brintos

brintos / linux-shallow public Read only

0
0
Text · 19.1 KiB · 3c9b024 Raw
655 lines · c
1// SPDX-License-Identifier: GPL-2.0-only2/*3 *4 * Copyright (c) 2009, Microsoft Corporation.5 *6 * Authors:7 *   Haiyang Zhang <haiyangz@microsoft.com>8 *   Hank Janssen  <hjanssen@microsoft.com>9 *   K. Y. Srinivasan <kys@microsoft.com>10 */11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt12 13#include <linux/kernel.h>14#include <linux/mm.h>15#include <linux/hyperv.h>16#include <linux/uio.h>17#include <linux/vmalloc.h>18#include <linux/slab.h>19#include <linux/prefetch.h>20#include <linux/io.h>21#include <asm/mshyperv.h>22 23#include "hyperv_vmbus.h"24 25#define VMBUS_PKT_TRAILER	826 27/*28 * When we write to the ring buffer, check if the host needs to29 * be signaled. Here is the details of this protocol:30 *31 *	1. The host guarantees that while it is draining the32 *	   ring buffer, it will set the interrupt_mask to33 *	   indicate it does not need to be interrupted when34 *	   new data is placed.35 *36 *	2. The host guarantees that it will completely drain37 *	   the ring buffer before exiting the read loop. Further,38 *	   once the ring buffer is empty, it will clear the39 *	   interrupt_mask and re-check to see if new data has40 *	   arrived.41 *42 * KYS: Oct. 30, 2016:43 * It looks like Windows hosts have logic to deal with DOS attacks that44 * can be triggered if it receives interrupts when it is not expecting45 * the interrupt. The host expects interrupts only when the ring46 * transitions from empty to non-empty (or full to non full on the guest47 * to host ring).48 * So, base the signaling decision solely on the ring state until the49 * host logic is fixed.50 */51 52static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)53{54	struct hv_ring_buffer_info *rbi = &channel->outbound;55 56	virt_mb();57	if (READ_ONCE(rbi->ring_buffer->interrupt_mask))58		return;59 60	/* check interrupt_mask before read_index */61	virt_rmb();62	/*63	 * This is the only case we need to signal when the64	 * ring transitions from being empty to non-empty.65	 */66	if (old_write == READ_ONCE(rbi->ring_buffer->read_index)) {67		++channel->intr_out_empty;68		vmbus_setevent(channel);69	}70}71 72/* Get the next write location for the specified ring buffer. */73static inline u3274hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)75{76	u32 next = ring_info->ring_buffer->write_index;77 78	return next;79}80 81/* Set the next write location for the specified ring buffer. */82static inline void83hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,84		     u32 next_write_location)85{86	ring_info->ring_buffer->write_index = next_write_location;87}88 89/* Get the size of the ring buffer. */90static inline u3291hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)92{93	return ring_info->ring_datasize;94}95 96/* Get the read and write indices as u64 of the specified ring buffer. */97static inline u6498hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)99{100	return (u64)ring_info->ring_buffer->write_index << 32;101}102 103/*104 * Helper routine to copy from source to ring buffer.105 * Assume there is enough room. Handles wrap-around in dest case only!!106 */107static u32 hv_copyto_ringbuffer(108	struct hv_ring_buffer_info	*ring_info,109	u32				start_write_offset,110	const void			*src,111	u32				srclen)112{113	void *ring_buffer = hv_get_ring_buffer(ring_info);114	u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);115 116	memcpy(ring_buffer + start_write_offset, src, srclen);117 118	start_write_offset += srclen;119	if (start_write_offset >= ring_buffer_size)120		start_write_offset -= ring_buffer_size;121 122	return start_write_offset;123}124 125/*126 *127 * hv_get_ringbuffer_availbytes()128 *129 * Get number of bytes available to read and to write to130 * for the specified ring buffer131 */132static void133hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,134			     u32 *read, u32 *write)135{136	u32 read_loc, write_loc, dsize;137 138	/* Capture the read/write indices before they changed */139	read_loc = READ_ONCE(rbi->ring_buffer->read_index);140	write_loc = READ_ONCE(rbi->ring_buffer->write_index);141	dsize = rbi->ring_datasize;142 143	*write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :144		read_loc - write_loc;145	*read = dsize - *write;146}147 148/* Get various debug metrics for the specified ring buffer. */149int hv_ringbuffer_get_debuginfo(struct hv_ring_buffer_info *ring_info,150				struct hv_ring_buffer_debug_info *debug_info)151{152	u32 bytes_avail_towrite;153	u32 bytes_avail_toread;154 155	mutex_lock(&ring_info->ring_buffer_mutex);156 157	if (!ring_info->ring_buffer) {158		mutex_unlock(&ring_info->ring_buffer_mutex);159		return -EINVAL;160	}161 162	hv_get_ringbuffer_availbytes(ring_info,163				     &bytes_avail_toread,164				     &bytes_avail_towrite);165	debug_info->bytes_avail_toread = bytes_avail_toread;166	debug_info->bytes_avail_towrite = bytes_avail_towrite;167	debug_info->current_read_index = ring_info->ring_buffer->read_index;168	debug_info->current_write_index = ring_info->ring_buffer->write_index;169	debug_info->current_interrupt_mask170		= ring_info->ring_buffer->interrupt_mask;171	mutex_unlock(&ring_info->ring_buffer_mutex);172 173	return 0;174}175EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);176 177/* Initialize a channel's ring buffer info mutex locks */178void hv_ringbuffer_pre_init(struct vmbus_channel *channel)179{180	mutex_init(&channel->inbound.ring_buffer_mutex);181	mutex_init(&channel->outbound.ring_buffer_mutex);182}183 184/* Initialize the ring buffer. */185int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,186		       struct page *pages, u32 page_cnt, u32 max_pkt_size)187{188	struct page **pages_wraparound;189	int i;190 191	BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));192 193	/*194	 * First page holds struct hv_ring_buffer, do wraparound mapping for195	 * the rest.196	 */197	pages_wraparound = kcalloc(page_cnt * 2 - 1,198				   sizeof(struct page *),199				   GFP_KERNEL);200	if (!pages_wraparound)201		return -ENOMEM;202 203	pages_wraparound[0] = pages;204	for (i = 0; i < 2 * (page_cnt - 1); i++)205		pages_wraparound[i + 1] =206			&pages[i % (page_cnt - 1) + 1];207 208	ring_info->ring_buffer = (struct hv_ring_buffer *)209		vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP,210			pgprot_decrypted(PAGE_KERNEL));211 212	kfree(pages_wraparound);213	if (!ring_info->ring_buffer)214		return -ENOMEM;215 216	/*217	 * Ensure the header page is zero'ed since218	 * encryption status may have changed.219	 */220	memset(ring_info->ring_buffer, 0, HV_HYP_PAGE_SIZE);221 222	ring_info->ring_buffer->read_index =223		ring_info->ring_buffer->write_index = 0;224 225	/* Set the feature bit for enabling flow control. */226	ring_info->ring_buffer->feature_bits.value = 1;227 228	ring_info->ring_size = page_cnt << PAGE_SHIFT;229	ring_info->ring_size_div10_reciprocal =230		reciprocal_value(ring_info->ring_size / 10);231	ring_info->ring_datasize = ring_info->ring_size -232		sizeof(struct hv_ring_buffer);233	ring_info->priv_read_index = 0;234 235	/* Initialize buffer that holds copies of incoming packets */236	if (max_pkt_size) {237		ring_info->pkt_buffer = kzalloc(max_pkt_size, GFP_KERNEL);238		if (!ring_info->pkt_buffer)239			return -ENOMEM;240		ring_info->pkt_buffer_size = max_pkt_size;241	}242 243	spin_lock_init(&ring_info->ring_lock);244 245	return 0;246}247 248/* Cleanup the ring buffer. */249void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)250{251	mutex_lock(&ring_info->ring_buffer_mutex);252	vunmap(ring_info->ring_buffer);253	ring_info->ring_buffer = NULL;254	mutex_unlock(&ring_info->ring_buffer_mutex);255 256	kfree(ring_info->pkt_buffer);257	ring_info->pkt_buffer = NULL;258	ring_info->pkt_buffer_size = 0;259}260 261/*262 * Check if the ring buffer spinlock is available to take or not; used on263 * atomic contexts, like panic path (see the Hyper-V framebuffer driver).264 */265 266bool hv_ringbuffer_spinlock_busy(struct vmbus_channel *channel)267{268	struct hv_ring_buffer_info *rinfo = &channel->outbound;269 270	return spin_is_locked(&rinfo->ring_lock);271}272EXPORT_SYMBOL_GPL(hv_ringbuffer_spinlock_busy);273 274/* Write to the ring buffer. */275int hv_ringbuffer_write(struct vmbus_channel *channel,276			const struct kvec *kv_list, u32 kv_count,277			u64 requestid, u64 *trans_id)278{279	int i;280	u32 bytes_avail_towrite;281	u32 totalbytes_towrite = sizeof(u64);282	u32 next_write_location;283	u32 old_write;284	u64 prev_indices;285	unsigned long flags;286	struct hv_ring_buffer_info *outring_info = &channel->outbound;287	struct vmpacket_descriptor *desc = kv_list[0].iov_base;288	u64 __trans_id, rqst_id = VMBUS_NO_RQSTOR;289 290	if (channel->rescind)291		return -ENODEV;292 293	for (i = 0; i < kv_count; i++)294		totalbytes_towrite += kv_list[i].iov_len;295 296	spin_lock_irqsave(&outring_info->ring_lock, flags);297 298	bytes_avail_towrite = hv_get_bytes_to_write(outring_info);299 300	/*301	 * If there is only room for the packet, assume it is full.302	 * Otherwise, the next time around, we think the ring buffer303	 * is empty since the read index == write index.304	 */305	if (bytes_avail_towrite <= totalbytes_towrite) {306		++channel->out_full_total;307 308		if (!channel->out_full_flag) {309			++channel->out_full_first;310			channel->out_full_flag = true;311		}312 313		spin_unlock_irqrestore(&outring_info->ring_lock, flags);314		return -EAGAIN;315	}316 317	channel->out_full_flag = false;318 319	/* Write to the ring buffer */320	next_write_location = hv_get_next_write_location(outring_info);321 322	old_write = next_write_location;323 324	for (i = 0; i < kv_count; i++) {325		next_write_location = hv_copyto_ringbuffer(outring_info,326						     next_write_location,327						     kv_list[i].iov_base,328						     kv_list[i].iov_len);329	}330 331	/*332	 * Allocate the request ID after the data has been copied into the333	 * ring buffer.  Once this request ID is allocated, the completion334	 * path could find the data and free it.335	 */336 337	if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {338		if (channel->next_request_id_callback != NULL) {339			rqst_id = channel->next_request_id_callback(channel, requestid);340			if (rqst_id == VMBUS_RQST_ERROR) {341				spin_unlock_irqrestore(&outring_info->ring_lock, flags);342				return -EAGAIN;343			}344		}345	}346	desc = hv_get_ring_buffer(outring_info) + old_write;347	__trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;348	/*349	 * Ensure the compiler doesn't generate code that reads the value of350	 * the transaction ID from the ring buffer, which is shared with the351	 * Hyper-V host and subject to being changed at any time.352	 */353	WRITE_ONCE(desc->trans_id, __trans_id);354	if (trans_id)355		*trans_id = __trans_id;356 357	/* Set previous packet start */358	prev_indices = hv_get_ring_bufferindices(outring_info);359 360	next_write_location = hv_copyto_ringbuffer(outring_info,361					     next_write_location,362					     &prev_indices,363					     sizeof(u64));364 365	/* Issue a full memory barrier before updating the write index */366	virt_mb();367 368	/* Now, update the write location */369	hv_set_next_write_location(outring_info, next_write_location);370 371 372	spin_unlock_irqrestore(&outring_info->ring_lock, flags);373 374	hv_signal_on_write(old_write, channel);375 376	if (channel->rescind) {377		if (rqst_id != VMBUS_NO_RQSTOR) {378			/* Reclaim request ID to avoid leak of IDs */379			if (channel->request_addr_callback != NULL)380				channel->request_addr_callback(channel, rqst_id);381		}382		return -ENODEV;383	}384 385	return 0;386}387 388int hv_ringbuffer_read(struct vmbus_channel *channel,389		       void *buffer, u32 buflen, u32 *buffer_actual_len,390		       u64 *requestid, bool raw)391{392	struct vmpacket_descriptor *desc;393	u32 packetlen, offset;394 395	if (unlikely(buflen == 0))396		return -EINVAL;397 398	*buffer_actual_len = 0;399	*requestid = 0;400 401	/* Make sure there is something to read */402	desc = hv_pkt_iter_first(channel);403	if (desc == NULL) {404		/*405		 * No error is set when there is even no header, drivers are406		 * supposed to analyze buffer_actual_len.407		 */408		return 0;409	}410 411	offset = raw ? 0 : (desc->offset8 << 3);412	packetlen = (desc->len8 << 3) - offset;413	*buffer_actual_len = packetlen;414	*requestid = desc->trans_id;415 416	if (unlikely(packetlen > buflen))417		return -ENOBUFS;418 419	/* since ring is double mapped, only one copy is necessary */420	memcpy(buffer, (const char *)desc + offset, packetlen);421 422	/* Advance ring index to next packet descriptor */423	__hv_pkt_iter_next(channel, desc);424 425	/* Notify host of update */426	hv_pkt_iter_close(channel);427 428	return 0;429}430 431/*432 * Determine number of bytes available in ring buffer after433 * the current iterator (priv_read_index) location.434 *435 * This is similar to hv_get_bytes_to_read but with private436 * read index instead.437 */438static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)439{440	u32 priv_read_loc = rbi->priv_read_index;441	u32 write_loc;442 443	/*444	 * The Hyper-V host writes the packet data, then uses445	 * store_release() to update the write_index.  Use load_acquire()446	 * here to prevent loads of the packet data from being re-ordered447	 * before the read of the write_index and potentially getting448	 * stale data.449	 */450	write_loc = virt_load_acquire(&rbi->ring_buffer->write_index);451 452	if (write_loc >= priv_read_loc)453		return write_loc - priv_read_loc;454	else455		return (rbi->ring_datasize - priv_read_loc) + write_loc;456}457 458/*459 * Get first vmbus packet from ring buffer after read_index460 *461 * If ring buffer is empty, returns NULL and no other action needed.462 */463struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)464{465	struct hv_ring_buffer_info *rbi = &channel->inbound;466	struct vmpacket_descriptor *desc, *desc_copy;467	u32 bytes_avail, pkt_len, pkt_offset;468 469	hv_debug_delay_test(channel, MESSAGE_DELAY);470 471	bytes_avail = hv_pkt_iter_avail(rbi);472	if (bytes_avail < sizeof(struct vmpacket_descriptor))473		return NULL;474	bytes_avail = min(rbi->pkt_buffer_size, bytes_avail);475 476	desc = (struct vmpacket_descriptor *)(hv_get_ring_buffer(rbi) + rbi->priv_read_index);477 478	/*479	 * Ensure the compiler does not use references to incoming Hyper-V values (which480	 * could change at any moment) when reading local variables later in the code481	 */482	pkt_len = READ_ONCE(desc->len8) << 3;483	pkt_offset = READ_ONCE(desc->offset8) << 3;484 485	/*486	 * If pkt_len is invalid, set it to the smaller of hv_pkt_iter_avail() and487	 * rbi->pkt_buffer_size488	 */489	if (pkt_len < sizeof(struct vmpacket_descriptor) || pkt_len > bytes_avail)490		pkt_len = bytes_avail;491 492	/*493	 * If pkt_offset is invalid, arbitrarily set it to494	 * the size of vmpacket_descriptor495	 */496	if (pkt_offset < sizeof(struct vmpacket_descriptor) || pkt_offset > pkt_len)497		pkt_offset = sizeof(struct vmpacket_descriptor);498 499	/* Copy the Hyper-V packet out of the ring buffer */500	desc_copy = (struct vmpacket_descriptor *)rbi->pkt_buffer;501	memcpy(desc_copy, desc, pkt_len);502 503	/*504	 * Hyper-V could still change len8 and offset8 after the earlier read.505	 * Ensure that desc_copy has legal values for len8 and offset8 that506	 * are consistent with the copy we just made507	 */508	desc_copy->len8 = pkt_len >> 3;509	desc_copy->offset8 = pkt_offset >> 3;510 511	return desc_copy;512}513EXPORT_SYMBOL_GPL(hv_pkt_iter_first);514 515/*516 * Get next vmbus packet from ring buffer.517 *518 * Advances the current location (priv_read_index) and checks for more519 * data. If the end of the ring buffer is reached, then return NULL.520 */521struct vmpacket_descriptor *522__hv_pkt_iter_next(struct vmbus_channel *channel,523		   const struct vmpacket_descriptor *desc)524{525	struct hv_ring_buffer_info *rbi = &channel->inbound;526	u32 packetlen = desc->len8 << 3;527	u32 dsize = rbi->ring_datasize;528 529	hv_debug_delay_test(channel, MESSAGE_DELAY);530	/* bump offset to next potential packet */531	rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;532	if (rbi->priv_read_index >= dsize)533		rbi->priv_read_index -= dsize;534 535	/* more data? */536	return hv_pkt_iter_first(channel);537}538EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);539 540/* How many bytes were read in this iterator cycle */541static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,542					u32 start_read_index)543{544	if (rbi->priv_read_index >= start_read_index)545		return rbi->priv_read_index - start_read_index;546	else547		return rbi->ring_datasize - start_read_index +548			rbi->priv_read_index;549}550 551/*552 * Update host ring buffer after iterating over packets. If the host has553 * stopped queuing new entries because it found the ring buffer full, and554 * sufficient space is being freed up, signal the host. But be careful to555 * only signal the host when necessary, both for performance reasons and556 * because Hyper-V protects itself by throttling guests that signal557 * inappropriately.558 *559 * Determining when to signal is tricky. There are three key data inputs560 * that must be handled in this order to avoid race conditions:561 *562 * 1. Update the read_index563 * 2. Read the pending_send_sz564 * 3. Read the current write_index565 *566 * The interrupt_mask is not used to determine when to signal. The567 * interrupt_mask is used only on the guest->host ring buffer when568 * sending requests to the host. The host does not use it on the host->569 * guest ring buffer to indicate whether it should be signaled.570 */571void hv_pkt_iter_close(struct vmbus_channel *channel)572{573	struct hv_ring_buffer_info *rbi = &channel->inbound;574	u32 curr_write_sz, pending_sz, bytes_read, start_read_index;575 576	/*577	 * Make sure all reads are done before we update the read index since578	 * the writer may start writing to the read area once the read index579	 * is updated.580	 */581	virt_rmb();582	start_read_index = rbi->ring_buffer->read_index;583	rbi->ring_buffer->read_index = rbi->priv_read_index;584 585	/*586	 * Older versions of Hyper-V (before WS2102 and Win8) do not587	 * implement pending_send_sz and simply poll if the host->guest588	 * ring buffer is full.  No signaling is needed or expected.589	 */590	if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)591		return;592 593	/*594	 * Issue a full memory barrier before making the signaling decision.595	 * If reading pending_send_sz were to be reordered and happen596	 * before we commit the new read_index, a race could occur.  If the597	 * host were to set the pending_send_sz after we have sampled598	 * pending_send_sz, and the ring buffer blocks before we commit the599	 * read index, we could miss sending the interrupt. Issue a full600	 * memory barrier to address this.601	 */602	virt_mb();603 604	/*605	 * If the pending_send_sz is zero, then the ring buffer is not606	 * blocked and there is no need to signal.  This is far by the607	 * most common case, so exit quickly for best performance.608	 */609	pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);610	if (!pending_sz)611		return;612 613	/*614	 * Ensure the read of write_index in hv_get_bytes_to_write()615	 * happens after the read of pending_send_sz.616	 */617	virt_rmb();618	curr_write_sz = hv_get_bytes_to_write(rbi);619	bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);620 621	/*622	 * We want to signal the host only if we're transitioning623	 * from a "not enough free space" state to a "enough free624	 * space" state.  For example, it's possible that this function625	 * could run and free up enough space to signal the host, and then626	 * run again and free up additional space before the host has a627	 * chance to clear the pending_send_sz.  The 2nd invocation would628	 * be a null transition from "enough free space" to "enough free629	 * space", which doesn't warrant a signal.630	 *631	 * Exactly filling the ring buffer is treated as "not enough632	 * space". The ring buffer always must have at least one byte633	 * empty so the empty and full conditions are distinguishable.634	 * hv_get_bytes_to_write() doesn't fully tell the truth in635	 * this regard.636	 *637	 * So first check if we were in the "enough free space" state638	 * before we began the iteration. If so, the host was not639	 * blocked, and there's no need to signal.640	 */641	if (curr_write_sz - bytes_read > pending_sz)642		return;643 644	/*645	 * Similarly, if the new state is "not enough space", then646	 * there's no need to signal.647	 */648	if (curr_write_sz <= pending_sz)649		return;650 651	++channel->intr_in_full;652	vmbus_setevent(channel);653}654EXPORT_SYMBOL_GPL(hv_pkt_iter_close);655