define LGUEST_NET_F_NOCSUM 0x4000 /* Don't bother checksumming */ #define LGUEST_DEVICE_F_RANDOMNESS 0x8000 /* IRQ is fairly random */ + /* This is how the Guest reports status of the device: the Host can set + * LGUEST_DEVICE_S_REMOVED to indicate removal, but the rest are only + * ever manipulated by the Guest, and only ever set. */ u16 status; /* 256 and above are device specific. */ #define LGUEST_DEVICE_S_ACKNOWLEDGE 1 /* We have seen device. */ @@ -58,9 +109,12 @@ struct lguest_device_desc { #define LGUEST_DEVICE_S_REMOVED_ACK 16 /* Driver has been told. */ #define LGUEST_DEVICE_S_FAILED 128 /* Something actually failed */ + /* Each device exists somewhere in Guest physical memory, over some + * number of pages. */ u16 num_pages; u32 pfn; }; +/*:*/ /* Write command first word is a request. */ enum lguest_req